Gartner published a finding this week that should change how every AI steering committee in Hong Kong runs its next meeting. The organisations getting the most value from AI are not the ones launching the most pilots. They are the ones that decided, in writing and in advance, when they would stop.
According to the Gartner 2027 CIO and Technology Executive Agenda survey of 2,002 CIOs, released on 19 October 2026, only 13% say their enterprise has realised significant or transformational value from AI. Companies with formal rules for halting weak pilots were 2.3 times more likely to be in the top-performing group. Yet only 31% of CIOs have such rules.
This guide explains what AI stopping rules are, why they are becoming a board-level control, and how to write them for your own portfolio before the next budget cycle.
What are AI stopping rules?
AI stopping rules are pre-agreed, written criteria that decide when an AI initiative must be paused, redesigned or shut down. They are set before a pilot starts, tied to a measurable value hypothesis, and owned by a named business sponsor. They turn "should we keep going?" from a debate into a test.
Think of them as the AI equivalent of a stop-loss order. An investment committee would never approve a trading strategy without a maximum drawdown. Stopping rules apply the same discipline to AI spend, where costs and benefits are far less predictable than in a conventional IT project.
Gartner's research identifies two moments where stopping rules matter most:
--- Pilot gate: criteria for deciding that a pilot should not move into production.
--- Production gate: criteria for retiring an AI system that is already live but no longer earning its keep.
Organisations with production stopping rules were 2.6 times more likely to be top performers, but only 17% of CIOs report having them. That gap is the opportunity.
Why are AI stopping rules a board issue in 2026?
Boards are shifting from asking how many AI projects exist to asking which ones pay back. Gartner reports that 44% of CEOs struggle to demonstrate AI financial returns to stakeholders, while 58% of CIOs face more pressure to deliver cost improvement. Stopping rules are the mechanism that protects returns.
The Gartner 2027 CEO Survey of 203 CEOs, published on 20 October 2026, found that 75% plan to increase technology investment. More money is coming. The question is whether it flows to initiatives that work or gets trapped in programmes nobody is willing to end.
Gartner VP Analyst Alicia Mullery framed it directly at the Orlando Symposium: "Volume is not value." Her point was that AI is exposing organisations that lack two basic disciplines, defining use-case value ahead of time and agreeing in advance when to walk away.
For Hong Kong leaders, there is a governance layer too. The Privacy Commissioner's Model Personal Data Protection Framework for AI expects continuous monitoring and review of AI systems. A documented stopping rule is one of the clearest ways to show that review is real rather than ceremonial.
Why do AI pilots keep running after they stop making sense?
Weak AI pilots survive because nobody defined failure at the start. Without a written threshold, every disappointing result can be explained away as early data, a training issue or a model that will improve next quarter. Sunk cost and sponsor reputation then keep the project alive long after the evidence has turned.
Three forces are at work in most organisations:
--- Moving goalposts. A pilot that was meant to cut handling time by 30% is quietly reframed as an "innovation learning exercise" once it delivers 8%.
--- Probabilistic costs. Gartner notes that AI costs are determined at run time, unlike predictable SaaS licences. 51% of CIOs expect AI to increase total cost of ownership. Pilots that looked cheap at 50 users can become expensive at 500.
--- No one owns the off switch. IT can report metrics, but only the business can declare that value has not arrived. If that authority is not assigned, the default is to continue.
The result is a portfolio that grows in count but not in contribution, which is exactly the pattern behind Gartner's one-in-five use-case success rate.
How do you write an AI stopping rule before a pilot starts?
A usable stopping rule has four parts: a value hypothesis with a number, a kill threshold below which the pilot ends, a fixed time box, and a named decision owner. Write all four into the pilot charter and have the business sponsor and finance sign it before any build work begins.
The four-part stopping rule framework
--- 1. Value hypothesis. State the business outcome in measurable terms. Example: "Reduce average claims triage time from 22 minutes to 12 minutes for motor claims."
--- 2. Kill threshold. Set the minimum acceptable result. Example: "If triage time is above 17 minutes at week 10, the pilot does not proceed to production." Include a cost ceiling, such as a maximum monthly inference spend per 1,000 cases.
--- 3. Time box. Fix the evaluation date. Eight to twelve weeks is typical for a contained workflow. Extensions require a new written case, not a verbal agreement.
--- 4. Decision owner. Name the executive who must make the go or stop call, and the date they must make it.
Add one more element if the pilot touches personal data: a privacy stop condition. If the system cannot meet your PDPO obligations, such as data minimisation or access controls, it stops regardless of performance.
What should a production stopping rule look like?
A production stopping rule monitors whether a live AI system still earns its cost. It tracks unit economics, adoption, output quality and risk on a fixed review cycle, usually quarterly. If two or more indicators breach their thresholds for consecutive reviews, the system is redesigned or retired.
Production rules matter more than pilot rules because the money is larger and the inertia is stronger. Once a tool is embedded in daily work, switching it off feels disruptive even when it is no longer worth the spend.
Four indicators to track for every live AI system
--- Unit economics: cost per transaction, case or document, compared with the pre-AI baseline.
--- Adoption: the percentage of eligible staff or cases actually using the system each month.
--- Quality: error rate, rework rate or human override rate on AI outputs.
--- Risk: incidents, complaints or compliance exceptions linked to the system.
Pair these with clear accountability. Our earlier piece on who is accountable when an AI agent fails sets out how to assign that ownership at board level.
How do AI stopping rules work in a Hong Kong enterprise?
In practice, stopping rules turn a quarterly AI review into a short decision meeting. Each initiative is scored against its pre-agreed thresholds, and the committee confirms continue, fix or stop. Freed budget is then redeployed to the initiatives that cleared their bar.
Consider a 400-person Hong Kong logistics group running three AI pilots: a customer enquiry assistant, an invoice-matching tool and a route-planning model.
--- The enquiry assistant set a goal of resolving 40% of WhatsApp enquiries without staff. At week ten it resolves 46%. It proceeds.
--- The invoice-matching tool aimed for 90% auto-match. It reaches 71%, below its 80% kill threshold, because supplier invoices arrive in inconsistent formats. It stops, and the team redirects effort to standardising supplier data first.
--- The route-planning model hits its accuracy target, but inference costs run 2.4 times the agreed ceiling. It moves to a fix track with a six-week deadline to cut costs or close.
No one has to argue. The thresholds were agreed in advance, so the decisions take minutes. The CFO sees a portfolio that is actively managed, which makes the next funding request far easier to approve.
What mistakes do leaders make with AI stopping rules?
The most common mistakes are setting thresholds after results arrive, measuring activity instead of outcomes, leaving the off switch with IT, and treating a stop as a failure. Each one weakens the discipline and lets weak initiatives survive. A stop decision is a successful control, not a failed project.
--- Retrofitting thresholds. Rules written after the data comes in are rationalisations, not rules.
--- Counting usage, not value. Prompts sent or logins per week say nothing about business impact. Gartner explicitly advises measuring unit economics, decision velocity and customer outcomes instead.
--- No business owner. If only the technology team can stop a project, stopping becomes a political act rather than a governance step.
--- Punishing the stop. If sponsors are penalised when a pilot is closed, they will stop proposing honest thresholds. Reward clear decisions in either direction.
--- Ignoring run-time cost. A model that works but costs too much at scale still fails the business case.
How do you present AI stopping rules to your board or CFO?
Present stopping rules as capital discipline, not technology process. Show the board one page listing each AI initiative, its value hypothesis, kill threshold, current status and decision date. This proves the portfolio is governed and gives the CFO confidence that new AI funding will not become open-ended spend.
Key facts to cite in your board paper
--- Only 13% of CIOs report significant or transformational AI value (Gartner, October 2026).
--- AI use-case success remains roughly one in five (Gartner, October 2026).
--- Organisations with pilot stopping rules are 2.3 times more likely to be top performers; only 31% have them.
--- Organisations with production stopping rules are 2.6 times more likely to be top performers; only 17% have them.
--- 44% of CEOs struggle to demonstrate AI financial returns (Gartner 2027 CEO Survey).
Frame the ask simply: approve the framework, approve the review cadence, and approve the principle that money released from stopped initiatives is reinvested in those that cleared their thresholds.
The strategic takeaway
AI stopping rules are not about doing less AI. They are about making sure the AI you keep is the AI that pays. Gartner's data is clear that the organisations pulling ahead are those willing to define value early and walk away when it does not arrive.
Start with your current pilots. Write the four-part rule for each one this month, assign a decision owner and set a date. Within one quarter you will know which initiatives deserve the next round of budget.
We understand AI. We understand you. With UD by your side, AI never feels cold.
Find the AI Initiatives Worth Funding
Now that you have the framework, the next step is identifying which of your workflows are ready for AI and which are not. We'll walk you through every step, from AI readiness assessment to pilot design, stopping rules and performance tracking, backed by 28 years of enterprise experience in Hong Kong.
Reviewed by the UD enterprise AI team.