At a Southeast Asian manufacturer running just-in-time inventory, an invoice-matching agent on Business Central was approved in January as a 90-day pilot. Day 90 landed on month-end close, so the review moved. At the rescheduled review the sponsor said the results were trending the right way, the vendor offered a two-quarter extension at the same rate, and nobody in the room could name the number that would have made the answer no. The extension got signed. In October the controller sees the agent's consumption on the invoice again and asks who approved the spend. Nobody did. The pilot renewed itself.
A pilot with no stop rule is a subscription. It ends when someone chooses to end it, and nobody holds that choice. The sponsor carries the upside of continuing and pays nothing for waiting. The controller sees the invoice and has no authority over a project someone else sponsored.
AI pilot stop criteria are a decision rule written before launch. They name who is allowed to end a pilot, the review date, the KPI and threshold that force the decision, and what gets switched off on a stop. A pilot stop rule blocks a move to production. A production stop rule retires live AI that no longer earns its cost.
A pilot with no stop criteria renews by default
The assumption inside the pilot charter is that a failing pilot fades out by itself: the sponsor moves on, the budget runs out. Neither is a decision, and neither arrives on a schedule. The pilot keeps running on its last approval.
Extending a pilot takes no meeting and no number. Stopping one takes both, plus someone willing to tell a sponsor that the project missed. With that asymmetry, extension wins any review where the rules are silent.
Two things make a stop real: a date and a number. A name makes it enforceable. Your AI Pilot Succeeded. Your Operations Didn't. covers why pilots stall after a technical success. Its fifth root cause, success never defined in operational terms, is where this piece starts. Assume you have the metric. The open question is who acts when the metric says no, and by when.
You can spot a subscription without opening a dashboard. Look for a review date that has moved at least once, and a KPI that appears in the vendor's deck but nowhere in the approval record. If the invoice also posts with no named approver, the stop decision has no owner.
Gartner's 2026 CIO survey: stop criteria separate the vanguard
Gartner's 19 October 2026 release reports that only 13% of CIOs say their enterprise has realized significant or transformational value from AI so far, and that "the likelihood of use case success remains one-in-five." The data comes from Gartner's 2027 CIO and Technology Executive Agenda survey: 2,002 CIOs and technology executives, May to July 2026. The release states no revenue band for respondents, so the figures describe CIOs in general, with no mid-market cut.
Gartner labels CIOs who deliver better business outcomes while reporting high CEO-perceived AI value "vanguards." Organizations with guidelines for when an AI pilot should not move to production are 2.3 times more likely to be vanguards. Organizations with guidelines for abandoning low-value AI already in production are 2.6 times more likely. Only 31% and 17% of CIOs, respectively, have those practices.
That is a correlation from one survey, and a stop rule does not manufacture a vanguard. The pairing does locate the discipline: the vanguards are more likely to have decided in advance what failure looks like.
Alicia Mullery, VP Analyst at Gartner, frames the stakes: "Volume is not value. AI is creating a reckoning for organizations that lack fundamental disciplines, such as defining use-case value ahead of time and establishing criteria for when to walk away."
Defining use-case value ahead of time is your KPI threshold. Establishing criteria for when to walk away is your stop rule. The first earns a slide in the pilot proposal. The second needs a named person willing to sign it, which makes it the harder one to write.
The stop-rule gate: who decides, on what date, against which KPI
Write five fields into the pilot approval before the first credit is consumed:
- One KPI, one threshold, in the pilot's own units. Invoices posted without correction, exceptions closed inside a business day, whatever the agent was bought to move. If the threshold needs a footnote, it is not a threshold.
- A review date on two calendars. The sponsor's and the controller's, set on approval day. A slipped review defaults to stop. Extension is the action that needs a signature.
- The two-period trigger. Two consecutive periods under threshold force the decision ahead of the review date. One miss is noise or ramp-up. Two in a row is a result.
- One named approver per outcome. Continue to production, extend with a new KPI, or stop. A continue or an extend takes the sponsor's signature and the controller's. A stop takes either one.
- The switch-off list. Which agent to disable in Business Central, which connector access to revoke, which credit budget to release, and who holds the admin rights to do it. Write it before launch, because on the day of a stop nobody remembers which connector the pilot used.
A slipped review defaults to stop, and extension needs a signature. Every other field exists to make that default survivable.
At the review date, or after a second consecutive miss, the gate runs the five stages any governed decision runs:
| Stage | What the gate does | What it replaces |
|---|---|---|
| Signal | The pilot or agent reaches its review date, or misses its agreed KPI two periods running. | A review someone remembers to schedule. |
| Route | The decision goes to the pilot sponsor and the finance controller together. | A status note to the sponsor alone. |
| Approve | A named approver signs one of three outcomes: continue to production, extend with a new KPI, or stop. | An extension that renews because nobody signed anything else. |
| Execute | On stop: disable the agent in Business Central, revoke connector access, release the credit budget set aside for it. | Access and consumption that keep running because switching them off is nobody's task. |
| Audit | KPI against target, the decision, the named approver, credits consumed to date. | A slide in the next steering deck. |
The audit line carries four facts, and the last one prices the next decision. A controller who sees credits consumed to date can weigh an extension against what the pilot has already cost. An extension signed without that line commits money against a number nobody has read.
This gate is a different control from the one in Your AI Agent Has an Approval Gate. It Doesn't Have a Stop Button. That article covers halting an agent mid-run, a power held by whoever notices the problem. The stop rule ends an agent's tenure on a date, and the sponsor and the controller hold it. An agent can pass every safety review and still fail its stop rule.
Production stop rules for AI that is already switched on
Pilots get a review date because someone planned a pilot. Live AI gets none. Check the agent that went live a year ago. Does anyone still review its output every week? Does its cost sit on its own line, or inside run-rate?
A production stop rule reuses the gate and changes two fields. The review recurs on a cycle the controller owns, such as quarter close, instead of arriving once. And the KPI gains a cost term: the outcome the agent produces divided by the consumption it took to produce it. Consumption varies by run, so the same agent can pass the quarter's outcome target and fail on cost per outcome.
Before anyone stops a live agent, check the fallback. The safe-state argument in the stop-button article applies here: if the manual process has decayed, stopping the agent creates an outage, and an owner who knows that will not pull the trigger. A production stop rule without a rehearsed manual path stays on paper.
What a stop rule releases: the shared Copilot Credit pool
Gartner's cost data explains why the release matters. 58% of CIOs report increased pressure to deliver cost improvement from AI, and 51% expect AI to increase the total cost of ownership of technology solutions. Mullery puts the cost problem this way: "AI costs are probabilistic and determined at run time." Costs set at run time have no ceiling in the contract. The ceiling gets written somewhere, and the stop rule is where.
On Business Central the meter is Copilot Credits. Microsoft's Copilot Credits licensing guidance says the credits "pool at the tenant level" and are consumed by both prebuilt and custom Dynamics 365 agents, so they "should be allocated to environments to manage usage across workloads." Pay-as-you-go credits cost $0.01 each, and Microsoft states that pay-as-you-go billing means "no interruptions due to overages." ERP Software Blog, a trade blog covering Business Central licensing, reported in October 2026 that "nothing is included free with your Business Central license" and that Copilot Studio agents, SharePoint agents and other Microsoft services on the same tenant draw from the same balance. A tenant without pay-as-you-go billing is the exception: the same blog reports that when capacity is depleted, "the AI capability becomes unavailable until more capacity is added." Our OpsGrid vs. Copilot comparison for Business Central sets out Microsoft's own billing terms.
A pilot sitting in review limbo keeps drawing from that balance while it waits for a decision. On pay-as-you-go that shows up as spend the controller never approved. On a prepaid balance it shows up as less capacity for another team's agent. The Execute step of the gate turns the pilot's draw off and releases the credit budget the controller reserved for it. Because the pool is tenant-wide, a stop avoids further spend and frees headroom instead of producing a refund, and it gives the controller a closed line to report upward.
The rework side of AI cost, what it takes to undo an agent's unverified output, is covered in AI Agent Costs: The Real Bill Isn't Tokens, It's Rework. The stop rule catches the case where consumption and rework together outrun the value, whichever line shows it first.
Frequently asked questions
What are AI pilot stop criteria?
AI pilot stop criteria are a decision rule written before launch. They name who is allowed to end a pilot, the review date, the KPI and threshold that force the decision, and what gets switched off on a stop. Gartner's 2026 CIO survey found that organizations with guidelines for when an AI pilot should not move to production are 2.3 times more likely to be vanguards, and only 31% of CIOs have them.
Who should decide when an AI pilot stops?
The pilot sponsor and the finance controller, together. The sponsor holds the upside of continuing and the controller holds the invoice, so neither decides well alone. Route the decision to both at the review date, name one approver for each outcome (continue to production, extend with a new KPI, or stop), and let a slipped review default to stop rather than to extend.
What is the difference between a pilot stop rule and a production stop rule?
A pilot stop rule decides whether a pilot moves to production. A production stop rule decides when to abandon AI that is already live and no longer earns its cost. Gartner reports that organizations with guidelines for abandoning low-value AI already in production are 2.6 times more likely to be vanguards, and only 17% of CIOs have them. The gate is the same. The production version runs on a recurring cycle and adds a cost term to the KPI.
What does stopping a Business Central agent pilot release?
Three things: the agent is disabled in Business Central, the connector access it was granted is revoked, and the credit budget the controller reserved for it is released. Copilot Credits pool at the tenant level, according to Microsoft's licensing guidance, so a stop avoids further spend and frees headroom rather than producing a refund. The audit line records the KPI against target, the decision, the named approver and the credits consumed to date.
Where the stop rule runs once it works
Run the gate on a calendar invite and a signed line in the pilot charter for the first two reviews. A gate that survives two reviews earns software. Decision infrastructure is the layer that routes a decision to named owners, holds the approval, executes against the system of record and keeps the record. OpsGrid, in active beta for Dynamics 365 Business Central, is built to carry the route, approve, execute and audit stages of a gate shaped like this one. It does not ship a stop-rule gate as a packaged feature today. Disabling an agent and revoking a connector stay admin actions in the platforms that hold them: the gate decides, the admin acts, and the audit line records both.
Twelve questions score the five stages this gate runs on. Take it with your oldest pilot in mind.
Take the diagnostic →