←
AI-Ready Foundation
Capable · M1 · lesson 1 of 5 · in progress
Preview — browse every lesson free. Enroll to mark lessons complete, open partner links and save your progress. Login & enroll →
AI Pilot Design: Baselines, Success Metrics, and the Go or No-Go Decision
📖
now learning

AI Pilot Design: Baselines, Success Metrics, and the Go or No-Go Decision

15 min

An AI pilot is supposed to end with a decision: scale it, stop it, or give it one more fixed period for a stated reason. Most never get there. They sit on a steering committee slide, permanently green, because nobody wrote down what number the pilot was meant to move, what that number was before it started, what result would justify the production bill, or what result would get it stopped. Good AI pilot design fixes all four before the first user logs in, on one page, and this lesson gives you that page.

The AI Pilot Problem: 37 Started, 3 Succeeded, and Nobody Can Say Which

Picture the decision meeting for a claims-triage pilot at a 2,000-person insurer (an invented scene, but you will recognize every line). Eight weeks are up. The slide says the assistant cut handling time by 22%. The vendor's account manager calls it a clear win. Then finance asks two questions: 22% against what, and how much would it need to be to pay for the rollout?

The claims lead says the old figure was "about twenty minutes". Nobody measured it. Nobody wrote down a target either, and nobody checked whether claims got easier over those eight weeks for reasons that had nothing to do with AI. Forty minutes later the room agrees to "keep going for another quarter and see". The pilot has just become permanent.

That meeting is the normal case. IDC surveys, reported by CIO.com on 5 December 2024, found that the average organization had launched 37 AI proofs of concept. About 5 reached production. 3 were judged successful.

Most people conclude from that funnel that AI pilots mostly fail. That reading uses the less interesting number. The sharper finding sat beside it: 30% of CIOs did not know what share of their proofs of concept had met their target KPIs, and nearly half either did not know whether the AI applications they had already put into production were successful, or said it was too early to tell.

It is not that the pilots lost. It is that nobody kept score. IDC's Daniel Saroff put it in one sentence:

"When we determine the proof of concept, we didn't have a measure of success".

Stopping 34 out of 37 experiments is not a scandal; experiments are meant to be cheap ways of finding out. The scandal is that the organization cannot tell which ones deserved to stop and which deserved the budget. A pilot is a decision instrument. If it cannot produce a decision, it was a demo with a budget.

The state where pilots are neither scaled nor stopped is called pilot purgatory, and it is caused by missing decisions, not by bad ones. Lesson 3, on choosing your first AI use case and avoiding pilot purgatory, stopped it at the selection stage with a scoring session and a kill date. This lesson stops it at the measurement stage.

Why do the published AI pilot failure rates disagree?

Because each study counts something different, which is itself the point of this lesson. The studies agree on the direction and disagree on the size. Do not average them, and do not quote one of them in a board paper without its definition attached.

Source and dateWhat it countedFinding
S&P Global Market Intelligence, reported by CIO Dive, 14 March 2025 (more than 1,000 respondents)Proofs of concept scrapped before production; companies abandoning most initiatives46% of proofs of concept scrapped on average; 42% of companies abandoned most of their AI initiatives in 2025, up from 17% in 2024
IDC with Lenovo, reported by CIO.com, 25 March 2025 (sample not stated)Proofs of concept that did not reach wide-scale deployment88%; of every 33, 4 graduated to production
Deloitte, State of AI in the Enterprise 2026, 21 January 2026Organizations that moved a large share of pilots into productionOnly 25% have moved 40% or more of their AI pilots into production
MIT NANDA, "The GenAI Divide", August 2025, as covered by GovInfoSecurity, 4 September 2025 and othersMeasurable profit-and-loss impact, a much stricter testWidely reported as 95% seeing zero return. Disputed, see below

A word on that last row, because you will be quoted it. The "95%" is contested. Critics point out that the method has not been disclosed in a way outsiders can check, that the 5% of successes refers narrowly to custom enterprise tools, and that a pilot which delivered some benefit but no sustained profit gain still counted as a failure. Oxford's Ajit Jaokar called it "a clever marketing gimmick" (GovInfoSecurity). We have read only the coverage, so treat the figure as a claim, not a fact.

Its real use is as a teaching example. Change the definition of success and the "failure rate" moves from 46% to 88% to 95%. If three research houses cannot agree on what a successful pilot is, your steering committee will not either, unless you write the definition down before you start. These figures are also dated: the newest is Deloitte's, from January 2026, and we found nothing more recent that measures pilot-to-production rates. The point they make has not aged, but quote them with their dates.

What does a pilot with no decision cost?

It costs the pilot budget, then a far larger sum nobody estimated. Gartner, in a press release of 29 July 2024, put the deployment approaches for generative AI at $5 million to $20 million, and analyst Rita Sallam added that "costs aren't as predictable as other technologies". (The same release forecast that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. That was a prediction, and we found no scoring of whether it came true.)

The December 2024 CIO.com article describes one proof of concept, an assistant for internal IT service requests, which EchoStor's Daniel Clydesdale-Cotter said could take "upwards of a million dollars potentially" to run in a data center. That changed the business case, and it was a gift, because it arrived early. Usually finance sees the production bill for the first time after the pilot "succeeds".

Why do organizations keep getting this wrong?

The reasons in the research are consistent, and none is about model quality. IDC's Ashish Nadkarni told CIO.com in March 2025 that "Most of these gen AI initiatives are born at the board level", often underfunded and without a strong business case. CSET's Mina Narayanan (GovInfoSecurity, September 2025) notes that pilots designed without the right people miss critical workflows. And the first of five root causes in RAND's report RR-A2680-1 (13 August 2024, 65 interviews) is that leadership misunderstands the problem to be solved and the metrics of success. All of that can be settled before launch, in a two-hour meeting.

How It Works: An AI Pilot Built to Produce a Decision

Four plain definitions first.

  • An AI proof of concept tests whether the technology can do the task at all, usually on sample data and without real users.
  • A pilot is a time-boxed test with real users on real work that exists to produce one decision: scale it, stop it, or extend it once for a stated reason. Many organizations call both a "POC", which blurs what decision each one owes.
  • A baseline is the measured value of your key number before the pilot starts. A comparison group is a set of comparable users doing the same work over the same weeks without the AI.
  • A go or no-go decision is the pre-agreed rule that turns the pilot's results into that decision.

The analogy: a pilot is a small clinical trial

Medicine solved this long ago. Before the first patient is enrolled in a clinical trial, a document called the protocol is written and signed, so nobody can look at the results first and then choose what success meant. The protocol fixes five things, and each has a direct twin in an AI pilot.

  • The endpoint. One outcome that counts, named in advance. If blood pressure did not move, you do not get to announce that patients felt more cheerful. Your twin is one business KPI.
  • The control arm. Patients on the standard treatment over the same months, so the trial measures the drug and not the season or the fact that everyone is being watched more closely. Your twin is the comparison group.
  • The smallest effect worth having. Trial designers decide in advance how big a benefit would justify the drug's cost and side effects. Your twin is a go threshold derived from the business case, not a round number.
  • Ethics approval. No patient is enrolled until an ethics committee has signed off on the risks. Your twin is the pre-pilot clearance.
  • The stopping rules and who decides. What ends the trial, and who makes the call. Your twin is the go, no-go and extend bands, and a named decision owner.

One more thing follows, and it matters most. A trial that ends with "the drug did not beat the control arm" is a completed trial, not a failed one: it answered its question and saved everyone a costlier mistake. A clear, evidenced no-go is your pilot doing its job.

The insurer's meeting at the top of this lesson had an endpoint of sorts, handling time, and broke every other rule: no measured baseline, no control arm, no threshold tied to cost, no stopping rule. The 22% might have been real. Nobody in the room could know.

What does a real pilot KPI look like?

A real pilot KPI is one operational number, measured against a baseline, over a stated period. The cleanest published example in our sources is from LinkedIn. In a paper posted to arXiv on 26 April 2024, its engineers reported that after deploying a retrieval system in the customer service team for about six months, median per-issue resolution time fell by 28.6% against the prior baseline. (It is company-authored, so read it as a pattern, not an audit.)

Copy the shape, not the size: one metric the business already cared about, a "before", a period. Compare "most pilot users found the assistant helpful", which tells you nothing about whether the work got faster, cheaper or better. The same shape works for pilots that are not question answering: invoices processed per hour for an extraction pilot, share of claims routed correctly first time for a triage pilot.

What should an AI pilot charter contain?

A pilot charter is a one-page agreement, signed before launch, on what the pilot will measure and how the result will be turned into a decision. We recommend eight parts. This is our synthesis of what the sources above say goes missing, plus the clearance checks a privacy leader will insist on; it is not a published standard, and we found no authoritative template.

  1. One business KPI with a measured baseline, and a comparison group where possible.
  2. A quality gate on the Lesson 8 golden set, so speed is never bought with wrong answers.
  3. A cost-per-task figure.
  4. A production cost estimate and the break-even it implies, done before the pilot.
  5. A clearance record: privacy, AI Act classification, disclosure and worker representation.
  6. A time box with the decision meeting already in the diary.
  7. A named decision owner: one person, not a committee.
  8. Go, no-go and extend conditions, as numbers, agreed before the start.

The Workshop: The One-Page AI Pilot Charter

This is the last lesson of Level 2 (AI Capable) on the five-level ladder, and it is where the previous four artifacts get cashed in. Lesson 6 gave you a data readiness scorecard, Lesson 7 a retrieval pipeline and its configuration sheet, Lesson 8 a golden test set and a calibrated judge. Lesson 9, on versioned prompts that survive model changes, gave you a prompt record under change control. The charter decides what happens once real users see the system, and the next lesson, LLMOps in practice: shipping, monitoring and rolling back AI in production, takes over the day the charter says go.

Who is in the room, and how long does it take?

Seat the people who own the answers the charter needs: the executive sponsor (the decision owner), the process owner (the KPI and its baseline), a finance business partner (the break-even arithmetic, and this seat is not optional), the AI or engineering lead (the quality gate and cost per task), the data protection officer or privacy lead with security (the clearance record), the HR or employee relations lead (worker monitoring and representation), and one frontline user, who will tell you the KPI you chose is gamed every Friday afternoon. For a vendor tool, add procurement for the contract terms.

The time is modest: a 2-hour charter session, a baseline period before launch, a 30-minute weekly review, and a 1-hour decision meeting fixed in advance. Start the clearance checks the day you book the session, because they slip most often. Bring the four earlier artifacts.

The eight steps

  1. State the decision, the owner and the date. One sentence: "This pilot exists so that [name] can decide whether to roll the assistant out to [the wider group] by [date]." The date fixes the time box. If you cannot name a single person, stop here. A pilot without a decision owner will not end.
  2. Clear the pilot before anyone logs in. Four checks, each signed by its owner and written into the charter. Privacy: if the pilot touches personal data, ask your data protection officer whether a data protection impact assessment (DPIA) is needed; under GDPR Article 35 one is required, before processing starts, where processing using new technologies is likely to result in a high risk to people's rights and freedoms. AI Act classification: record whether the use case is prohibited, high-risk under Annex III (areas such as employment, education and credit), or neither. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744 (in force 27 July 2026), Annex III duties apply from 2 December 2027, possibly while the system you pilot now is running. Disclosure: Article 50(1) of the AI Act, applying since 2 August 2026, requires providers to design AI systems that interact directly with people so that those people are told they are dealing with an AI, unless that is obvious from the context. Employee representation: if the pilot logs or measures staff, ask HR whether worker representatives must be consulted or must agree first. In Germany, section 87(1) no. 6 of the Works Constitution Act (as amended 19 July 2024) gives the works council co-determination over technical devices designed to monitor employee behaviour or performance. Then confirm data access for exactly the pilot users against your Lesson 6 scorecard. Confirm each point against the Official Journal text with legal; this is a checklist, not legal advice.
  3. Choose one business KPI. Pick a number the process owner already reports, so nobody has to be persuaded it matters. One. Then add at most two guardrail measures, which are numbers that must not get worse while the KPI improves (error rate or escalation rate, for example).
  4. Measure the baseline, and keep a comparison group where you can. Same population, same method, same data source you will use during the pilot. Write down the period and the count of cases. A baseline remembered ("it usually takes about twenty minutes") is not a baseline. Where you can, keep a concurrent comparison group: people on the same queue, over the same weeks, without the assistant. Judge the pilot on the difference between the two groups' change from baseline, not on the pilot group's before and after alone. Where you cannot, check that the baseline weeks match the pilot weeks for seasonality and case mix, and say so in the charter.
  5. Set the quality gate, including what happens to an early error. The golden-set pass rate, the regression suite result, and a plain definition of a critical error, checked in a weekly expert sample of live outputs. Then write down which weeks the gate is judged on and what an earlier critical error requires, or the decision meeting will argue about wording instead of evidence.
  6. Estimate cost twice, and derive the go threshold from it. First, cost per task during the pilot. Second, the full production cost: licenses, infrastructure, integration, support, and the people who will review outputs. Take it to finance now; if it kills the business case, you have saved the whole pilot budget. Then derive the go threshold: the KPI improvement at which the production cost estimate breaks even within the budget cycle, plus a safety margin. Write the arithmetic in the charter, so the threshold can be defended.
  7. Write the go, no-go and extend conditions as numbers. Limit "extend" to once, for a fixed period, with a named fix. The three bands must cover every possible outcome, so that no result leaves the room wondering.
  8. Run it, review weekly, and decide on the date. Check the KPI, gate and cost every week, and read a sample of real transcripts: the numbers tell you what happened, the transcripts tell you why. On the scheduled date, record the decision, the evidence and the reason (including for a no-go) where the next pilot team will find it.

How long should the time box be, and how many users do you need?

There is no authoritative standard for either, and we will not invent one. Collect enough real work to compare fairly with your baseline period, size the pilot and comparison groups alike, and fix the decision date on day one. The only reference point, about 90 days from pilot to deployment for mid-market firms against nine months or more at Fortune 500 companies, comes from the disputed MIT NANDA report, so do not lean on it.

A worked example: the warranty assistant

This is an invented illustration, carried through from the earlier lessons. A 900-person industrial equipment distributor has built a service desk assistant for warranty and parts questions, with a 40-task golden set and a 15-task regression suite. The pilot is chartered for 25 agents over 8 weeks, with 25 agents on the same queue as the comparison group. The VP of service is the decision owner. Both baselines cover the four weeks before launch, from the handling-time field the service desk already reports.

Clearance took two weeks. The queries contain customer names and serial numbers, so the data protection officer ran a short DPIA. The use case is not on the Annex III list, but the HR lead spotted a trap: using the logs to rank individual agents would be monitoring and evaluating the performance of workers, which Annex III point 4 lists as high-risk. So the logs feed team-level KPIs only. The distributor built the assistant, so it is the provider and the Article 50(1) duty is its own: the assistant's window says plainly that it is an AI system.

Finance wrote the threshold arithmetic into the charter. The full rollout covers 120 agents handling about 90,000 warranty queries a year at a baseline of 18 minutes each. At a loaded cost of $0.75 a minute, every 1% cut in handling time is worth $12,150 a year. The production estimate is $243,000 a year (licenses, infrastructure, integration spread over three years, support and output review), so break-even is 243,000 divided by 12,150, which is 20%. The room added five points of margin for the novelty effect fading and set go at 25%, and no-go below 10%, half of break-even, where no single fix is likely to close the gap. (Valuing saved minutes at loaded cost assumes the freed time is used; Lesson 19, on measuring AI ROI honestly, shows how to test that.)

The charter, with week 8 results:

Charter elementAgreed before launchResult at week 8
Business KPIMedian handling time per warranty query. Baseline: pilot group 18.0 minutes, comparison group 17.5 minutesPilot group 14.1 minutes (21.7% lower); comparison group 16.8 minutes (4.0% lower); net effect 17.7%
GuardrailEscalation rate must not rise above the baseline of 12%11%
Quality gateGolden set at least 36 of 40; regression 15 of 15; zero critical errors (a wrong warranty entitlement) in the weekly samples of 50 in weeks 5 to 8. Any critical error earlier must be fixed, added to the golden set and passing within one week37 of 40; 15 of 15; one critical error in week 3, fixed, added and passing in week 4; none in weeks 5 to 8. Met
ClearanceDPIA signed by the DPO; not high-risk, logs used for team KPIs only, agreed with worker representatives; AI disclosure in the assistant windowAll signed before launch
Cost per taskNo more than $0.20 per query$0.12
Production cost and break-even$243,000 a year; $12,150 per 1% of net reduction; break-even 20%Estimate unchanged
Time box8 weeks, decision meeting bookedHeld on the date
GoNet reduction at least 25% (break-even plus 5 points), all gate conditions metNot met: 17.7%
No-goNet reduction less than 10%, or any quality gate condition not met at week 8Not triggered
Extend (once, 4 weeks)Net reduction between 10% and 25%, gates met, and a named fix; if no fix can be named, the result is no-goTriggered

Walk through the decision meeting. The pilot group's handling time fell 21.7%. On a before-and-after view alone, that clears break-even: 21.7 times $12,150 is about $264,000 a year against a $243,000 cost. But the comparison group got 4.0% faster over the same weeks without the assistant (here, a quiet spell after a product recall closed). Net of that, the assistant's effect is 17.7%, worth about $215,000 a year. Without the comparison group, finance would have been told the rollout pays for itself. It does not, yet.

So the extend band fires. Decision recorded: extend once, for four weeks. The named fix comes from the transcripts: agents wait on slow part-number lookups, a retrieval problem routed back to the Lesson 7 pipeline. If the net reduction is still under 25% at week 12, the charter already says the decision is no-go.

Now notice what did not happen. Nobody argued about whether 17.7% was "pretty good" or proposed switching the headline metric to agent satisfaction. Nobody argued about the week 3 critical error either: the weekly sample caught a wrong warranty entitlement, the team fixed it and added it to the golden set, and the gate said exactly what that required. Without that sentence, half the room would have read "zero critical errors" as "zero, ever" and half as "zero at the end". The VP of service did not have to be brave. The protocol decided.

The vendor-tool variant: piloting someone else's product

Many pilots at this level are a vendor copilot or an AI feature inside a SaaS product you already pay for. Every row of the charter still applies. Five things change.

  • Read the data terms before the clearance step. Find out from the contract, not the sales deck, whether your data trains the vendor's models, how long inputs and outputs are retained, and where they are processed. Your DPO needs those answers.
  • Bring your own golden set as the acceptance test. Run your Lesson 8 golden set before and after the pilot, including the should-decline tasks that check it abstains when your sources hold no answer. The vendor's demo set measures the vendor's content, not yours.
  • Get the production price in writing at the start. Pilot pricing is often discounted. Break-even must use the full-rollout price at your expected volume.
  • Measure the KPI from your own systems. The vendor's dashboard counts logins. Your ticket system and finance data are the only sources the decision meeting accepts.
  • Write the exit before you enter. Agree what you get back on a no-go: your data, configuration and logs, and confirmation of deletion. If you are an EU financial entity and the tool would support a critical or important function, DORA (Regulation (EU) 2022/2554, applying since 17 January 2025) already requires, under Article 28(8), a documented and periodically tested exit strategy for it.

The AI Act roles flip too: the vendor is the provider and you are the deployer, so the Article 50(1) design duty sits with the vendor. Check that the disclosure actually appears in your configuration, and record it in the clearance row.

Your artifact: the Pilot Charter and Decision Record

You leave with the Pilot Charter and Decision Record: one page of charter plus a short decision note at the end. Fill the right-hand column in the room.

FieldWhat to write
Use case, decision and decision ownerOne sentence each; a named individual
Business KPIDefinition, source, baseline, period, count of cases; comparison group or why none
Guardrail measuresAt most two, each with its baseline
Quality gateThresholds, critical-error definition, sampling plan, weeks judged, rule for earlier errors
Clearance recordPrivacy or DPIA, AI Act class, disclosure, worker representation, data access; who signed, when
Cost per task targetA ceiling, in currency per task
Production cost and break-evenThe estimate, the date finance signed it off, and the arithmetic behind the go threshold
User group and time boxWho, how many, start date, end date, decision meeting date
Go, no-go and extend conditionsThree numeric bands that cover every outcome; extend allowed once
Vendor terms (vendor tools only)Data use and retention, production price, exit and deletion terms
Decision record and lessonsDecision, date, evidence, reasons; three lines at most for the next pilot

AI Pilot Failure Modes, How to Spot Them, and the Metrics to Track

Each has a tell you can check in under a minute against every live pilot.

Failure modeHow to spot it
No baselineThe "before" number is an estimate from memory, or was measured after launch
Before-and-after onlyOnly the pilot group was measured, with no check of season or case mix
Round-number thresholdThe go threshold has no break-even arithmetic behind it in the charter
Metric shoppingThe KPI on the final slide is not the KPI in the charter, or success is reported as logins or satisfaction
Ambiguous gateTwo readers disagree on whether a mid-pilot error breaks the gate
Uncleared launchThe privacy review, AI Act classification or disclosure check is dated after the first user logged in
Endless extensionThe pilot has passed its time box with no recorded decision, or has been extended more than once
Production cost surpriseFinance first sees a production estimate, or the vendor's full price, after the go decision
Portfolio sprawlThe number of live pilots exceeds the number of named decision owners with time to run them

On that last one, HBR authors Nathan Furr and Andrew Shipilov (reported by GovInfoSecurity, September 2025) argue that the real problem is resources spread thinly across unfocused pilots. Three pilots with three owners who turn up beat eleven with half-attentive sponsors.

How do you measure whether your pilot discipline is improving?

Track four numbers at portfolio level, and report them to the steering committee each quarter.

  • Decision rate. Pilots that ended with a recorded go, no-go or extend on the scheduled date, divided by pilots whose time box has ended. It is the best single measure of leaving purgatory.
  • Evidence coverage. Live pilots that had a measured baseline, a comparison group or a written reason for none, and a signed clearance record before launch, divided by all live pilots.
  • Cost variance. Actual cost per task and production estimate against what the charter said.
  • Pilot-to-production conversion, alongside the share of no-go decisions with written reasons. Report the two together.

One warning on that last metric. Do not set a target conversion rate. A portfolio where every pilot scales is not testing anything; it is rubber-stamping. What you want to rise is the decision rate, not the go rate.

What does your organization look like when Level 2 is done?

An organization that has worked through Lessons 6 to 10 has something it lacked at Level 1: a pilot register in which every entry has a charter, a decision and a reason. The steering committee sees decision rates, not green dots. Finance co-writes the go threshold instead of discovering the bill. The team can say "no-go" out loud without anyone losing face, because the charter said what no-go meant before anyone was invested. That is AI Capable: not that every pilot works, but that every pilot tells you something you can act on.

Frequently Asked Questions

What is the AI pilot failure rate?

There is no single figure: published rates run from 46% of proofs of concept scrapped (S&P Global Market Intelligence, reported March 2025) to 88% not reaching wide-scale deployment (IDC with Lenovo, reported March 2025), because each study defines failure differently. MIT NANDA's widely quoted 95% (August 2025) uses a stricter profit-and-loss test and is disputed, so do not quote it as fact.

What is the difference between an AI proof of concept and an AI pilot?

A proof of concept tests whether the technology can do the task at all, usually on sample data and without real users; a pilot tests whether it is worth doing, with real users on real work, and ends in a go, no-go or extend decision. Many organizations call both a "POC", which blurs what decision each one owes.

How do you set the go threshold for an AI pilot?

Derive it from the business case: find the KPI improvement at which the production cost estimate breaks even within the budget cycle, add a safety margin, and write the arithmetic into the charter. In the lesson's illustrative example, a $243,000 production cost and $12,150 of value per 1% of handling time put break-even at 20% and the go threshold at 25%.

Do you need a control group to judge an AI pilot?

Not always, but use one wherever you can: a concurrent comparison group doing the same work over the same weeks without the AI. Judge the pilot on the difference between the two groups' change from baseline; where no comparison group is possible, check that the baseline weeks match the pilot weeks for seasonality and case mix, and say so in the charter.

Should a vendor run the evaluation of its own AI pilot?

Let the vendor run the tool, but not the scoring. Measure the KPI from your own systems, test the tool against your own golden set (including tasks it should decline), and build the business case on the production price in writing, not the pilot discount.

Key Takeaways

  • An AI pilot is a decision instrument. If it cannot end in scale, stop or one fixed extension, it is a demo with a budget.
  • The damaging number is not how many pilots stop. It is that 30% of CIOs in IDC's surveys (CIO.com, December 2024) could not say how many of theirs met their KPIs.
  • Run the pilot like a small clinical trial: one endpoint, a measured baseline, and a concurrent comparison group wherever you can have one.
  • Derive the go threshold from break-even on the production cost, plus a margin, and write the arithmetic into the charter.
  • Clear the pilot before launch: privacy review or DPIA, AI Act classification, and Article 50(1) disclosure.
  • For a vendor tool, use your own golden set and KPI data, the production price and a written exit.
  • Track decision rate and evidence coverage across the portfolio, and never set a target conversion rate. A recorded no-go with reasons is a result.