>
Voice of Volge is a series of books, and the firm's thought leadership on the solutions and the domains Volge specialises in. Each edition takes one of them and sets out what the technology makes possible, what it makes mandatory, and what an operator has to put in place before either is true of them.
Pilots impress. Production pays. Most automation stops at a departmental win. Scaling it takes a framework the whole business runs on, agents included.
Figures carry their basis and their date. Where published estimates disagree, the range is given and the basis named rather than averaged away, and where a widely quoted number does not survive checking it is corrected here rather than repeated. Every chart plots a figure stated in the text, and nothing is projected beyond what a named source projects.
Automation has a curious record. Nearly every enterprise has done it, and very few have collected the return. The typical organisation has automated the same class of work several times over, a departmental workflow here, a point tool there, a proof of concept that impressed a steering committee and then quietly expired. Activity is abundant. Compounding return is rare.
The reason is not the technology, and it has not been for years. A workflow built for one team fails to become a system the business runs on because it is never given what production demands, which is integration with the systems of record, governance a risk committee will sign, and a deterministic floor beneath the probabilistic parts. The newest version of the failure wears an AI badge, the agent demo that dazzles and then stalls, but the pattern predates it, and so does the cure.
This report is for the executive who approved that spend and now has to justify it, the COO, CIO or transformation lead who has watched the distance between a working demo and a working business swallow a budget. It is written to be read once, cover to cover, in about the time it takes to fly between two offices. The evidence is drawn from Gartner, MIT, S&P Global and Deloitte, dated and quoted, and the architecture is described plainly enough to argue with.
One platform, Nintex, supplies the running example, and Volge discloses the interest as its delivery partner. In shape, the argument is platform-agnostic. Automation pays a dividend only when a workflow stops being a departmental win and becomes how the operation runs. One framework the whole business runs on, agents included.
The Voice of Volge desk
Nintex is a process-and-automation software company. What began as SharePoint-era workflow is now a single platform that maps work, automates it and, since 2025, orchestrates AI agents against it. That was built out over two decades and a handful of pointed acquisitions.
The recent chapters matter most to a buyer today: Agentic Business Orchestration (September 2025), locally-hosted AI in Automation K2 5.9.1 (April 2026), Agent Designer and Nintex Orchestration shipping inside the consolidated Nintex CE Platform (September 2026), and Nintex Processes (26 October 2026).
Source: Nintex corporate and product materials, 2006–2026. Ownership is majority-held by TPG Capital (2021), with Thoma Bravo a minority holder. Ratings run G2 4.2/5 across 1,326 reviews and TrustRadius 8.0/10, read 6 November 2026. Automation K2 5.9.1 of April 2026 is still the current release. Nintex lists 5.10 as the next minor release, with a community release event on 17 November 2026.
This report covers Workflow Automation, the Build & run layer, where the work is built and run at scale across the business.
| Understand | Process Manager · Nintex Processes |
|---|---|
| Build & run Focus of this report | Automation CE · Automation K2 (on-prem) · RPA · DocGen · Forms · eSign · Solutions |
| Orchestrate | Agentic Business Orchestration · Agent Designer · Nintex Orchestration |
Source: G2 and TrustRadius aggregate ratings, read 6 November 2026; The Forrester Wave: Agentic Process Orchestration Platforms, Q3 2026; Forrester Total Economic Impact of Nintex Workflow, commissioned by Nintex and conducted in late 2022.
This is not a young product. It is a two-decade platform that has just added governed AI and a locally-hosted engine, mature where it must be, current where it counts, and now built to orchestrate agents at scale.
The demo is quick, impressive, and inexpensive. The production system is governed, integrated, and paid back. This book is about crossing from the first to the second.
The model was never the problem. A probabilistic system was asked to do a deterministic job, without a governed floor beneath it. Put the floor in, and the demo becomes a production line.
Begin with the uncomfortable number. Over 40% of agentic-AI projects will be cancelled by the end of 2027, Gartner said on 25 June 2025, on escalating cost, unclear value and inadequate risk controls. That is not a forecast about bad models. It is a forecast about bad plumbing. The models arrived; the governance, the integration, and the deterministic substrate did not.
The pattern repeats across every serious study of the past eighteen months. The GenAI Divide, published under the MIT NANDA masthead in July 2025, found that roughly 95% of organisations are getting zero return from generative AI. S&P Global's 451 Research reported on 13 January 2025 that 42% of companies had abandoned most of their AI initiatives before reaching production, up from 17% a year earlier, and that organisations scrap around 46% of projects between proof of concept and broad adoption. Deloitte, taking the more generous view, still found only a quarter of its respondents had moved at least 40% of their pilots into production (State of AI in the Enterprise 2026, 21 January 2026).
Read together, these are not four opinions. They are one finding, measured four ways. The pilot is easy and the production system is hard, and almost everyone is stuck on the wrong side of that gap.
Amit Mathradas, then chief executive of Nintex, Business Wire, 24 September 2025
The reason the survivors ship is unglamorous. They stop treating the agent as the whole system and start treating it as one component inside a governed workflow. Judgement goes to the AI where judgement adds value. Precision goes to a deterministic workflow where compliance and repeatability are non-negotiable. A human sits at the control points that matter. That is the shape of the systems that reach production, and it has a name.
Agent pilots fail as pilots and succeed as orchestration, on a deterministic-workflow substrate carrying probabilistic AI, under one governed framework, with a human in the loop where it counts.
Nintex calls the answer Agentic Business Orchestration, or ABO. It holds structured workflow executions, AI-driven decision-making and human control together in a single framework. Unveiled on 24 September 2025, with its capabilities announced on 17 March 2026, it is a deliberate refusal of the binary that has wrecked so many programmes, the belief that a system must be either fully agentic or fully hard-coded.
Niranjan Vijayaragavan, then chief product and technology officer, Nintex, 17 March 2026
The rest of this report is the case for that sentence, built out for a reader who signs cheques. Part I sets the failure data on the table, because a board that has not seen the numbers will underrate the risk of doing nothing well. Part II explains the hybrid and draws it, so the mechanism is legible rather than asserted. Part III walks the ladder from pilot to enterprise and names the substrate that carries the weight. Part IV does the economics, taking in return and proof as well as the buy-versus-build question that quietly decides most of these programmes. Part V is the operator's read, with a checklist, our role and the annex.
Two of those figures deserve a moment. Buying from a specialist succeeds about twice as often as building in-house, a two-to-one advantage that survives most attempts to argue it away. And of the crowd of vendors selling the word "agentic", Gartner counts only around 130 that mean it. The market is loud, and the signal is thin. Choosing well is most of the work.
Every chart and table in this report carries its own data and its own source line. The prose makes the argument; the exhibits carry the evidence. Where a number is an estimate rather than a measured result, we say so on the exhibit itself.
The failure data is now abundant, consistent, and dull in its repetition. Boards underrate it because each study reads like a headline rather than a pattern. Set side by side, the studies describe one problem: the demo is cheap and the production system is expensive, and most organisations stall in the gap between them.
The case against doing nothing well is quantitative, and it is grimly consistent. Four independent bodies measured the same thing, the share of AI ambition that never reaches production, and arrived at figures that rhyme. We put them on one axis below, and note where each is measuring pilots, projects, or companies, because the denominators differ.
The Gartner line is the sharpest, because it is about agentic projects specifically and it is a commitment about the near future, not a description of the past. Over 40% of agentic-AI
projects will be cancelled by the end of 2027, the firm said. The causes it named were escalating cost, unclear value and inadequate risk controls. Those are governance failures, not model failures.
How to read it. The bars measure different denominators, organisations, projects, companies and survey respondents, so they are not additive. The pattern is what matters. The further along the production pipeline you measure, the smaller the surviving share. The one green bar, Deloitte's 25%, is the good news, and it is a minority. Sources: MIT NANDA, The GenAI Divide, July 2025; S&P Global 451 Research Market Insight, 13 January 2025; Gartner, 25 June 2025; Deloitte, 21 January 2026.
A single bad year could be a blip. A worsening trend is a signal. S&P Global's finding is not merely that 42% of companies abandoned most of their AI initiatives before production; it is that the figure rose from 17% the year before. Enthusiasm collided with delivery, and delivery won.
The abandonment rate roughly doubled in a year. The number that should worry a board most is not the level but the direction. Source: S&P Global 451 Research Market Insight, Alex Johnston, 13 January 2025 (n=1,006, fielded October to November 2024).
Deloitte's State of AI in the Enterprise 2026 (21 January 2026) offers the most flattering framing available, and it still lands hard. Only a quarter of its 3,000-plus executive respondents have moved at least 40% of their pilots into production. Put the other way, three-quarters have not, which means most organisations are running a lab, not a business line, and paying for both.
None of this argues against agents. It argues against a particular way of buying and building them, as isolated pilots, without governance, without integration into the systems that run the company, and without a deterministic floor to stand on. We turn to that diagnosis next.
A widely circulated claim that "89% of AI projects never scale" appears in several vendor decks. We cannot trace it to a primary source we trust, so we leave it out. The figures above are enough, and they are attributable.
Pilots stall for four reasons, and none of them is model quality. They stall on hype, on poor integration, on the absence of governance, and on the missing deterministic substrate. The first is a marketing problem; the other three are engineering and operating problems, and they are the ones that cost money.
Hype sets the trap. A demo that works on a curated example creates the impression that the hard part is done, when in fact the hard part has not started. The demo runs once, cleanly, in a sandbox. Production runs ten thousand times against live systems, messily and with auditors watching. The gap between those two conditions is exactly where the 95% of organisations with no measurable return are stranded.
An agent that cannot reach the ERP, the CRM, the document store, and the approval chain is an expensive chatbot with API access. Most pilots are built against mock data or a single system, then fail when asked to reach across the six systems a real process touches. Integration is not a feature you add later; it is the substance of the work, and pilots that defer it are deferring the moment of failure, not avoiding it.
A probabilistic system makes different choices on different runs. That is tolerable in a brainstorming tool and intolerable in a payments run, a claims decision, or a regulatory filing. Without controls, who approved this, on what data, under which policy, with what audit trail, a promising agent becomes an unmanageable liability the moment it touches anything that matters. Gartner's "inadequate risk controls" is this, precisely.
The deepest reason is architectural. Teams reach for a purely agentic design because it demos well. Then they discover that the exact parts of the process, the compliance checks and the mandatory sequencing of steps, cannot be left to a model that improvises. Lacking a deterministic floor to stand on, the whole system inherits the variance of its least reliable
component. This is the failure the hybrid is built to prevent, and it is the subject of Part II.
Hype creates the false start. Poor integration breaks the demo against live systems. Missing governance makes the result unusable where it matters. And the absent deterministic substrate makes the whole thing unreliable by construction. Only the first is cosmetic.
Every stall point has a corresponding fix, and the fixes are not exotic. They are the ordinary disciplines of running production software, applied to a new kind of component. The table maps the diagnosis to the remedy that Agentic Business Orchestration supplies.
| Stall point | Why the pilot dies | What production actually needs |
|---|---|---|
| Hype | A curated demo implies the hard work is finished when it has barely begun. | An honest maturity model that treats the demo as rung one of several. |
| Poor integration | Built against mock data or one system; breaks across the six a real process touches. | Native connectors and a workflow engine that reaches live systems of record. |
| No governance | Non-deterministic choices with no audit trail, unusable where compliance bites. | A governed, low-code environment with approvals, policy, and traceability. |
| No deterministic floor | The exact steps are left to a model that improvises, so variance spreads. | Conditional workflow control where precision and compliance are required. |
The remedies are the components of ABO. Nintex describes ABO as structured workflow executions plus AI-driven decision-making plus human control, one remedy per stall point. Source: Nintex, Agentic Business Orchestration, 24 September 2025.
Notice that three of the four remedies have nothing to do with the model and everything to do with the system around it. This is why the buy-versus-build gap is so wide. The model is the commodity, and the governed substrate that makes it safe to run is the scarce and expensive specialist part. Building that substrate in-house is where most internal programmes lose the two-to-one bet against them.
If the failures are governance and architecture failures, the fix is a governance and architecture fix, meaning a deterministic workflow substrate carrying AI decisioning, under one
framework, with a human at the controls. That is Agentic Business Orchestration, and it is where we go next.
The stall points from the last section resolve into five recurring failure modes. Each is a way a pilot dies, and each has a single answer inside Agentic Business Orchestration. The card set names the failure, states the stall in a line, then gives the answer.
The stall. The exact steps and their sequencing, including mandatory checks and compliance gates, are left to a model that improvises, so variance leaks into work that has to be exact.
The answer. Structured workflow executions run the mandatory steps the same way every time; the model is asked to judge, never to improvise the floor.
The stall. Built against mock data or a single system, the agent breaks when a real process reaches across the six systems of record it actually touches.
The answer. Nintex Automation CE reaches the ERP, the CRM, and the document store through native connectors, so the orchestration runs against live systems rather than a sandbox.
The stall. An agent allowed to act without approval gates, policy, or an audit trail turns into a liability the moment it touches a payment or a filing.
The answer. Agent Designer builds inside a governed, low-code environment, so approvals, policy, and lineage are inherited by default rather than added after the fact.
The stall. A pilot with no measured profit-and-loss line cannot defend itself when the budget is reviewed, which is where Gartner says the money is pulled.
The answer. The orchestration is scoped to one named process and measured against it on cycle time, exception rate and cost per case, so value is reported, not asserted.
The stall. An agent asked to reason over inconsistent, unlabelled, or unreachable data inherits the mess and repeats it at speed.
The answer. Nintex Forms captures structured input at intake, and the deterministic workflow validates and logs it before the AI lane reads anything.
Every autopsy we have read reduces to one of these five. A programme that can place its own failure on this card set can also find its remedy on it, which is the point of naming them.
When Gartner said over 40% of agentic-AI projects would be cancelled by the end of 2027, it named three reasons, which were escalating cost, unclear value and inadequate risk controls (25 June 2025). Those three are the five failure modes, grouped. The table maps each stated reason to the failure beneath it and the ABO control that retires it.
| Gartner's stated reason | The failure beneath it | The ABO control that answers it |
|---|---|---|
| Escalating cost | Integration rebuilt for every pilot; sprawl compounding the cost to run. | Connectors built once on Nintex Automation CE and reused; one framework consolidates the sprawl. |
| Unclear value | Pilots with no measured P&L line, unable to survive a budget review. | Process-scoped measurement against a named baseline, cycle time, exception rate, cost per case. |
| Inadequate risk controls | Ungoverned autonomy over a non-deterministic floor, with no audit trail. | Agent Designer governance, a deterministic workflow floor, human control points, end-to-end audit. |
Read the right-hand column as the specification for a production system. Gartner named the diseases, and the controls are the treatment. Reasons per Gartner, 25 June 2025; controls per Nintex ABO, 24 September 2025.
The mapping carries a quiet implication. None of the three reasons is a complaint about model quality, and none is answered by a better model. Each is answered by the system built around the model, its substrate, its governance and its measurement. This is why buying a proven platform beats building one, a point Part IV puts a number against.
Ask which of Gartner's three reasons your programme is most exposed to, then ask which control retires it. If the answer is "a better prompt", the programme is still on rung one.
The systems that reach production refuse the binary. They put AI where judgement adds value and conditional workflow control where precision and compliance are required, and they hold both under one governed framework with a human in the loop. Nintex calls it Agentic Business Orchestration. This part explains it and draws it.
One framework holds three things that most architectures keep apart, namely structured workflow executions, AI-driven decision-making and human control.
Agentic Business Orchestration is a hybrid, and the hybrid is the point. Nintex unveiled ABO on 24 September 2025 as the combination of structured workflow executions, AI-driven decision-making, and human control inside a single governed framework. The design principle is stated plainly. AI goes where judgement adds value, conditional workflow control where precision and compliance are critical.
The wager behind that principle is a refusal. Most failed programmes bet on one pole, either a fully agentic system that improvises everything, or a fully deterministic one that cannot adapt. ABO bets on neither, and puts a governed seam between them so that each does the work it is suited to.
Niranjan Vijayaragavan, then chief product and technology officer, Nintex, 17 March 2026
Structured workflow executions are the deterministic floor. They run the exact, repeatable and auditable steps the same way every time. That covers the sequencing, the mandatory checks and the compliance gates. This is the part a model must never be allowed to improvise.
AI-driven decision-making is where judgement lives. When a step requires reading unstructured input, weighing ambiguous evidence, or choosing among options, the AI decides, and hands its decision back to the workflow to execute. Judgement is used deliberately, not everywhere.
Human control is the third leg, not an afterthought. People sit at the control points where accountability must rest with a person, approving work and handling exceptions and escalations. The human is in the loop by design, at the places the design chooses.
An orchestra is many instruments under one score and one conductor. ABO is deterministic workflow, probabilistic AI, and human judgement under one governed framework. The value is not in any one player; it is in the coordination, which is exactly the capability isolated pilots lack.
The choice is usually framed as agentic versus deterministic, as if those were the only options. They are the two ways to fail. The table sets both against the hybrid, on the dimensions a production owner actually cares about.
| Dimension | Pure agentic | Pure deterministic | ABO hybrid |
|---|---|---|---|
| Handles ambiguity | Yes | No | Yes, via AI decisioning |
| Repeatable & exact | No: improvises | Yes | Yes, on the workflow floor |
| Auditable trail | Weak | Strong | Strong, governed end to end |
| Adapts to change | Yes | No: brittle | Yes, where change is expected |
| Compliance-safe | Risky | Yes | Yes, precision where it counts |
| Human accountability | Unclear | Rigid | Designed-in control points |
| Reaches production | Rarely | Often, but limited | Built for it |
The hybrid keeps the strengths of both columns and discards their failure modes. It is not a compromise; it is a division of labour. Framework per Nintex, Agentic Business Orchestration, 24 September 2025.
The pure-agentic column is where the demos come from and where the zero-return 95% collect. The pure-deterministic column is safe but cannot cope with the ambiguity that real processes throw at it. The hybrid column is the only one that answers "yes" down the whole list, and answering yes down the whole list is what shipping to production requires.
ABO also does something a board will value directly: it consolidates. Instead of a drawer of disconnected AI pilots, each with its own risk profile and its own integration debt, one framework carries them all. We return to that consolidation argument in Part III; here it is enough to note that the hybrid is the thing that makes consolidation possible.
A claims-style process, drawn as ABO runs it. The deterministic workflow carries the sequence; AI decisioning handles judgement; a human holds the control point. Governance wraps the whole run.
One governed framework wraps every lane. It carries identity and policy, approvals and data lineage, and a full audit trail across the whole run, so a probabilistic decision inside a deterministic process is always accountable to a person and a policy.
Read across, not down. The workflow lane never improvises; the AI lane never executes mandatory steps unchecked; the human lane owns the moments that require accountability. The three run as one orchestration. Illustrative process; framework per Nintex ABO, 24 September 2025.
The diagram makes the earlier abstraction concrete. Judgement sits in the AI lane and nowhere else; precision is guaranteed in the workflow lane; the human lane carries accountability. Nothing is left to improvisation that should not be, and nothing is hard-coded that needs to flex.
A supplier invoice that does not match its purchase order, the everyday exception that clogs accounts payable, run through the hybrid, so the division of labour is concrete rather than asserted.
An invoice arrives quoting a price the purchase order does not. A pure deterministic system kicks it to a human queue; a pure agentic system guesses. ABO does neither: it reads the mismatch, judges whether it sits inside tolerance, clears what it can, and pulls a person in only for what it should.
Every clearance, by the AI within tolerance or by the approver above it, is written to the same audit trail, against the policy and the person accountable for it.
Illustrative process. The three-way match and the ERP post never improvise; the AI never posts an unmatched invoice on its own; the approver owns only the exceptions that need a person. Framework per Nintex ABO, 24 September 2025.
Where control holds. The three-way match is deterministic and stays that way. An invoice clears only when the purchase
order, the goods receipt, and the invoice agree within a policy the workflow enforces identically on every run. The posting to the ERP and the audit record are equally fixed. None of this is a matter of judgement, so none of it is handed to the model.
Where AI decides. Judgement begins where the numbers disagree. The AI lane reads the invoice, classifies the mismatch as price, quantity, or tax, and weighs it against the tolerance and the supplier's history. A £40 overcharge from a supplier with a clean record and a standing tolerance is a different object from the same overcharge on a first invoice from an unknown vendor. The AI proposes a decision and a confidence, then hands both back to the workflow.
Where a person is pulled in. The approver sees only what should reach a person: the out-of-tolerance cases, the novel suppliers, the recommendations the AI itself flags as low-confidence. The approver approves, adjusts, or rejects, and the reason is logged. Everything the AI cleared within tolerance never troubles a human at all, which is the entire economic point.
| Step | Who acts | Kind of work | Control |
|---|---|---|---|
| Receive & match | Workflow | Deterministic | Same rule every run |
| Read & classify mismatch | AI lane | Judgement | Confidence returned |
| Assess against tolerance | AI lane | Judgement | Policy-bounded |
| Clear within tolerance | Workflow | Deterministic | Logged, auditable |
| Out-of-tolerance case | AP approver | Human control | Reason recorded |
| Post & close | Workflow | Deterministic | Written to ERP + audit |
Two of six steps are judgement; one is a person; the rest are exact. The proportions are illustrative, but the shape is the point: judgement is used narrowly, precision holds the floor, and the human is spent only where accountability requires it.
Swap invoice-exception for claims triage, customer onboarding, or contract intake and the lanes do not move. The deterministic floor carries the mandatory steps, the AI lane carries the ambiguous ones, and the person carries the accountable ones. The process changes; the division of labour does not.
A framework is only as good as the environment in which people build on it. Two capabilities carry ABO from principle
to practice. Both were announced on 17 March 2026 for selected customers, moved to open beta in the Q2 2026 update of 3 June 2026, and are described by Nintex as shipping parts of the CE Platform from the Q3 2026 release of 29 September 2026.
Agent Designer is the governed, low-code environment required for mission-critical operations. "Governed" and "low-code" are doing equal work in that phrase. Low-code means the people who understand the process can build the agent without a platform team as a bottleneck. Governed means what they build inherits policy, approvals, and audit by default, rather than bolting them on later, which, as Part I showed, is where pilots fail.
Nintex Orchestration provides modular, phase-based workflows. A process is broken into phases that can move forward, jump back, or repeat, the control flow that real operations need, rather than the straight line that only demos require. Nintex described it as in open beta through mid-2026 and now presents it in the present tense as part of the CE Platform, so we no longer call it a beta.
↶ jump back to an earlier phase when new evidence arrives · ↻ repeat a phase until a condition is met
Move forward, then jump back and repeat. Real processes loop and revisit; a phase-based model expresses that directly. Source: Nintex Orchestration, announced 17 March 2026, in open beta from 3 June 2026, described as shipping in the Nintex CE Platform release of 29 September 2026.
Together, Agent Designer and Orchestration are the reason ABO is a platform rather than a slogan. One gives builders a governed place to work; the other gives processes a shape that matches how they actually run. Neither is exotic. Both are the ordinary machinery of production software, which is exactly what pilots were missing.
Sources. Nintex Agentic Business Orchestration unveiled 24 Sep 2025, capabilities announced 17 Mar 2026; Q2-2026 update 3 Jun 2026; Nintex CE Platform consolidation 29 Sep 2026, by which point Agent Designer and Nintex Orchestration are described as shipping.
Of the three reasons Gartner gave for cancelled agentic projects, the one an operator can do least about after the fact is inadequate risk controls (25 June 2025). Cost can be renegotiated and value can be re-measured, but a system built without controls cannot have them retrofitted once it is touching money. Controls are an architecture decision, taken at the start or not at all.
ABO takes that decision at the start through Nintex Agent Designer, the governed, low-code environment required for mission-critical operations. Three controls do most of the work, and each maps to a question an auditor will actually ask.
Every decision in a run is recorded. The record holds what was decided, on what data, under which policy, and by the AI or by a person. A probabilistic choice inside a deterministic process is only acceptable if it leaves a trail, and the trail is written by default rather than reconstructed later. It answers one question. Show me why this happened.
Defined points in the flow require a sign-off before execution continues. A gate is not a suggestion the workflow may skip under load; it is a stop the deterministic floor enforces. Gates sit where policy or value demands a person, and nowhere else, so they add friction only where friction is the point. It answers one question. Who agreed to this before it ran.
Accountability rests with a named person at the control points, where work is approved and exceptions and escalations are handled. The human is placed by the design, not summoned by a failure, and the action is logged as the accountable act. It answers one question. Who owns the outcome.
Low-code lets the people who understand the process build the agent. Governed means what they build inherits audit, gates, and human control by default. Take either word away and Agent Designer becomes the thing that failed in Part I.
Controls are worth listing precisely, because a board signs against a list, not a principle. The table sets each control beside what it does and the evidence it leaves, the record an auditor, a regulator, or a post-incident review will ask to see.
| Control | What it does | The evidence it leaves |
|---|---|---|
| Identity & policy | Binds every action to an authenticated actor and the policy in force. | Who acted, under which rule. |
| Approval gate | Halts execution until a required sign-off is given. | The approver, the timestamp, the decision. |
| Human control point | Routes exceptions and escalations to a named owner. | The person, the override, the reason. |
| Data lineage | Traces the inputs a decision was made on. | The source records, versioned. |
| Audit trail | Writes the whole run, AI and human alike, to one record. | The end-to-end account of what happened. |
The right-hand column is what survives the meeting. Controls that leave no evidence are not controls. Governance per Nintex Agent Designer; framing against Gartner's “inadequate risk controls”, 25 June 2025.
: The governance principle behind ABO
The controls are deliberately ordinary. They are the disciplines any regulated operation already applies to its people, extended to a new kind of actor. That is the reassuring part. Making an agent auditable does not require a new theory of control, only that the control be built in rather than promised.
Three claims carry into Part III. First, the production-grade architecture is a hybrid, not a pure design. Second, the hybrid needs a governed, low-code build environment, Agent Designer, or it recreates the governance gap that killed the pilots. Third, real processes need phase-based control flow, which Nintex Orchestration supplies.
The ABO design principle, Nintex, 24 September 2025
The architecture is settled. The remaining question is the one boards actually lose sleep over: can it scale from a single working process to the enterprise, without the whole thing turning back into a pile of pilots? That is Part III, and it depends on a substrate strong enough to build and run on.
A governed hybrid solves the architecture problem. Scaling it solves the operating problem, and scaling needs a build-and-run substrate, a maturity path, and a way to consolidate rather than multiply the pilots. We turn to all three next.
An audit trail that satisfies an architect and an audit trail that satisfies a supervisor are different documents, and the gap is usually discovered during an examination. A supervisor is not checking that the system worked. The question is whether the firm can account for a particular decision taken on a particular day, and account for it without the people who built the system.
Seven fields answer that question for a single automated decision. They are unremarkable individually, and the absence of any one of them is what turns a reconstruction into an estimate.
Field five is the one most often missing. A confidence value returned by a model and discarded by the workflow removes the only evidence that the routing decision was correct on the information available. Keep it, and a low-confidence case that went wrong is a defensible outcome. Discard it, and the same case is an unexplained one.
Field one is the sharpest in practice. A process redesigned twice in a year generates decisions under three designs, and a trail that does not record which design was live cannot explain why two apparently identical cases were handled differently. Version the process definition, and bind the version to the run.
None of this is exotic, and that is the reassuring part. Making an automated decision accountable does not require a new theory of control. It requires that the record be written as the run happens, because a trail reconstructed afterwards is an argument rather than evidence.
Scale is not one big leap; it is a ladder climbed one rung at a time, on a substrate strong enough to carry the weight. This part sets out the maturity model, names the build-and-run engine, Nintex Automation CE, and shows how the enhancers compose, and how one framework consolidates the sprawl that pilots create.
Deloitte found only a quarter of its respondents have moved 40% or more of their pilots into production. The other three-quarters are stalled on a rung. The ladder makes the rungs explicit, so a programme knows where it is and what the next step costs.
The gap that matters is rung 2 to rung 3. Integration and governance are the wall most programmes hit. ABO is designed to carry a process across it rather than leaving it stranded. Rung-4 threshold per Deloitte, State of AI in the Enterprise 2026, 21 Jan 2026.
We built these five rungs as a planning heuristic, not a validated maturity model. Use them to locate a programme honestly, not as an industry benchmark, and not as a number to quote back to us. The Deloitte and Gartner figures around them are independent; the rungs themselves are Volge's framing.
Climbing is not automatic. Each rung adds a discipline the last one lacked, and each has an owner. The table is a planning aid: it tells a programme what to add next and who is accountable for adding it.
| Rung | State | What you add to climb | Owner |
|---|---|---|---|
| 1 | Demo | Real, representative data instead of a curated example. | Process team |
| 2 | Contained pilot | Live connectors to systems of record; the integration work. | Platform engineering |
| 3 | Integrated pilot | Governance: approvals, policy, audit, human control points. | Risk and compliance |
| 4 | Production at scale | Monitoring and P&L measurement; run-book and support. | Operations |
| 5 | Orchestrated enterprise | Shared framework across processes; consolidation of pilots. | COO / CIO |
The wall sits between rungs two and three, which is where the failure data clusters. Integration and governance are the two disciplines a pilot can be built without and a production system cannot.
The ladder also reframes the buy-versus-build question. Every rung above the demo is substrate work. Connectors, governance and monitoring are exactly what specialist platforms have already built and internal teams have to build from scratch. The 67% versus 33% success gap between buying and building, which we detail in Part IV, is this ladder measured in outcomes.
If your programme cannot name which rung each process is on, it is on rung one. The value of an explicit ladder is that it makes "we have a pilot" a precise statement instead of a comforting one.
Every orchestration needs an engine to build on and run on, and for ABO that engine is Nintex Automation CE, the build-and-run substrate. It is the part that reaches the systems of record, executes the deterministic workflow, and gives the AI and human lanes somewhere to plug in. Nintex publishes no list price for it, so a business case has to be built from a quotation rather than from a published floor.
The substrate matters because it is precisely what internal builds underestimate. A team can stand up an impressive agent in a fortnight; standing up the governed, integrated, always-on engine beneath it is the work of many quarters, and it is the work Nintex Automation CE removes.
On top of the substrate sit capabilities that agents invoke as building blocks, rather than features to be admired in isolation. An orchestration composes them into a process; it does not treat each as a separate product. Four are worth naming.
Named, not over-featured. These are ingredients an orchestration calls, not destinations. The value is in the composition, which the substrate and Agent Designer make governable.
| Capability | What an agent uses it for |
|---|---|
| Nintex RPA | Task automation across applications. Two lines run today: Nintex RPA (formerly Kryon, release 26.6, July 2026) is the strategic go-forward product; Nintex Foxtrot RPA (formerly Nintex RPA Foxtrot and Nintex RPA LE) continues alongside it. |
| Nintex DocGen | Document generation from process data, including a DocGen Agentforce connector for Salesforce, so generation composes into a CRM-driven flow. |
| Nintex Forms | Structured data capture at the front of a process, feeding clean input into the workflow and the AI lane. |
| Nintex eSign | Electronic signature at the control points where a signature is the accountable act. |
One way to read the whole platform is as a stack. The substrate runs and reaches; the enhancers supply capability; Agent Designer governs the build; ABO orchestrates the three lanes into a process. The diagram stacks them in that order.
Read bottom-up. Nothing floats. The substrate carries the enhancers; Agent Designer governs how they are assembled; ABO orchestrates the assembled process. The isolated pilot is what you get when only the top layer exists.
Then-Nintex CEO Amit Mathradas framed the ambition on 24 September 2025 as orchestrating “across people, systems, and AI agents to deliver outcomes at scale.” (Nintex announced his move to Five9 on 17 December 2025, effective 2 February 2026; board chair Stephen Elop has since been named chief executive.) The stack is that sentence rendered as architecture.
An orchestration is only as capable as the parts it can call. Four Nintex capabilities sit on the substrate as building blocks, invoked by a flow rather than admired as products. Named, not over-featured.
Nintex RPA (formerly Kryon) is the strategic go-forward line, current release 26.6, announced on 13 July 2026. It gives an agent hands: attended and unattended automation across applications that expose no API, so screen-level work can be delegated rather than left undone.
Nintex Foxtrot RPA, formerly Nintex RPA Foxtrot and Nintex RPA LE, is maintained alongside it, with no published end-of-life date. Existing desktop automations keep running and stay supported; new work belongs on the go-forward line, not here.
Turns process data into a finished document, a quote, a contract, a letter, without a person retyping what the system already holds. The DocGen Agentforce connector composes generation into a Salesforce-driven flow, so a decision taken in the CRM produces the paperwork that follows it.
Each capability is something an orchestration invokes at the right step, not a destination a user visits. The value is in the composition, which the substrate runs and Agent Designer governs.
Captures structured, validated input at the front of a process, feeding clean data into both the workflow and the AI lane. This is where the “data not ready” failure is retired, at intake, before a decision is ever asked for.
Adds a signature at the control point where a signature is the accountable act. The human’s approval becomes a signed, logged record, closing the audit trail the governance layer keeps.
Read left to right, a customer-onboarding flow calls each in turn. Forms captures the application, RPA pulls a record from a system with no API, the AI lane judges the risk, DocGen produces the agreement, and eSign closes it, with the workflow sequencing the lot and a person approving where policy requires.
Illustrative flow. The capabilities are ingredients, not steps a user runs by hand; the orchestration calls them and governs the call. Capabilities per Nintex; composition per ABO, 24 September 2025.
Sprawl is the tax that pilots levy after the fact. Each one arrives with its own tools, its own risk profile, its own integration debt, and its own quiet maintenance cost. Nintex positions ABO explicitly against "software sprawl and the risks of disconnected AI pilots", so consolidation is a selling point rather than a side effect.
| Dimension | A drawer of disconnected pilots | One governed framework |
|---|---|---|
| Governance | Per pilot, inconsistent, often absent | Shared, applied by default |
| Integration | Rebuilt each time; duplicated debt | Built once on the substrate, reused |
| Risk profile | Many, each unmanaged | One, monitored centrally |
| Vendor count | Many, overlapping | Consolidated on one platform |
| Cost to run | Compounds quietly with each pilot | Amortised across the framework |
| Path to scale | Each pilot restarts the climb | Rungs already built, reused |
Sprawl is not free; it is deferred cost. Positioning per Nintex: ABO consolidates against software sprawl and the risks of disconnected AI pilots, 24 September 2025.
The financial logic is simple. Ten pilots on ten stacks pay for governance ten times and still have less of it than one framework provides once. The consolidation argument is therefore also a cost argument, and a risk argument, and a speed argument, which is why it belongs in the board pack rather than the architecture review.
Every rung of the ladder built inside one framework is a rung the next process inherits for free. Sprawl throws that reuse away. Consolidation banks it. That compounding is the difference between a programme that accelerates and one that merely repeats.
A platform that works gets used, and use without a release route produces the condition this chapter warns against from the opposite direction. Forty workflows built by nine teams to four conventions is not sprawl across vendors. It is sprawl inside one framework, and it is harder to see because the consolidation argument has already been won.
Three disciplines prevent it, and all three cost less at the first workflow than at the fortieth. A naming and structure convention, written down and enforced in review, so that a workflow can be read by someone who did not build it. A promotion path from development to test to production, with the same artefact moving through each rather than being rebuilt. And a reusable component library, so that the connector written for the first process is called by the second rather than written again.
The third is where the consolidation dividend is actually banked. A connector built once and reused is the whole financial argument of this chapter, expressed as an engineering practice. Built nine times by nine teams, the saving never arrives, and the platform bill arrives anyway.
Every scaled automation programme ends up with a small central function, and the ones that plan for it fare better than the ones that discover it. The function is not a build team. It owns the conventions, the component library, the promotion path and the platform's own health, and it reviews what the business builds rather than building it.
Source: Volge analysis, 2026. The split is a working division of responsibility rather than an organisational design.
The useful part of that table is the right-hand column. Most programmes fail at scale because the central function starts building, which makes it a bottleneck, or because the
| Function | What it owns | What it does not own |
|---|---|---|
| Platform operations | Availability, connector credentials, releases, version currency | The processes themselves |
| Standards and review | Conventions, the component library, design review before promotion | Approval of the business case |
| Process ownership | The business outcome, the baseline, the control points | The platform |
| Risk and compliance | The control set, the audit trail requirement, the approval authority matrix | The build sequence |
business starts promoting its own workflows, which makes the conventions optional. Either failure is visible within two quarters.
A production system earns its keep or it does not survive the next budget round. This part sets out the return, customer-proven and analyst-modelled, the buy-versus-build evidence that favours a proven platform, and the market context, with every estimate labelled as an estimate.
Return on a production orchestration is measurable, and it has been measured. Forrester's Total Economic Impact study of Nintex Workflow, commissioned by Nintex and conducted in late 2022, modelled a 213% return with a nine-month payback for a composite organisation. The return comes from the same place in every case, with a slow manual base replaced by a governed automated one.
A single modelled study is still a model, so its limits belong next to it. The study was commissioned by Nintex, it covers Nintex Workflow rather than the agentic capabilities this report describes, and it runs on a composite organisation rather than a named client. The vintage is worth labelling, because these studies age. Read it as the generalisable middle of the range, not as a measurement of your own programme.
These are platform returns, established before ABO existed, which makes them a floor for the argument rather than a ceiling. The orchestration layer is meant to extend that record into processes that pure workflow could not previously reach, the ones needing judgement as well as precision.
| Lens | Return | Payback / period | Vintage |
|---|---|---|---|
| Forrester TEI of Nintex Workflow (modelled) | 213% ROI | 9-month payback | Conducted late 2022 |
| Forrester Wave, Agentic Process Orchestration Platforms | Strong Performer | A placement, not leadership | Q3 2026 |
A modelled return and an analyst placement, side by side. Neither is ABO-specific. The Total Economic Impact study covers Nintex Workflow, and the Wave rates agentic process orchestration platforms as a class. Both are dated accordingly.
MIT NANDA's The GenAI Divide measured the choice that quietly determines success. Generative-AI tools sourced from external partners reached deployment about 67% of the time, against about 33% for tools built internally (July 2025). The paper's own table reads 66% and heads the column “% of Deployments”, so this is a share of deployments rather than a success rate, and the paper states that the correlation does not prove causation. The gap is still a doubling of the odds, and it points at a proven platform.
Bought tools reached deployment roughly twice as often as built ones. The substrate, governance, and integration that a specialist has already built are exactly what an internal team must build from scratch. Source: MIT NANDA, The GenAI Divide, July 2025. The measure covers generative-AI tools generally, not agentic capability specifically.
The result is intuitive once the ladder from Part III is in view. Buying starts you several rungs up, with the substrate, the governance and the connectors already in place. Building starts you at the bottom, paying to construct the very machinery whose absence killed the pilots. The 67/33 split is the ladder, measured in outcomes.
Two caveats, because we prefer them to slogans. First, the 67/33 split is one paper, published under the MIT NANDA masthead in July 2025, which disclaims institutional attribution to MIT and drew methodological criticism; read it as directional, not decisive. Second, we sell the "buy" side, so weigh our fondness for a figure that favours buying accordingly. And buying only wins if the platform is genuinely agentic. Gartner
counts about 130 of those among the thousands claiming the label. The advantage is contingent on choosing well, which Part V describes.
Build, and you take a one-in-three bet while paying to construct the substrate. Buy a proven platform, and you take a two-in-three bet with the substrate already built. For a system that must run the business, the arithmetic is not close.
Proof of a platform is not only its return; it is its footprint and its reputation among the people who run it. Nintex's own About page says more than 7,000 organisations across over 100 countries run on it (Nintex, September 2026), though its newsroom boilerplate dropped the count on 22 October 2026 in favour of “thousands”. User ratings are solid rather than inflated. G2 puts it at 4.2 out of 5 across 1,326 reviews, TrustRadius at 8.0 out of 10 (both read 6 November 2026). We report those without dressing them up.
The context is a market growing fast enough to make inaction its own risk. The agentic-AI market is estimated at about $5.2 billion in 2024, and projected to reach roughly $196.6 billion by 2034, a compound annual growth rate of about 43.8%. That figure is a Market.us estimate, and we label it as an estimate rather than as a Gartner or Forrester number, because it is one.
~43.8% CAGR, an estimate, not an analyst-house forecast. Directionally, the market is large and compounding; precisely, treat the endpoints as a modelled estimate. Source: Market.us estimate.
We make no claim that Nintex holds a Gartner Magic Quadrant or Forrester Wave leadership position; those are specific designations and we will not assert them without the citation to match. Forrester did name Nintex a Strong Performer in The Forrester Wave: Agentic Process Orchestration Platforms, Q3 2026, which is a placement rather than leadership. The case here rests on the architecture, the modelled return, the buy-versus-build evidence and the footprint, which is enough.
No board should buy on the customer headline alone. We frame the return as a bear case, a base case and a bull case, anchored on the sourced figures, so the case survives scrutiny rather than flattering it.
The honest planning number is the base case. The bull case is real but conditional; the bear case is survivable because the return stays positive even when execution disappoints. A business case built on the base, aware of the bull, and hedged against the bear is one a board can sign without embarrassment.
Return has two sides and the business cases that fail in year two are the ones that modelled only the first. Five cost lines carry a production orchestration, and three of them do not appear in a platform quotation.
The last line is the regional one, and it is specific to the ground chapter five describes. A Gulf or Indian programme running the core workflow cloud is running it offshore, which is a legal and documentary cost rather than a technical one. Where classification rules out the multi-tenant cloud, the alternative is a self-hosted engine, and that carries infrastructure, patching and availability costs the cloud subscription had absorbed.
Model both routes before selecting one. A business case built on cloud economics and delivered on self-hosted infrastructure is the most common way a regional automation programme misses its number, and the divergence is visible at design time rather than at the first invoice.
One discipline keeps the whole case honest. Measure the baseline before the build, in system data, and report the result against it on the same definitions. Everything else in this chapter is an estimate. That comparison is a measurement, and it is the only figure a second budget round will accept.
Nintex Automation Cloud has no Middle East region and no India region. Nintex does run a United Arab Emirates instance, and it serves process mapping and analytics only. So the process map can sit in the Gulf and the workflow that acts on it cannot, which is the single most consequential regional fact for a programme of this kind. Oman requires classification before migration and an authorisation before contracting. Kuwait writes exit and portability into law. India places the liability for an AI system's output on the firm that deploys it.
Every Gulf and Indian customer of the core Nintex workflow platform runs offshore, and the vendor's own operational register says so. Nintex's status page, read on 8 November 2026, lists Nintex Automation Cloud in seven regions: Australia, Europe, the United States, Canada, the United Kingdom, Singapore and US Government. There is no Middle East region. There is no India region. The nearest region to India is Singapore, and the nearest to the Gulf is Europe.
The same register records one Gulf exception, and it is precise. Nintex operates a United Arab Emirates instance for two services only, Nintex Process Manager production and Nintex Analytics. That is the process-mapping and process-intelligence half of the portfolio. The workflow-execution half has no in-region home anywhere in the Gulf or India.
Read those two facts together and the architectural consequence is unavoidable. The documented process can live in the United Arab Emirates. The workflow that reads that process, reaches the systems of record and acts on live data cannot. A programme whose residency story rests on the UAE instance has described its map and said nothing about its engine.
Volge: on the Nintex regional footprint, read 8 November 2026
One further distinction has to be kept. The regional footprint above is operational fact drawn from a live status register, not a contractual commitment. Nintex's trust centre is a gated portal that returned no text, and no residency article could be located in its product documentation. So no contractual data
residency commitment by Nintex is asserted here. Where a residency obligation is binding, the commitment has to be obtained in the contract, by name, and not inferred from an operations page.
Nintex ships three delivery models, and its own copy draws the sovereignty line. Nintex Automation CE is described by Nintex as a unified multi-tenant cloud platform. Nintex Automation K2 is self-hosted, and Nintex's own footer describes it as delivering "full control and data sovereignty". Nintex for Salesforce runs inside the customer's Salesforce org. That is a vendor characterisation and is quoted as one, but it is the vendor's architecture statement rather than a customer claim, and it is reliable at that level.
Set the vendor's position against Oman's procurement chain and the two line up unusually well. The Cloud Computing First Circular, Ministry of Transport, Communications and Information Technology Summary 10/2026 of 28 June 2026, requires government entities to classify data before storing or processing it in a cloud environment, and requires providers to hold approved certifications before contracting. The Cloud First Policy published on 1 July 2026 requires contracting only with providers holding a valid authorisation issued by the Telecommunications Regulatory Authority, and binds government bodies excluding security and military agencies.
An Omani government body that must classify before it migrates, and contract only with an authorised provider, has a plausible route to the self-hosted engine and no plausible route to the multi-tenant cloud. That is not a criticism of either product. It is a sourcing conclusion that follows from two published instruments and one vendor architecture page, and it should be reached before a platform is selected rather than during a procurement review.
Oman has a third instrument that bears directly on orchestration. Ministerial Circular No. 16/2026 on the Electronic Integration Policy, issued 7 September 2026, together with the General Electronic Integration Policy, governs interoperability and application-to-application integration across government units. A workflow platform is an integration layer. In Oman it is an integration layer with a policy attached.
Nintex builds government editions when a market compels it, and it has built one. Nintex Workflow Government Edition is documented in the vendor's own help centre, alongside a FedRAMP edition of DocGen for Salesforce. There is no Gulf government edition and no Indian one. Nintex RPA
runs in the United States only, as does Nintex Licensing and Permitting.
The American edition is the useful part of that finding. It establishes that the vendor will build a jurisdictional edition where procurement rules demand it, which makes the absence of a Gulf edition a statement about demand rather than about capability. For a buyer, it is also the right question to put in a procurement dialogue. Ask what it would take, and ask for the answer in writing.
Kuwait is the one jurisdiction in this set that legislates portability, and the clauses read as though drafted with an orchestration platform in view. Chapter 6 of the Communication and Information Technology Regulatory Authority's Cloud Computing Regulatory Framework places data ownership with the subscriber at 6.4.9, limits use to service provision at 6.4.10, permits sub-contracting with ultimate responsibility resting on the provider at 6.4.14, and at 6.4.17 requires the provider to return content in a customary format or migrate it directly to another provider.
Two clauses go further than anything else in the region. Clauses 6.4.17.5 and 6.4.17.6 prohibit a provider from withholding government data or obstructing migration on grounds of unsettled payments or an ongoing dispute, and guarantee service continuity for at least six months from the onset of a dispute. Chapter 7 voids unfair limitation-of-liability terms.
For a workflow programme the relevance is immediate. An orchestration holds process definitions, in-flight cases and the audit trail of everything the platform has done. Those are the three assets a dispute would strand. A Kuwaiti public body can cite a clause. Everyone else has to negotiate one, and the Kuwaiti text is the best available template for doing so.
The hosting side of the Kuwaiti framework is where the platform choice is actually decided. Clause 2.1 permits only a party licensed by the Authority, with a data centre and infrastructure inside Kuwait, to contract to provide cloud services to the public sector, and the licence is not available to foreign operators directly. The framework also prohibits hosting Level 3 and Level 4 data outside Kuwait, along with individuals' personal data held by government bodies, private-sector companies or service providers, and requires Level 3 data to sit in a private or hybrid cloud under clauses 4.2.1.1 and 3.1.4.3 of the Arabic primary text.
Dubai adds a second gate specific to this book's subject. The Dubai AI Seal operates as an assurance mark for artificial intelligence procured into the emirate's public sector, alongside the Dubai AI Security Policy. An orchestration platform
marketed on its agentic capability meets both a cloud gate and an AI gate in Dubai, and they are assessed separately.
Four accreditation gates in the region bind the supplier rather than the workload, and an automation programme that ignores them stalls at procurement. Saudi Arabia requires registration with the Communications, Space and Technology Commission, at Class A for public government data, Class B for confidential and Class C for secret or top secret. In Dubai, the Dubai Electronic Security Center's Cloud Service Provider Security Standard is mandatory, in the Center's own words, for all providers wishing to offer cloud services to Dubai government and semi-government entities, with annual surveillance audits and recertification every three years. Kuwait requires the Authority's cloud licence. Oman requires a Telecommunications Regulatory Authority authorisation.
| Jurisdiction | Nintex workflow cloud in region | Nintex process mapping in region | The supplier gate |
|---|---|---|---|
| Saudi Arabia | No | No | CST registration at Class A, B or C by data classification |
| United Arab Emirates | No | Yes, Process Manager production and Analytics | DESC Cloud Service Provider Security Standard for Dubai government, plus the Dubai AI Seal for AI procurement |
| Qatar | No | No | Not established from a primary source |
| Kuwait | No | No | CITRA cloud licence, requiring a data centre inside Kuwait; not available to foreign operators directly |
| Oman | No | No | TRA authorisation before contracting, plus classification before migration |
| Bahrain | No | No | No national accreditation scheme; international certifications specified instead |
| India | No; nearest region is Singapore | No | MeitY empanelment and STQC audit status for SEBI regulated entities under the 2023 cloud circular |
Sources: Nintex Services status register, read 8 November 2026, for regional availability. Legal and procurement instruments as cited in the text, each with its issuing body and date. Compilation by Volge, 2026.
That provision is the regulatory form of this book's architecture argument. If the deploying firm owns the output
whatever the source of the model, then the human control point is not a governance nicety. It is where the firm's own liability is discharged, and the audit trail is the evidence that it was. A design with no named owner at the control point has left the liability in place and removed the record of it.
India holds the only provision in these jurisdictions that allocates liability for an artificial intelligence system's output, and it is the most important regional fact in this book after the absence of a region. The Securities and Exchange Board of India (Intermediaries) (Amendment) Regulations, 2025, notification SEBI/LAD-NRO/GN/2025/226, inserted Chapter IIIB and regulation 16C on responsibility for the use of artificial intelligence tools. A regulated person using such tools, whether designed in-house or procured from a third-party technology service provider, is solely responsible for the privacy, security and integrity of investor and stakeholder data, for the output arising from the use of such tools, and for compliance with applicable law, "irrespective of the scale and scenario of adoption". Sources give 6 February 2025 and 10 February 2025 for gazette publication, and the discrepancy could not be resolved from a primary page.
The Reserve Bank of India's companion position arrives through a committee report rather than a direction. The Report of the Committee on a Framework for Responsible and Ethical Enablement of Artificial Intelligence in the Financial Sector, chaired by Dr Pushpak Bhattacharyya and issued on 13 August 2025, sets out seven principles and twenty-six recommendations, and proposes expanding seven existing master directions to cover artificial intelligence. Nothing in it binds a regulated entity, and no implementing direction has been issued. Cite it as a committee report, because that is what it is.
The clearest Indian rule for a workflow platform is not an AI rule at all. The Reserve Bank of India's Information Technology Governance Directions of 2023 required, at section 17, that for critical applications there be no manual intervention or modification in data transferred between processes or applications, and that the transfer mechanism be properly tested, securely automated with necessary checks and balances, and properly integrated through what the Direction calls Straight Through Processing methodology, "with appropriate authentication mechanism and audit trails". Those provisions are carried into the Reserve Bank of India's consolidated Directions for commercial banks of 2026, at Chapter V
section S, paragraphs 141 and 142, following the consolidation of 31 July 2026.
That is a workflow connector described in prudential language. An orchestration that moves a payment instruction between two systems of record is a data transfer between applications, and the Direction requires it to authenticate and to leave a trail. Most platform evaluations treat connector authentication as a technical detail. In an Indian bank it is a supervisory requirement with a paragraph number.
Three further Indian obligations shape the operating design. The Master Direction on Outsourcing of Information Technology Services, RBI/2023-24/102 of 10 April 2023 and effective 1 October 2023, requires at clause 16(7) storage of data only in India where regulatory requirements apply, and at clause 17(8) that a cyber incident be reported to the Reserve Bank within six hours of detection by the third-party service provider. Chapter X prohibits a provider from erasing, purging, revoking, altering or changing any data during an exit transition unless advised by the regulator or the regulated entity. Chapter IX adds four cross-border requirements, including monitoring of country risk and a contractual right for the Bank to inspect a foreign-based provider.
CERT-In's Directions of 28 April 2022, No. 20(3)/2022-CERT-In, are in force and unamended, and they reach every service provider, intermediary, data centre, body corporate and government organisation. Direction (ii) requires reporting of listed incidents within six hours of noticing them. Direction (iv) requires logs of all information and communications technology systems to be enabled and maintained securely for a rolling period of 180 days, "maintained within the Indian jurisdiction". CERT-In's own frequently asked questions of May 2022 say at question 35 that logs may be stored outside India provided the obligation to produce them is met in reasonable time. The binding text says India. The clarification says anywhere, if production is timely. Cite both, decide deliberately, and record the decision.
For an orchestration running in Singapore on behalf of an Indian entity, that is the operative tension. The audit trail this book has argued for throughout is a log under direction (iv). Where it sits, and how quickly it can be produced, is a question to settle in the design rather than during an incident.
India's data-protection rules attach to classes of data rather than to tenants, and most of the commercially consequential provisions are not yet live. The Digital Personal Data Protection Act, 2023 is phasing in. The Digital Personal Data Protection Rules, 2025, notified on 13 November 2025 as G.S.R. 846(E), commence in tranches: rules 1, 2 and 17 to 21 on
notification, rule 4 on consent managers from 13 November 2026, and rules 3, 5 to 16, 22 and 23 from 13 May 2027. Until then section 43A of the Information Technology Act, 2000 and the rules on sensitive personal data remain the operative Indian data security provisions.
Two of the coming rules bear on an orchestration. Rule 6(1) requires encryption, obfuscation, masking or virtual tokens, appropriate access control for computer resources, and visibility of access through logs, monitoring and review. Rules 6(1)(e) and 8(3) impose a minimum one-year retention of personal data, associated traffic data and processing logs, binding on a processor engaged by a fiduciary as well as on the fiduciary. That is a retention floor, and it cuts against aggressive expiry of workflow history. Rule 7 requires breach intimation to each affected data principal without delay, and to the Data Protection Board without delay and again in detail within seventy-two hours. There is no de minimis threshold.
In the Gulf the binding instruments are shorter and sharper. Oman's Personal Data Protection Law, Royal Decree No. 6/2022, has been fully enforceable since 5 February 2026, and its Article 23 penalty for unlawful cross-border transfer runs from 100,000 to 500,000 Rial Omani against bands of 500 to 20,000 for almost everything else. The United Arab Emirates federal Personal Data Protection Law, Federal Decree-Law No. 45 of 2021, has been in force since 2 January 2022 with no Executive Regulations as at 8 November 2026, so it carries no commencement date, no statutory notification deadline and no published penalty schedule. Against that, Article 13 of Federal Law No. 2 of 2019 prohibits storing, processing, generating or transferring health data relating to health services provided inside the State outside the State, free zones included, with a fine under Article 24 of not less than AED 500,000 and not more than AED 700,000.
A workflow that moves a UAE patient record to a European region is therefore the one design in this book with a criminal provision attached. Classify before you orchestrate.
The regional ground changes the sequence and sometimes the product, rather than the architecture. Classify the data each process touches before selecting a deployment model, because the classification decides whether the multi-tenant cloud is available at all. Establish the supplier gate next, because in four of these jurisdictions it decides whether the platform can be bought by a public body regardless of where the data sits. Separate the process map from the workflow engine in every residency document, because the vendor's own register separates them. And obtain any residency commitment contractually, by name, because the operational footprint is not a commitment.
A regional position earns its keep once it sits on paper in three places, in identical wording. A difference between the three is what an examiner finds.
The line usually missing is the plainest one. State that the workflow engine is not in region, and name where it is. That fact does not improve by being omitted, and it is the premise of every other position in this chapter.
The decision is a sequencing problem rather than an architecture one. What must be in place before the first build, the order the work has to run in, the controls and how each is tested, the ways a production orchestration fails, and what to measure once it is live.
If a programme cannot answer these plainly, it is closer to rung one than it thinks. Each maps to a failure the data in Part I already documented.
A programme that answers all seven crisply is on the two-in-three side of the buy-versus-build bet. A programme that answers vaguely is, statistically, among the 95% of organisations getting zero return, and does not yet know it.
Most stalls at the integration wall are prerequisite failures wearing a technical costume. Six things must exist before a process is scoped, and only one of them is software.
The second item deserves its own sentence. Credentials are an organisational artefact, not a technical one, and they are obtained from the team that owns the system of record. Ask for them in week one, not in the week the integration is scheduled.
The order of work is fixed by dependency. Qualification decides which process is worth building. Design divides the labour between the deterministic floor, the judgement lane and the person. Integration connects the systems of record, and governance is built before scale rather than after it. Each stage has a test that says whether it is finished, and the test is not a milestone date.
| Stage | What it produces | The exit test |
|---|---|---|
| 01 Qualify | One named process, with a measured baseline and a target | The baseline figures come from system data rather than from an interview, and the process owner has signed them |
| 02 Design | The division of labour, step by step, with control points placed | Every step is assigned to the workflow, the judgement lane or a named person. No step is unassigned |
| 03 Integrate | Live connections to every system of record in scope | The end-to-end flow runs on live data in a non-production environment, including the two hardest exceptions |
| 04 Govern | Approvals, policy binding, lineage and the audit trail | An auditor can reconstruct a completed run from the record alone, without asking the build team a question |
| 05 Operate | A measured, supported production service | The process has run a full cycle of live volume, and the measured result is reported against the stage-one baseline |
Source: Volge analysis, 2026. The five stages are a working sequence rather than a validated methodology, and the exit tests are the part worth copying.
Stage three carries the most risk and the least glamour. The instruction to include the two hardest exceptions is deliberate. A flow that handles the straightforward cases on live data has proved the connection and nothing about the design, because exceptions are where real volume lives.
A governed orchestration stands on a small number of controls, and each one can be tested in an afternoon. An untested control is a claim, and claims do not survive a post-incident review.
Source: Volge analysis, 2026. Control names are generic rather than a product's, so the set applies to whichever platform a programme runs.
The replay test is the one worth running first. A deterministic floor that varies is the failure this book opened with, arriving quietly, and twenty replays of one case will find it in an hour.
| Control | What it is | How you test it |
|---|---|---|
| Identity binding | Every action carries an authenticated actor, human or not, and the policy in force at the time. | Take one completed run and name the actor behind every action in it. Any action attributed to the platform rather than to an identity is a gap. |
| Deterministic floor | Mandatory steps and compliance gates execute identically on every run. | Replay the same case twenty times and compare the mandatory steps. Any variance is a step that was left to judgement by accident. |
| Approval gate | Execution halts until a required sign-off is given. | Attempt to complete a gated run without the approval. The attempt should fail and be logged as a failure. |
| Human control point | Exceptions and escalations route to a named owner with authority. | Check the approval authority matrix against the live routing configuration. They diverge within months of go-live. |
| Confidence threshold | The judgement lane routes low-confidence decisions to a person rather than acting. | Submit cases engineered to sit below the threshold and confirm each routes out. Then check what the threshold was last changed to, and by whom. |
| Data lineage | The inputs a decision rested on are traceable and versioned. | Pick a decision from last quarter and retrieve the exact input records it used. |
| Audit trail completeness | The whole run, automated and human alike, is written to one record. | Reconstruct a run end to end from the trail alone. If the build team is needed to interpret it, the trail is incomplete. |
| Exception-rate monitoring | Departures from the expected path are counted and trended, not merely handled. | Ask for the exception rate by type for the last three months. An orchestration that handles exceptions silently is accumulating a design problem. |
| Credential and secret rotation | Connector credentials expire and are rotated on a schedule. | Ask when each connector credential was last rotated. The answer is usually never, and an expired credential stops the floor rather than the judgement lane. |
A production orchestration fails in a handful of recognisable ways, and none of them presents as a failure. Each looks like a working system producing results nobody has reason to question.
01 Silent threshold drift
A confidence or tolerance threshold is widened to reduce the volume reaching a person, and the widening is never reviewed. In production this looks like a falling human-touch rate, which reads as success on a dashboard. The control is to treat the threshold as a governed parameter with a change record and an owner, not as a configuration setting.
02 The exception path that became the main path
An exception route built for rare cases starts carrying real volume because an upstream system changed. In production this looks like stable cycle time and a quietly rising exception rate. The control is the exception rate by type, trended rather than totalled.
03 Approval fatigue
A gate placed too widely sends more work to an approver than the approver can read, so approvals become a reflex. In production this looks like a complete audit trail of approvals that were not decisions. The control is to measure median time spent at each gate and to narrow any gate whose median approaches zero.
04 The orphaned workflow
The person who understood a workflow leaves, and the workflow keeps running. In production this looks like an automation nobody will change, so processes are redesigned around it rather than through it. The control is an owner per workflow, reviewed when people move, and a documented process definition that is not the implementation.
05 Credential expiry on the floor
A connector credential expires and the deterministic floor stops mid-case. In production this looks like a sudden backlog in one process and a healthy platform everywhere else. The control is scheduled rotation with an alert on approaching expiry, which is dull and prevents the loudest failure in the list.
06 Sprawl by success
The platform works, so every team builds on it without a release route, and within a year there are forty workflows and four conventions. In production this looks like a successful programme that has become difficult to change. The control is the change and release route from the prerequisite list, applied from the first workflow rather than the fortieth.
Five measures carry a production orchestration, and four of them are available from the platform's own record. The fifth has to be taken from finance, which is why it is usually the one missing.
| Measure | Definition | Direction | Cadence |
|---|---|---|---|
| Cycle time, median and ninetieth percentile | Elapsed time from intake to close, per process | Falling, with the tail watched more closely than the median | Weekly |
| Exception rate by type | Cases leaving the expected path, as a share of volume, split by cause | Falling, and attributed | Weekly |
| Human-touch rate | Share of cases reaching a person, against the share the design intended | Matching the design, in either direction | Monthly |
| Measure | Definition | Direction | Cadence |
|---|---|---|---|
| Straight-through completion | Cases completed with no manual intervention and no retry | Rising | Monthly |
| Cost per case | Fully loaded cost of running the process, including platform and support | Falling against the stage-one baseline | Quarterly |
Source: Volge analysis, 2026. Definitions are stated so that a figure means the same thing in two consecutive quarters, which is most of the value of a measure.
Note what is absent. There is no automation count, because the number of workflows built is a measure of activity and this book is about return. And the human-touch rate is read against the design rather than driven downward, because a design that intended twelve per cent of cases to reach a person and now sends two has not improved. It has stopped doing what it was approved to do.
An orchestration is finished, in the only sense that matters, when four statements are true and evidenced. The process runs on live systems of record with no manual bridge. A completed run can be reconstructed from the audit trail by someone who did not build it. The measured result stands against the baseline taken before the build. And a named person owns each control point, with the routing configuration matching the approval authority matrix.
Nothing on that list is a go-live date. Three of the four can only be demonstrated after a cycle of real volume, which is why a programme that declares victory at go-live has described a rung and called it a ladder.
Workflow automation is the layer with the most moving parts to procure. Two lists follow, the Nintex elements to licence, and the non-Nintex pieces that must already be in place, so a build-and-run programme does not stall on a missing prerequisite. Tiers and figures are indicative for 2026; confirm current SKUs with Nintex.
Nintex elements to licence
| Nintex Automation CE | The cloud workflow platform, within the Nintex CE Platform. Subscription, Standard or Enterprise tier; priced by users or workflow-execution volume, on quotation. |
|---|---|
| Nintex Forms | Digital intake forms. Included in the platform. |
| Nintex DocGen | Document generation. Licensed as an add-on module. |
| Nintex eSign | Compliant e-signature. Licensed separately; lightweight signature is now included in Nintex Documents. |
| Nintex RPA | Bridges legacy screens that have no API. Licensed separately. |
| Nintex Automation K2 | On-premises / hybrid engine for regulated, mission-critical builds. Separate on-prem licence. |
| Nintex Solutions | Governed containers, with Solution Designer and Solutions ALM, from the Q3 2026 release. Part of the CE Platform. |
| Premium connectors | SAP, Oracle, Salesforce and similar. Licensed per connector. |
What else must be in place (non-Nintex)
| Microsoft 365 and/or Salesforce | The systems the workflows act on, plus API access to any ERP/CRM in scope. |
|---|---|
| Identity / single sign-on | Microsoft Entra ID or an equivalent SSO provider. |
| On-prem infrastructure (for K2) | Windows Server and SQL Server, plus the hosting infrastructure and network. |
| Change & release governance | So scaled workflows do not fork per team as they spread. |
| Owners and a run team | Process owners and an admin or centre of excellence to run the platform after go-live. |
The terms used in this report, defined for a generalist reader.
| Term | Definition |
|---|---|
| Agentic Business Orchestration (ABO) | Nintex's framework combining structured workflow executions, AI-driven decision-making, and human control in one governed system. Unveiled 24 September 2025; capabilities announced 17 March 2026. |
| Nintex CE Platform | The consolidated platform of the Q3 2026 release, 29 September 2026, holding workflow, apps, forms, orchestration, agents and Solutions. |
| Deterministic workflow | A process that runs the same exact steps in the same order every time: the floor that guarantees precision and compliance. |
| Probabilistic AI | A model that may make different choices on different runs; suited to judgement, unsuited to mandatory exact steps. |
| Human in the loop | A designed control point where a named person approves, overrides, or escalates before execution resumes. |
| Nintex Agent Designer | The governed, low-code environment for building mission-critical agents. Announced 17 March 2026; described by Nintex as shipping in the CE Platform. |
| Nintex Orchestration | Modular, phase-based workflows that move forward, jump back, or repeat. Announced 17 March 2026; described by Nintex as shipping in the CE Platform. |
| Nintex Automation CE | The build-and-run substrate for ABO, within the Nintex CE Platform. Nintex publishes no list price. |
| Nintex Processes | Process management native to the CE platform, launched 26 October 2026 and presented as the next evolution of Process Manager rather than a replacement. |
| Nintex RPA | Task automation across applications. Nintex RPA (formerly Kryon; release 26.6, July 2026) is the strategic go-forward line; Nintex Foxtrot RPA (formerly Nintex RPA Foxtrot and Nintex RPA LE) is maintained alongside. |
| Nintex DocGen | Document generation from process data, including a DocGen Agentforce connector for Salesforce. |
| Nintex Forms / eSign | Structured data capture at the front of a process; electronic signature at accountable control points. |
| Maturity ladder | The five rungs from demo to orchestrated enterprise used in Part III to locate a programme's true state. |
| Total Economic Impact (TEI) | Forrester's modelled cost-and-benefit methodology; here, 213% ROI and a nine-month payback for Nintex Workflow, conducted late 2022. |
| Proof of concept (PoC) | A limited build to test feasibility; organisations scrap ~46% of projects between proof of concept and broad adoption (S&P Global, 13 January 2025). |
| Software sprawl | The accumulation of disconnected tools and pilots, each with its own cost, risk, and integration debt. |
Every figure in the argument traces to one of the entries below; the About-Nintex page (folio 04) additionally draws on Nintex corporate and customer materials, 2022–26. Estimates are labelled as estimates. We deliberately omit the unverified "89% never scale" claim.
Gartner: over 40% of agentic-AI projects projected to be cancelled by end-2027 on escalating cost, unclear value and inadequate risk controls; of the thousands of agentic-AI vendors, Gartner estimates only about 130 are real. 25 June 2025.
MIT NANDA: The GenAI Divide, July 2025: ~95% of organisations are getting zero return from generative AI; externally sourced tools reached deployment ~67% of the time against ~33% for internally built tools, a share of deployments rather than a success rate. The paper disclaims institutional attribution to MIT, and 18 August 2025 is the date of Fortune's write-up, not the paper's.
S&P Global 451 Research Market Insight, Alex Johnston, 13 January 2025 (n=1,006, fielded October to November 2024): 42% of companies abandoned most AI initiatives before production, up from 17%; organisations scrap ~46% of projects between proof of concept and broad adoption. 14 March 2025 is the date of CIO Dive's story, not the primary.
Deloitte: State of AI in the Enterprise 2026: only 25% of its 3,000-plus executive respondents have moved ≥40% of AI pilots into production. 21 January 2026; fieldwork August to September 2025.
Nintex: Agentic Business Orchestration unveiled; structured workflow + AI decisioning + human control; positioned against software sprawl and disconnected pilots; 7,000+ organisations across 100+ countries. 24 September 2025.
Nintex: Niranjan Vijayaragavan, then chief product and technology officer: “AI for modern business cannot be purely agentic, nor purely deterministic; it must support both.” 17 March 2026.
Nintex: then chief executive Amit Mathradas on orchestrating across people, systems, and AI agents to deliver outcomes at scale. Business Wire, 24 September 2025. (Departure announced 17 December 2025, effective 2 February 2026; Stephen Elop is now chief executive.)
Nintex: Agent Designer (governed, low-code) and Nintex Orchestration (phase-based: move forward, jump back, repeat) announced 17 March 2026; open beta from the Q2-2026 update, 3 June 2026; described as shipping parts of the Nintex CE Platform from 29 September 2026.
Nintex: Automation CE build-and-run substrate, within the Nintex CE Platform; Nintex publishes no list price. Composable toolkit: RPA (formerly Kryon, release 26.6, 13 July 2026, strategic go-forward; Nintex Foxtrot RPA, formerly Nintex RPA Foxtrot and Nintex RPA LE, maintained), DocGen (incl. DocGen Agentforce connector for Salesforce), Forms, eSign.
Nintex: Nintex Processes launched 26 October 2026, native to the CE platform and presented as the next evolution of Process Manager; the Process Manager MCP Server has been in open beta since August 2026, with MCP authentication for eleven clients, among them ChatGPT, Claude, Gemini and Microsoft 365 Copilot, in the October 2026 Process Manager release, 4.26.23. Nintex Process Discovery was put end-of-life on 4 May 2026: end of life 1 May 2027, end of support May 2028, with customers migrated to Process Manager.
Forrester: The Total Economic Impact of Nintex Workflow, commissioned by Nintex and conducted in late 2022: 213% ROI, nine-month payback, modelled on a composite organisation. Separately, Forrester named Nintex a Strong Performer in The Forrester Wave: Agentic Process Orchestration Platforms, Q3 2026.
G2 / TrustRadius: aggregate user ratings, read 6 November 2026: G2 4.2/5 across 1,326 reviews; TrustRadius 8.0/10 across 530 reviews.
Market.us (estimate): Agentic-AI market ~$5.2B (2024) → ~$196.6B by 2034, ~43.8% CAGR. Labelled as an estimate, not a Gartner or Forrester forecast.
Note on claims we do not make. We do not assert a Gartner Magic Quadrant or Forrester Wave leadership position for Nintex; Forrester's Q3 2026 Agentic Process Orchestration Platforms Wave placed it as a Strong Performer, which is a placement and not leadership. Where a source is a single customer story, we say so; where a figure is modelled or estimated, we label the method and the year.
Volge takes automation programmes across the wall between an integrated pilot and a production service, and the work is engineering rather than advice. The practice sits where this book locates the failures, in the integration into systems of record and in the governance a risk committee will sign. Delivery is by senior engineers who build inside the client's own tenant and environment, through to a measured result reported against the baseline taken before the build.
The work has a usual shape. Qualification picks one process on evidence rather than enthusiasm, and takes the baseline from system data because a return with nothing to compare it to is an assertion. Design divides the labour step by step, so that every step is assigned either to the deterministic floor, to the judgement lane or to a named person, with none left unassigned. Integration comes next and connects the systems of record on live data, including the two hardest exceptions, because that is where real volume lives. Governance is built before scale. The measured result closes the engagement, not the go-live.
Who does the work. A deployment strategist holds the business outcome and the governance cadence with the client's operations and risk leadership. A transformation architect holds the division of labour, the control-point design and the sequence. Senior engineers own the integrations and build inside the client's environment, including the connector credentials and the release route. The same people stay from qualification through the first full cycle of live volume, which is what makes the control tests in chapter six drills the team has run rather than procedures it has recommended.
The regional position is often what a Gulf or Indian client needs first, and it is less comfortable here than in most subjects. Volge is headquartered in the Gulf, and the finding in chapter five is treated as a design input rather than a footnote: the process map can sit in the United Arab Emirates and the workflow engine cannot sit anywhere in the region. Where a classification or an accreditation gate rules out the multi-tenant cloud, as the Omani and Kuwaiti instruments can, the self-hosted route is scoped on its own terms and the difference in cost and effort is stated before a platform
is selected. Where a residency commitment is needed, it is obtained contractually rather than inferred from an operations page.
An engagement starts with a scoped assessment, usually four to six weeks. It produces a ladder position for each process already in flight, a measured baseline for the first candidate, a division-of-labour design with the control points placed, a residency and classification position for the data each process touches, and a staged plan whose first stage already carries an exit test. Volge is a Nintex delivery partner and benefits commercially if a client buys, which a reader should weigh against the recommendations in this book, including its preference for buying over building. The conversation begins wherever the programme stands today.
Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.
Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.
Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.
This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.