The most expensive AI projects we have seen were not the ones that failed. They were the ones that took eighteen months to admit they had failed, because nobody had agreed in advance what failure would look like.
The commitment problem
AI initiatives attract executive attention, and executive attention makes them hard to stop. Six months in, with a sponsor who has described the project in a board paper, 'this is not working' becomes a career conversation rather than a technical one.
The solution is not better judgement later. It is agreeing the criteria before anyone is invested — while stopping is still a neutral outcome rather than an admission.
Decide what failure looks like while failing is still an acceptable answer.
What the gates actually are
| Stage | Gate question | Abandon if |
|---|---|---|
| Opportunity mapping | Is there a use case with quantifiable value? | No candidate clears the value threshold |
| Data feasibility | Can we reach the data lawfully, at quality? | Retrieval quality on a real sample is below the floor |
| Proof of value | Does it work well enough for real users? | Evaluation metrics miss gates after two improvement cycles |
| Production readiness | Can it be operated safely and affordably? | Cost per interaction or risk profile is unacceptable |
Each gate has a named decision-maker and a scheduled date agreed at kick-off.
The dates matter as much as the criteria. A gate without a date drifts, and a gate that drifts is not a gate.
Where projects actually die
In our experience the second gate — data feasibility — accounts for most legitimate abandonments, and it arrives in week four. That is a good outcome: four weeks of work to discover that the data required is not reachable within the privacy constraints is a cheap and correct answer.
Without the gate, the same discovery happens in month seven, after a build has been commissioned on the assumption that the data problem would be solved by someone else.
The objection we hear
'Building in an exit signals a lack of confidence.' We understand the instinct, and it is backwards. Explicit gates are what allow an organisation to run several AI initiatives at once, because each has a bounded downside.
Without gates the rational strategy is to run one initiative very cautiously, which is a much worse portfolio position than running four with honest kill criteria.
Gates are not a prediction of failure. They are how the sponsor keeps the option to redirect budget to the initiative that is working, which is a much stronger position than defending one that is not.
What happens when a gate fails
Not much drama, if it was set up properly. The team writes up what was learned — usually a genuinely valuable artefact about data readiness — the budget returns to the portfolio, and the next candidate use case starts.
We have had engagements end at gate two and turn into a data platform engagement instead, because the constraint identified was the real problem worth solving. That is a good outcome for everyone, and it only happens when stopping is cheap.
Running an AI portfolio without gates?
We can help you define the criteria and the decision rights before the next initiative starts, which is the only time it is easy.