Contents
The Voice of Volge is Volge’s thought-leadership series, unsigned and institutional, every figure sourced, quantified and year-stamped. Volge is a Nintex delivery partner; interest disclosed.
volge.com/voice
About Nintex
Nintex is a process-and-automation software company. What began as SharePoint-era workflow is now a single platform that maps, mines, automates and, since 2025, orchestrates AI agents against work, built out over two decades and a handful of pointed acquisitions.
The chapters that matter to a buyer today: Agentic Business Orchestration (Sept 2025), locally-hosted AI in K2 5.9.1 (Apr 2026), and Agent Designer + Orchestration in beta.
This report covers Agentic Automation, the Orchestrate layer, where governed agents run inside the workflow.
| Understand | Process Manager · Process Intelligence |
|---|---|
| Build & run | Automation Cloud · Automation K2 (on-prem) · RPA · DocGen · Forms · eSign |
| Orchestrate Focus of this report | Agentic Business Orchestration · Agent Designer (beta) · Nintex Orchestration (beta) |
Source: Nintex customer stories and G2/TrustRadius, 2022–26. Customer outcomes as reported by Nintex.
A two-decade platform that has just added governed AI and a locally-hosted engine, mature where it must be, current where it counts.
Move the model, not the data.
Cloud agents move your data to the model. The sovereign alternative moves the model to your data, inside the boundary you already trust, and that single inversion is the subject of this book.
The question is not whether, but where
Most agentic-AI pitches answer a question regulated buyers are not asking. The interesting variable is not model quality. It is the address at which the model runs.
A government department, a systemically important bank, a defence supplier and a health system share a constraint that consumer-grade AI vendors tend to treat as a footnote. Their data may not leave a controlled boundary, not for latency or cost, but because the law says so. For an EU financial firm the instruments are specific: the Digital Operational Resilience Act, applicable since 17 January 2025, and GDPR's Chapter V transfer regime, hardened by the Court of Justice's Schrems II judgment of July 2020. For a defence supplier it is ITAR and the US Department of Defense impact levels; for a health system, sector privacy law. The common effect is a boundary regulated data may not cross.
The prevailing agent platforms were built the other way round. They assume a round trip to a large model behind someone else's API. That assumption is invisible in a demo and load-bearing in an audit.
Nintex's contribution, in the K2 5.9.1 release of 29 April 2026, is unglamorous and precisely aimed. It puts a small model inside the automation platform itself, so that routing, prioritisation and issue-detection decisions are made where the data already sits.
Our interest is disclosed: Volge is a Nintex delivery partner. We have kept that separate from the evidence. Where Nintex makes a claim we quote it and date it; where the constraint is legal we name the instrument; where we compare Nintex with Microsoft, AWS or Google, we label the comparison as ours.
The reader we have in mind runs risk for a living. They have heard "sovereign" used as a paint colour. So we keep the vendor's own caveat visible: the no-external-call claim holds during normal operation, and topology decides the rest.
Read the next three pages for the shape of the problem. The remaining forty are the detail an architect will want before signing.
The barrier for regulated firms hasn't been interest, it's been how to apply it responsibly.
Sources: Nintex, Nintex Automation K2 5.9.1 release materials, 29 Apr 2026. Volge editorial framing of vendor comparisons.
The wall and the agent
Regulated organisations want agentic AI and are structurally unable to buy the version on offer. Nintex's answer is to move the model, not the data.
Consider the position of a regulated firm, a national bank, in the middle of 2026. The board wants a plan for agentic automation, because competitors claim to have one. The chief operating officer wants the throughput it promises. The regulator, and the firm's own risk and security leaders, want an assurance that customer records never traverse an uncontrolled network. Those demands are usually answered by different vendors who do not talk to each other.
The friction is not political and it is not fashion. It is architectural. The agent platforms that dominate the market, Microsoft Copilot Studio, Amazon Bedrock Agents, Google Vertex AI Agents, are, in Volge's assessment, designed to reach a large model hosted by the provider. The agent reasons in the provider's cloud. To reason about your data, it must be shown your data.
For most enterprises that is a manageable trade. For a bank under a data-residency rule, a defence contractor on a classified network, or a government body bound by a sovereignty statute, showing the data to an external endpoint is the one thing that cannot happen. The prohibition is the whole point of the control.
Cloud agents move your data to the model. Nintex Automation K2 5.9.1 moves the model to your data. That single inversion is the subject of this book.
Nintex's framing, in its own words for K2 5.9.1, is deliberately modest: "AI-driven decisions, including routing, prioritization, and issue detection, run entirely within the customer environment, with no external dependencies required" (Nintex, 29 Apr 2026). The product page is blunter still: "No cloud dependency. No external API calls. Just AI-driven decisions running inside your environment."
We will hold that claim up to the light in Part II, caveats included. For now, note only what it makes possible. An agent can read a case, judge its severity, decide where it should go and flag an anomaly, without a single packet of the underlying record leaving the boundary you already trust.
Sources: Nintex Automation K2 5.9.1 release and product page, 29 Apr 2026. Named-vendor comparison is Volge analysis, not a Nintex claim.
The reflex response, "run an open-source model yourself", is where most sovereign-AI programmes go to die. Standing up a private model is not one project; it is a model, a serving stack, a security review, an identity integration, a workflow engine to call it, and an audit trail to defend it. Each is a team. Regulated firms rarely have the years.
What Nintex offers instead is the model arriving pre-integrated inside an automation platform those firms, in many cases, already run. The AI is not a bolt-on service to be secured separately. It is a feature of workflows and forms the organisation has been operating, in some cases, for over a decade.
Two dates matter. On 24 September 2025, Nintex unveiled Agentic Business Orchestration, deterministic workflows, AI decisioning and human oversight, governed together. On 29 April 2026, K2 5.9.1 shipped the locally-hosted model that makes the decisioning run on-premises. The second date is when the sovereign version stopped being a slide and became a release.
We will claim that K2 5.9.1 runs its AI decisions locally, because Nintex says so and dates it. We will not claim a Gartner or Forrester ranking, because we have none to cite. And we will keep Nintex's own scope, "during normal operation", attached to the sovereignty story rather than buried in a footnote.
"Air-gapped", "off the hyperscalers" and "own, don't rent" are Volge's editorial framing of a genuine capability. Nintex's own phrasing is the more careful "cloud-only alternatives". We keep both visible so you can tell marketing from mechanism.
Sources: Nintex, Agentic Business Orchestration launch, 24 Sep 2025; K2 5.9.1 release, 29 Apr 2026.
What "inside your walls" is worth
The value of local execution is not a feature list. It is the removal of a conversation you would otherwise have to have with a regulator, a data-protection officer and an internal audit function, repeatedly, for every use case.
When the model runs inside the boundary, the data-flow diagram that governs an AI use case gets shorter. There is no external processor to name in a record of processing. There is no cross-border transfer to justify. There is no third-party endpoint to add to a threat model.
None of that makes the system risk-free. It changes which risks you own. You take on model-lifecycle and capacity risk; you shed data-egress and third-party-dependency risk. For a regulated buyer, that is usually the trade worth making.
| Question | Cloud agent | K2 local |
|---|---|---|
| Does data leave? | Yes, to the model API | No* |
| Works air-gapped? | No | Yes |
| Third party to name? | The provider | None for AI |
The through-line
Everything that follows is an elaboration of one move: keep the reasoning where the data lives. Part I shows why the alternative stalls. Part II shows how Nintex does it. Parts III–V show that it is governed, proven and buyable.
Sources: Nintex, K2 5.9.1, 29 Apr 2026. Comparative characterisations are Volge analysis.
The round trip that cannot happen
A cloud agent's intelligence is somewhere else. To use it, your data has to go there. In regulated settings that single fact disqualifies the design.
Strip the marketing from a cloud agent and the mechanism is simple. A prompt, assembled from your data, is sent to a model hosted by the provider. The model reasons and returns a result. The reasoning happens on infrastructure you do not control, reached across a network you do not own.
For a retailer optimising promotions, that is fine and often excellent. For a bank, a ministry or a defence supplier, the assembled prompt is the problem. It contains the very records the control regime exists to contain.
The usual mitigations do not dissolve the issue; they relocate it. Private networking narrows the path but still terminates at the provider. Regional hosting satisfies residency but not a genuine air-gap. Redaction removes the useful signal along with the sensitive tokens. Each is a compromise an auditor will probe.
There is also a dependency question that has nothing to do with secrecy. An agent that must reach an external endpoint stops working when the endpoint is unreachable, by outage, by a severed link, or by design on a classified network. "No external dependencies required" (Nintex, K2 5.9.1, 29 Apr 2026) is as much an availability statement as a privacy one.
None of this is a criticism of the cloud platforms on their own turf. It is an observation that their turf has an edge, and that regulated data lives on the far side of it.
Nintex frames the same barrier in operator's terms. Its CPTO says the obstacle "hasn't been interest, it's been how to apply it responsibly", and that the fix is keeping AI "within the boundaries they already trust" (Nintex, 2026). We think that is the correct diagnosis.
The movement of data out of a controlled boundary to an external system. For a regulated buyer, an AI feature that requires egress is not a smaller version of a compliant feature. It is a different category of thing.
Sources: Nintex, K2 5.9.1, 29 Apr 2026. Characterisation of cloud-agent mechanics is Volge analysis.
The rules that put the wall there
The boundary is not Volge's framing: it is codified in law a regulated firm is already audited against. These instruments make data egress a compliance event, not a preference.
| Instrument | What it obliges | Status / date |
|---|---|---|
| EU DORA | Financial entities must manage ICT third-party and concentration risk, map dependencies and control where data sits | Reg. (EU) 2022/2554 · applicable 17 Jan 2025 |
| GDPR, Chapter V + Schrems II | Transfers of personal data outside the EU need adequacy or safeguards; a US-hosted model endpoint is such a transfer | Reg. (EU) 2016/679 · CJEU C-311/18, 16 Jul 2020 |
| EU AI Act | High-risk AI: data governance, automatic logging and traceability, human oversight: far easier to evidence in-boundary | Reg. (EU) 2024/1689 · high-risk duties from 2 Aug 2026 |
| NIST AI RMF · ISO/IEC 42001 | Voluntary control frameworks: the governance most US and international enterprises are asked to show against | RMF 1.0, Jan 2023; ISO 42001, Dec 2023 |
| US defence: ITAR · CMMC · DoD IL5/IL6 | Technical data and mission data must stay under US-person control on sovereign, isolated infrastructure | CMMC clauses in DoD contracts from 10 Nov 2025 |
Together these turn "keep the data inside the boundary" from a preference into a control a firm must evidence to a supervisor. An architecture that reasons over regulated data in a provider's cloud does not fail on engineering: it fails because it cannot produce that evidence.
Every claim that follows is measured against these rules, not a marketing promise. If a design cannot satisfy them, no benchmark rescues it.
Every external call is a governance event
The expensive part of a cloud agent in a regulated firm is not the licence. It is the standing overhead each external call imposes on assurance, legal and audit.
Once a design sends regulated data to an outside endpoint, a set of obligations attaches and never fully detaches. They are illustrative of the pattern, not a scored index; the point is that the list exists at all, and that local execution shortens it.
Read the exhibit as a shape, not a score. Four categories of standing work attach to an external-call design and recur for every new use case. Local execution does not zero them, but it collapses the ones driven by egress and third-party dependency.
This is why "run it locally" is a commercial argument as much as a technical one. The saving is the assurance cycle you do not repeat.
If a control exists to keep data inside a boundary, an AI design that sends data outside it inherits every question that control was written to ask. Removing the call removes the questions.
Sources: Volge analysis, 2026; Nintex, K2 5.9.1, 29 Apr 2026.
On-prem agents versus cloud agents
The two designs differ on one axis that dominates every other: where the model runs.
| Dimension | Nintex Automation K2 5.9.1 (local) | Cloud-only agent platforms |
|---|---|---|
| Where the model runs | Inside the customer environment | Provider cloud, behind an external API |
| External API calls for AI decisions | None during normal operation | Required |
| Runs fully air-gapped | Yes (Volge framing) | No |
| Data leaves the boundary | No* | Yes, to the model endpoint |
| Third party to name in processing records | None for the AI decision | The cloud provider |
| Deployment modes | On-prem, private cloud, managed hosting | Provider cloud |
| Named examples | K2 5.9.1 built-in local model | Copilot Studio · Bedrock Agents · Vertex AI Agents |
| Availability if link is severed | AI decisions continue | AI decisions stall |
The table is not a verdict that one architecture is better; it is a map of where each fits. If your data can move, a cloud agent may be faster. If it cannot, the right-hand column is closed to you: the control regime you operate under has made the choice.
Sources: Nintex, K2 5.9.1, 29 Apr 2026; Nintex, 24 Sep 2025. Named-vendor comparison: Volge analysis.
A map, not a mandate
For data that may leave the boundary, public content, low-sensitivity operational records, anything already destined for a SaaS, a cloud agent is often the shortest route to a working system. The provider carries the model lifecycle. You carry integration.
We say this plainly because a brief that pretended cloud agents were always wrong would be selling, not advising. On their own ground they are strong.
For data bound by residency, classification or a sovereignty statute, the model has to come to the data. Here K2 5.9.1's local decisioning is not a preference; it is the precondition for doing agentic work at all.
Nintex names the constituency directly: "industries where data sovereignty is non-negotiable", citing government and regulated sectors (Nintex, 29 Apr 2026).
The honest boundary
Many organisations are both. A single enterprise may run cloud agents for marketing and local agents for the parts of the business a regulator watches. The skill is drawing the line correctly, not pretending it does not exist. Part V returns to how to draw it in practice.
We do not argue cloud agents are obsolete. We argue that a large, well-defined set of workloads is closed to them by rule, and that this set is exactly where Nintex's local engine earns its keep.
Sources: Nintex, K2 5.9.1, 29 Apr 2026. Fit assessment is Volge analysis.
Sovereignty is also an uptime story
Data protection dominates the sovereign-AI conversation, but the same architecture answers a separate question a chief operating officer cares about: does the agent keep working when the outside world does not?
An external dependency is a shared fate. If the provider has an incident, your agent has an incident. If a submarine cable is cut, a data-centre region degrades, or a border control severs a link, the reasoning that lives on the far side of that link becomes unreachable.
For a consumer app, that is an inconvenience measured in minutes. For a hospital triaging cases or a control room prioritising incidents, an agent that stalls because a distant API is unreachable is worse than no agent, because a process was built to assume it.
K2's "no external dependencies required" (Nintex, 29 Apr 2026) reads differently through this lens. It means the intelligence and the process share the same fate as the rest of your on-premises systems, no better, but, more to the point, no worse.
This is the quiet reason air-gapped operators care. On a classified network there is no external endpoint to reach, by design. A local model is not a nicety there; it is the only kind of model that can exist.
Local execution removes the external dependency; it does not remove all dependency. You now depend on your own capacity planning, patching and resilience for the model. The risk moves inside, where you can govern it, rather than vanishing.
Sources: Nintex, K2 5.9.1, 29 Apr 2026; Volge analysis, 2026.
The stall, restated
Cloud agents reason where the provider's model lives. Regulated data lives behind a wall. The two facts do not reconcile, and no amount of private networking fully reconciles them.
That is the whole of Part I. It is not that hyperscaler platforms are weak. It is that a defining set of workloads sits, by rule, on the wrong side of the boundary those platforms assume they can cross.
| If you must answer… | …and the answer is "no" | Then |
|---|---|---|
| May regulated data leave the boundary? | No | Local model required |
| Can the agent depend on an external API? | No | Local model required |
| Is a third-party processor acceptable for AI? | No | Local model required |
Turn the page
Part II opens the engine. What exactly runs locally in K2 5.9.1, what the boundary looks like as an architecture, and how honest the "no external API calls" claim is under an operator's inspection.
Sources: Volge analysis, 2026; Nintex, K2 5.9.1, 29 Apr 2026.
What Nintex actually says
We quote the sovereignty claim in full so you can judge it, then translate it into architecture. The wording is careful, and the caveat is deliberate.
AI-driven decisions, including routing, prioritization, and issue detection, run entirely within the customer environment, with no external dependencies required.
No cloud dependency. No external API calls. Just AI-driven decisions running inside your environment.
Read those two sentences as an architect. They assert three things. First, a model exists inside the platform, "built-in, locally-hosted". Second, specific decision types run there: routing, prioritisation and issue detection. Third, those decisions need nothing outside the environment.
Nintex names where this appears: sentiment analysis and severity scoring "inside SmartForms and workflows" (Nintex, 2026). It is not an abstract capability. It is embedded in the forms and processes the platform already runs.
Nintex scopes the claim to normal operation. We keep that word attached. It signals that model updates, licensing checks or optional online features may involve connectivity: the routine AI decision does not. Confirm the exact boundary for your topology.
Sources: Nintex, K2 5.9.1 release and product page, 29 Apr 2026. Emphasis on "normal operation" is Nintex's own scope, restated by Volge.
Data that never leaves the boundary
The whole design is one enclosure. Data, model and decision sit inside it. The external model API sits outside, and the routine decision never crosses the line.
The diagram carries the argument better than prose. Capture, reasoning, decision, audit and human oversight all sit within a single dashed enclosure, the customer environment. The external model endpoint, drawn faint and struck through, is the thing a cloud agent depends on and this design does without.
Sources: Nintex, K2 5.9.1, 29 Apr 2026. Diagram is Volge's schematic, not a Nintex reference architecture.
What the model decides, in place
The local model is not a general chatbot. It makes three decisions, routing by content, prioritising by urgency, detecting anomalies, plus sentiment and severity scoring inside SmartForms. Each is a private judgement worth more than an open-ended conversation.
Each is a judgement a person once made by reading the item, high-volume, repetitive, bounded, precisely the shape of work where a small local model earns its cost and human attention is the scarce resource.
Because the judgement happens inside a SmartForm or workflow step, its output is a first-class part of the process: it can trigger a route, raise a flag, or set a priority, and it is logged like any other step.
For classification, scoring and routing on your own data, a modest model that never leaves the boundary often beats a frontier model reached across a network you cannot fully trust. The constraint sets the winner.
| Step | What happens | Where |
|---|---|---|
| 1 · Capture | Complaint submitted via SmartForm | Inside boundary |
| 2 · Score | Local model scores sentiment and severity | Inside boundary |
| 3 · Route | High-severity case routed to specialist queue; anomalies flagged for review | Inside boundary |
| 4 · Escalate | Agent brings a person in to decide | Inside boundary |
Sources: Nintex, K2 5.9.1, 29 Apr 2026.
The model does not work alone
The local engine is one part of a platform. Several long-standing Nintex capabilities make the sovereign story practical, because they too are built to run inside the boundary.
Nintex SmartForms and Forms capture the data at source and host the AI-scored fields. This is where sentiment and severity surface to a user, and where a routing decision becomes visible.
Nintex eSign keeps the signature step inside the same trust boundary as the rest of the process, rather than pushing it to an outside service.
Nintex DocGen assembles documents from process data. Combined with local decisioning, a case can be captured, judged, routed and turned into an output document without the data touching an external endpoint.
K2 5.9.1 simplifies identity with OIDC and SCIM, so the platform integrates with the directory you already run: with WCAG accessibility improvements in the same release (Nintex, 29 Apr 2026).
A local model is only as sovereign as the process around it. Because forms, signing and generation run in place too, the boundary holds end to end, not just at the moment of decision.
Sources: Nintex, K2 5.9.1, 29 Apr 2026.
Where "no external calls" ends
A brief that only repeated the marketing would be worthless to anyone who has to sign the deployment off, security, risk, IT and the board alike. So here is the boundary of the claim, in the plainest terms we can manage.
Nintex says the AI decision runs locally "during normal operation" and with "no external dependencies required". Both phrasings are precise, and precision has edges. We read the edges as follows, and recommend you confirm each against your own topology.
| Activity | Local by default? | What to confirm |
|---|---|---|
| Routine AI decision | Yes | The core claim; runs in-boundary during normal operation |
| Sentiment / severity scoring | Yes | Inside SmartForms and workflows |
| Model or platform updates | Depends | How updates are delivered to an air-gapped site |
| Licensing / activation | Depends | Whether any check requires connectivity |
| Optional online features | No | Anything you deliberately connect externally |
The sovereignty claim is strong and, for the routine AI decision, clean. "During normal operation" is not a weasel; it is Nintex declining to promise that update and licensing mechanics never touch a network. One limit we should own too: we have read the wording and the architecture, not run a packet analyser against your build. That verification is the checklist's real job, not a formality, and it is the part we cannot do for you.
Sources: Nintex, K2 5.9.1, 29 Apr 2026; Volge analysis, 2026.
The risks you now own
Moving the model inside the boundary does not delete risk. It relocates it to where you have authority. That is the trade, and it is worth stating without spin.
With a cloud agent, the provider carries the model's capacity, patching and lifecycle, and you carry the egress and third-party exposure. With a local engine, that flips. You gain control of the data path and inherit responsibility for the model's operation.
For a regulated firm this is usually the right way round. Capacity and patching are problems your operations team already solves for other on-premises systems. Data egress is a problem your regulator will not let you solve at all.
Sizing the model to peak decision volume; a patch and update path (including for air-gapped sites); capacity headroom; and resilience for the model alongside the workflow engine. None is exotic. All should be in the deployment plan.
One cost we will not tuck into a table: the model you run in-boundary is smaller than the frontier models a hyperscaler rents you. For routing and scoring the gap is immaterial. For open-ended reasoning it is not. If your case is the latter, this is the wrong tool: we would rather lose the argument here than in production.
| Risk | Cloud agent | K2 local engine |
|---|---|---|
| Data egress | Owned by you, hard to close | Removed |
| Third-party dependency | On the provider | Removed for AI |
| Model capacity / sizing | Provider's problem | Yours to plan |
| Patching / lifecycle | Provider's problem | Yours to run |
| Availability if link cut | Fails | Continues |
Sources: Volge analysis, 2026; Nintex, K2 5.9.1, 29 Apr 2026.
The engine, restated
K2 5.9.1 puts a built-in model inside the customer environment and runs routing, prioritisation, issue detection, sentiment and severity there: with no external API calls during normal operation.
The claim is careful, the caveat is honest, and the surrounding platform, forms, signing, document generation, identity, is built to keep the whole process in the same boundary. What you take on is model operation; what you shed is egress and third-party dependency.
Turn the page
A local model that made unaccountable decisions would be a new liability, not a solution. Part III shows how the decisions are governed: deterministic where they must be, AI where it helps, human where it matters, and audited throughout.
Sources: Nintex, K2 5.9.1, 29 Apr 2026.
Deterministic where it counts, AI where it helps
Nintex's frame, unveiled 24 September 2025, is that agents should sit inside governed processes, not replace them. The intelligence decides; the process still runs on rails.
The failure mode of naive agentic AI is a system that acts confidently and cannot explain itself. For a regulated firm that is not an edge case; it is a disqualifier. A decision a regulator can question must be a decision the system can reconstruct.
Agentic Business Orchestration answers this by combining three things that are usually sold separately. Deterministic workflows provide the rails: the steps, rules and approvals that must happen the same way every time. AI decisioning provides judgement inside those rails. Human-in-the-loop provides an off-ramp when judgement is not enough.
The combination is "governed and auditable" (Nintex, 24 Sep 2025). That is the operative pair of words. Governance means the agent's authority is bounded by the process. Auditability means every decision leaves a record.
The agent is not given the keys to the building. It can "escalate, ask for clarification, or bring a person into the process" (Nintex, 2025). Those three verbs describe an assistant that knows its limits, not an autonomous actor.
Deterministic workflow
The rails: fixed steps, rules and approvals that behave identically every time.
AI decisioning
Judgement inside the rails: routing, prioritisation, detection, scoring.
Human-in-the-loop
The off-ramp: escalate, clarify, or bring a person in to decide.
Sources: Nintex, Agentic Business Orchestration launch, 24 Sep 2025.
A decision you can reconstruct
In a regulated setting, an unexplainable decision is a finding waiting to happen. The value of running AI inside a deterministic process is that the AI's judgement becomes just another logged step.
Because the model's output feeds a workflow, setting a route, raising a flag, assigning a priority, that output is captured where every other process action is captured. There is no separate, opaque AI system to reconcile against the record. The decision and its trail live in the same place.
This matters twice over for a sovereign deployment. The record never leaves the boundary, so the audit trail is as contained as the data. And the trail exists at all, so a supervisor, an internal auditor or a regulator can ask "why did this go here?" and get an answer grounded in the process.
| Audit question | Answered by | Location |
|---|---|---|
| What did the AI decide? | Logged decision in the workflow | In-boundary |
| What was the input? | The SmartForm / case data | In-boundary |
| Was a human involved? | Human-in-the-loop step | In-boundary |
| Where did it go next? | Deterministic routing rule | In-boundary |
| Did the data leave? | No external call for the decision | In-boundary |
Sovereignty and auditability reinforce each other. A contained data path produces a contained audit trail; a governed process produces a decision worth auditing. Neither works as well without the other.
Sources: Nintex, 24 Sep 2025 and 29 Apr 2026.
The agent that knows its limits
The most important thing a governed agent does is stop. Nintex builds three exits into the design, so the model hands off rather than guesses when the stakes rise.
Escalate
Raise the case to a higher authority when it exceeds the agent's remit.
Ask for clarification
Request more information rather than act on an ambiguous input.
Bring a person in
Insert a human decision-maker into the process at the right point.
These verbs are Nintex's own (2025), and they describe a philosophy as much as a feature. An agent that can escalate, ask and defer is an agent whose authority is bounded by design. It does the high-volume judgement and hands back control at the edges, where a person should decide.
For a regulated buyer, this is the difference between augmentation and abdication. The organisation keeps a human accountable at the points that carry legal or safety weight, while the model absorbs the repetitive work that clogs a queue.
Agents can escalate, ask for clarification, or bring a person into the process.
Human oversight only helps if the human has time and context to exercise it. Designing the escalation thresholds, when the agent should defer, is real work, and getting them wrong turns oversight into rubber-stamping. This is a design task, not a default.
Sources: Nintex, 24 Sep 2025; Volge analysis, 2026.
Agentic versus "agentish"
A vendor warning against overclaiming in its own category is unusual and worth noting. Nintex has publicly cautioned against "agent-washing".
The agentic-AI market has a credibility problem, and much of it is self-inflicted. Products that automate a fixed script get rebranded as "agents"; chat interfaces that cannot act get sold as autonomous workers. The result is a buyer who cannot tell capability from vocabulary.
Nintex named this directly in a piece titled "agentic vs agentish" (9 April 2026), warning against "agent-washing": dressing ordinary automation in agentic language. We cite it not as a triumph but as useful sobriety from a vendor with an incentive to hype.
The distinction is the one that matters to an operator. A genuine agent makes a judgement and can alter its course, escalate, ask, defer. An "agentish" system runs a predetermined path and calls it intelligence. The first can be governed and audited; the second is automation with a costume. The buyer's filter is simple: ask what the agent decides that a fixed workflow could not, and what it does when unsure. Thin answers mean the "agent" is a script.
By publishing "agentic vs agentish" and warning against agent-washing, Nintex sets a bar it must then clear, a reason to take its governed, auditable framing more seriously, not less.
| Ask the vendor | Agentic | Agentish |
|---|---|---|
| What does it decide? | Judgement that changes the path | Nothing a rule couldn't |
| What if it's unsure? | Escalates / asks / defers | Proceeds anyway |
| Can you audit it? | Yes, in the process | Opaque |
Sources: Nintex, "agentic vs agentish", 9 Apr 2026.
Four layers that make an agent safe to deploy
Put the pieces together and a picture emerges of what makes this deployable in a regulated firm. It is not one feature. It is four layers that each answer a different objection.
| Layer | What it provides | Objection it answers |
|---|---|---|
| Sovereignty | Local execution, no external calls | "Our data can't leave." |
| Determinism | Fixed workflow rails | "It has to behave the same way every time." |
| Auditability | Every decision logged in-process | "We must explain it to a regulator." |
| Human oversight | Escalate, clarify, bring a person in | "A person must own the risky calls." |
Read down the right-hand column. Those four objections are, in our experience, exactly the ones that stop an agentic-AI programme at the risk committee. Each layer removes one.
Remove all four and the conversation changes from "can we do this safely?" to "which process first?" That shift is where the commercial case turns, and it is why the governance story is not a footnote to the sovereignty story but half of it.
Sovereignty gets you into the room. Determinism, auditability and human oversight let you stay. A local model without governance would clear the data-residency objection and fail every other one.
Sources: Nintex, 24 Sep 2025 and 29 Apr 2026; Volge analysis, 2026.
Governed, restated
A sovereign model that could not explain itself would trade one liability for another. Agentic Business Orchestration prevents that: deterministic rails, AI judgement inside them, human off-ramps, and a full audit trail: all in the boundary.
Nintex's public stance against agent-washing raises the bar it has to clear, and the four-layer stack is how it clears it. Sovereignty opens the door; governance is why a risk committee lets you walk through.
Governed and auditable, agents that escalate, ask for clarification, or bring a person into the process.
Turn the page
Claims and architecture are one thing; a track record is another. Part IV puts the K2 lineage on a timeline, presents the one hard customer outcome we will stand behind, and lays out how the software is actually deployed.
Sources: Nintex, 24 Sep 2025.
K2 was not retired. It was reinforced
A common misreading holds that "K2" is a discontinued brand. The opposite is true. Nintex removed the name, reconsidered, and put it back, and it now heads the current on-prem line.
The confusion is understandable. In 2023, with version 5.7, Nintex dropped "K2" and rebranded the on-premises product simply "Nintex Automation". To an outside observer that looks like a brand being retired ahead of a product being wound down.
It was not. In September 2024, with version 5.8, Nintex deliberately re-added "K2" to the name. The current product is "Nintex Automation K2", under the umbrella "Nintex Automation On-Prem (formerly K2)". The name came back on purpose.
The reinvestment is the point. A vendor abandoning a product does not ship a reinforced release with a built-in local AI model two years later. Nintex did exactly that: version 5.9 in December 2025, then 5.9.1 on 29 April 2026, the release this book is about.
For a buyer, this matters as a signal of commitment. The on-premises line is not a maintenance backwater kept alive for laggards. It is receiving the flagship capability, sovereign agentic AI, first.
K2 is the current, reinforced name of Nintex's on-premises platform. It was removed in 2023, re-added deliberately in 2024, and now heads version 5.9.1 (29 Apr 2026). Anyone telling you K2 is deprecated is working from stale information.
Sources: Nintex product-naming history, 2023–2026; K2 5.9.1 release, 29 Apr 2026.
From K2 blackpearl to a sovereign engine
The shape of the line matters more than any single point on it. An acquired platform, briefly rebranded, then given its name back and its most advanced capability first, is not a product in decline. It is one a vendor has chosen to build on.
Sources: Nintex, 2020–2026.
A regulated asset manager, 75% faster
We will stand behind one named result. Sarasin & Partners, a regulated asset manager, cut its incident-reporting cycle time by three-quarters on K2 on-premises.
The detail worth dwelling on is not the headline percentage; it is the sector. Sarasin & Partners is a regulated asset manager, exactly the kind of organisation for which data-residency is a hard constraint rather than a preference.
The outcome, a 75% cut in incident-reporting cycle time (Nintex, 10 Jul 2025), was achieved on K2 on-premises. This is the proof point that the on-prem line delivers material process improvement without asking a regulated firm to relax its boundary.
We present one result rather than a wall of logos deliberately. A single, dated, sector-relevant outcome is more useful to a risk committee than a dozen unattributed claims.
If a regulated asset manager can compress a compliance-adjacent process by three-quarters while keeping data on-premises, the pattern generalises to any organisation with the same constraint: government, healthcare, defence, and regulated finance more broadly.
One customer, one process, one figure, and note the date. This result is from July 2025, on K2 on-premises, which puts it before the locally-hosted AI model this book is actually about (April 2026). It proves the on-prem platform delivers; it is not a test of the 2026 engine. We use it because it is the honest, dated evidence we have, and we would rather flag the gap than let the number do work it has not earned.
Sources: Nintex, Sarasin & Partners outcome, 10 Jul 2025.
Three ways to run it, one boundary to keep
Sovereignty is not synonymous with a server in your own basement. The platform deploys three ways, and each can keep the AI decision inside a boundary you control.
| Model | Where it runs | Best when |
|---|---|---|
| On-premises | Your own data centre | Maximum control; air-gapped or classified networks |
| Private cloud | Your dedicated cloud tenancy | Control with elasticity; single-tenant isolation |
| Managed hosting | Hosted, managed on your behalf | You want the boundary without running the metal |
The choice is a control-versus-effort trade, not a sovereignty trade. On-premises gives the most control and suits air-gapped sites. Managed hosting hands the operational burden to someone else while keeping the data in a defined boundary.
The point of listing three is that "sovereign" need not mean "you run everything yourself". It means the data and the AI decision stay where your rules apply.
Choosing among these three, sizing them, and wiring identity is exactly the work Volge does. The next part turns the argument into a plan.
Sources: Nintex, 2026.
Not a first customer's risk
A sovereign-AI decision carries less career risk on a platform thousands of organisations already run. Scale is not proof of quality, but it is proof you will not be the pioneer taking every arrow.
Read the exhibit with the caveat in the source line: the bars are not on a shared scale and are there to hold the figures, not to be measured against one another. The substance is the numbers. Seven thousand organisations, a hundred-plus countries, one regulated proof point, three ways to deploy.
Put the four parts together and the case is complete in outline: the problem is real (Part I), the engine answers it (Part II), it is governed (Part III), and it is proven and buyable (Part IV). What remains is turning that into a decision, which is Part V.
Turn the page
Part V is the operator's read: the questions to ask before buying, an honest account of what could go wrong, and what a disciplined first deployment looks like.
Sources: Nintex, 24 Sep 2025; 10 Jul 2025; 2026.
What Nintex says about its own direction
K2 (5.9.1) is our first step in bringing AI directly into the platform in a way that fits how our customers operate today.
The barrier for regulated firms hasn't been interest, it's been how to apply it responsibly: by keeping AI within the boundaries they already trust.
We quote the CPTO because his framing is the honest one. "First step" is not the language of a finished revolution; it is the language of a vendor shipping a real, bounded capability and saying so. The value for a regulated buyer is precisely that modesty.
We do not name a current Nintex chief executive here. The former CEO left the company effective 2 February 2026; any 2025 remarks are attributable to the "then-CEO, September 2025", and we have not relied on them for the claims in this book.
Sources: Nintex, CPTO remarks on K2 5.9.1, 2026.
Eight questions worth asking first
A short interrogation separates a sound sovereign-AI deployment from an expensive disappointment. Ask these before signing, of us and of Nintex.
| # | Question | What a good answer looks like |
|---|---|---|
| 1 | Which processes actually need local AI? | A shortlist bounded by data-residency rules, not everything |
| 2 | What exactly stays in-boundary during normal operation? | The routine AI decision; update/licensing paths named separately |
| 3 | How are updates delivered to an air-gapped site? | A defined, offline-capable patch path |
| 4 | How is the model sized to peak volume? | Capacity plan tied to decision throughput |
| 5 | Where are escalation thresholds set? | Designed, documented, reviewed, not defaulted |
| 6 | What does the audit trail capture? | Input, decision, human step, routing: all in-boundary |
| 7 | Which deployment model, and why? | On-prem / private cloud / managed, matched to your control needs |
| 8 | How does identity integrate? | OIDC + SCIM into your existing directory |
Questions 3 and 5. Air-gapped update paths and escalation-threshold design are where sovereign-AI programmes quietly fail. Both are solvable, and both need attention before go-live, not after.
Sources: Volge, 2026, informed by Nintex K2 5.9.1 documentation.
Where this goes wrong, and how to avoid it
A brief that promised only upside would deserve the scepticism it earned. Here are the failure modes we have seen or expect, and the mitigation for each.
| Failure mode | Why it happens | Mitigation |
|---|---|---|
| Rubber-stamp oversight | Humans lack time or context to review | Design thresholds so escalations are rare and meaningful |
| Under-sized model | Peak decision volume underestimated | Capacity plan with headroom; load-test before go-live |
| Stale air-gapped model | No offline update path defined | Establish an offline patch cadence up front |
| Scope creep past the boundary | A tempting external feature gets bolted on | Keep an explicit in-boundary policy; review additions |
| Agentish deployment | Automating a script and calling it an agent | Apply the agentic-vs-agentish filter (Exhibit 12) |
None of these is a reason not to proceed. Each is a reason to proceed with a design phase rather than a demo and a purchase order. The technology removes the data-residency objection; it does not remove the need to engineer.
We sell delivery, so we would say a design phase matters. We also believe it, because the programmes that skip it are the ones we are later called in to rescue.
Sources: Volge analysis, 2026.
From objection to first governed process
A disciplined first deployment does not begin with a platform-wide rollout. It begins with one process, done properly, that proves the pattern to the people who must sign off the rest.
| Phase | Focus | Output |
|---|---|---|
| Weeks 1–3 | Scope & risk review | Shortlisted process; boundary and scope confirmed with risk/compliance |
| Weeks 3–6 | Design | Deployment model, sizing, identity, escalation thresholds, audit plan |
| Weeks 6–10 | Build & integrate | Local AI decisioning wired into one governed workflow |
| Weeks 10–13 | Validate & hand over | Load-tested, audited, defensible to a risk committee |
The deliverable at the end is not a slideware pilot. It is one real process, running local AI decisions inside your boundary, with the audit trail and human oversight a regulator would expect. That is the artefact that opens the way to the second process and the tenth.
Thirteen weeks is a shape, not a promise. Approval cycles in regulated firms are the usual long pole. We would rather set that expectation now than miss it later.
Sources: Volge, 2026.
Glossary
Plain definitions for the terms this book leans on. Where a term is Volge's editorial framing rather than a vendor claim, we say so.
- Agentic Business Orchestration
- Nintex's model combining deterministic workflows, AI decisioning and human-in-the-loop, governed and auditable. Unveiled 24 September 2025.
- Agent-washing
- Dressing ordinary automation in agentic language. Nintex cautioned against it in "agentic vs agentish", 9 April 2026.
- Air-gapped
- A network with no connection to external systems. Volge editorial framing of the setting K2's local engine suits.
- Data egress
- Movement of data out of a controlled boundary to an external system. The problem local execution removes.
- Deterministic workflow
- A process whose steps and rules execute the same way every time, providing the rails an AI decision sits inside.
- Human-in-the-loop
- A design where a person can be brought into a decision. Nintex: agents "escalate, ask for clarification, or bring a person into the process".
- K2
- The on-premises platform lineage, now the current name in "Nintex Automation K2". Stripped in 2023, deliberately re-added in 2024.
- Locally-hosted AI model
- A model that runs inside the customer environment. In K2 5.9.1 it makes routing, prioritisation and issue-detection decisions with no external API calls during normal operation.
- Nintex Automation On-Prem (formerly K2)
- The umbrella name for the on-premises product line.
- OIDC / SCIM
- Open standards for authentication (OIDC) and identity provisioning (SCIM). Simplified in K2 5.9.1.
- Off the hyperscalers
- Volge editorial framing for running AI without dependence on a public-cloud provider's model API.
- Own, don't rent
- Volge editorial framing for keeping the model and the decision inside your own boundary rather than consuming them as an external service.
Sources: Nintex, 2024–2026; Volge editorial framing noted inline.
Sources & dates
Every material claim about the sovereignty thesis traces to one of the entries below; platform and customer figures are attributed inline where they appear. Comparative and framing statements attributed to Volge are our analysis, not vendor claims.
- Nintex acquires K2 Software: acquisition completed 14 October 2020.
- K2 lineage: K2 blackpearl → K2 Five (2017), the on-premises workflow line.
- Rebrand to "Nintex Automation": "K2" stripped from the name, v5.7, 2023.
- "K2" re-added: deliberately restored, v5.8, 5 September 2024.
- Agentic Business Orchestration unveiled: deterministic workflows + AI decisioning + human-in-the-loop, governed and auditable; agents can "escalate, ask for clarification, or bring a person into the process"; 7,000+ organisations across 100+ countries. Nintex, 24 September 2025.
- Sarasin & Partners outcome: regulated asset manager cut incident-reporting cycle time 75% on K2 on-premises. Nintex, 10 July 2025.
- "Nintex Automation K2" 5.9: reinforced on-prem release, 2 December 2025.
- Nintex Automation K2 5.9.1: built-in, locally-hosted AI model; "AI-driven decisions, including routing, prioritization, and issue detection, run entirely within the customer environment, with no external dependencies required"; sentiment and severity in SmartForms and workflows; simplified OIDC + SCIM; WCAG improvements; STS, support to 31 March 2027. Released 29 April 2026.
- K2 5.9.1 product page: "No cloud dependency. No external API calls. Just AI-driven decisions running inside your environment." Nintex, 2026.
- CPTO remarks: Niranjan Vijayaragavan on K2 5.9.1: "first step in bringing AI directly into the platform…"; the barrier "hasn't been interest, it's been how to apply it responsibly", solved by keeping AI "within the boundaries they already trust." Nintex, 2026.
- "Agentic vs agentish": Nintex warns against "agent-washing", 9 April 2026.
- Deployment options: on-premises, private cloud, managed hosting. Nintex, 2026.
- Leadership note: former CEO Amit Mathradas departed for Five9, effective 2 February 2026; September 2025 remarks attributable as "then-CEO, September 2025". No current CEO named in this book.
- Volge analysis: all named-vendor comparisons (Microsoft Copilot Studio, AWS Bedrock Agents, Google Vertex AI Agents), governance-overhead and risk exhibits, and the framings "air-gapped", "off the hyperscalers" and "own, don't rent" are Volge's assessment, 2026, not Nintex claims.
Figures and quotations are drawn from the dated primary materials above and the named regulatory instruments.
What you need in place to run governed agents
Agents assume the most underneath them. Two lists follow, the Nintex elements to licence, and the non-Nintex pieces that must already be in place, because an agent with no governed workflow, no identity, and nothing to read is a demo, not a system. Tiers and figures are indicative for 2026; confirm current SKUs with Nintex.
| Nintex Agent Designer | Builds the agents (supervisor and multi-agent). Open beta on Automation CE and Nintex Workflow (2026); GA at Professional tier and above, later in 2026. |
|---|---|
| Nintex Orchestration | The layer agents run inside. Open beta (2026). |
| Automation CE / Workflow | The underlying platform. Agents run inside governed workflows, so the Build & run layer must be licensed first. |
| Nintex Automation K2 5.9.1 | On-premises, locally-hosted agentic AI, runs in your environment, no external calls. Separate on-prem licence. |
| Professional tier or above | Required for the GA agentic features. |
| A workflow footprint | An agent needs a governed process to act within: build or map it first. |
|---|---|
| On-prem infrastructure | For local agents: Windows Server, SQL Server, and compute for on-prem inference. |
| Agent identity & access | Microsoft Entra Agent ID or equivalent, each agent needs a first-class identity, least privilege, and revocation. |
| Grounding data | The systems and documents the agents read, with access controls in place. |
| Human-in-the-loop policy | Named owners, an approval and escalation path, and compliance sign-off for autonomous actions. |
Sources: Nintex product and release materials, 2025–2026 (Agent Designer availability per Nintex Q2 2026 update); Volge editorial framing noted inline.
One accountable team, from strategy to shipped software
Volge delivers agentic automation on the Nintex platform, in cloud, local, or mixed topologies, through one team and one arc. Four roles hold the work from the first workshop to operational handover.
One agentic approach, governed wherever the model runs
The question is not whether agents, but where they run and who governs them. The same five phases answer it for the Nintex platform, in cloud, local, or mixed topologies.