Judging AI proposals and pilot reports

Why this is where it comes together

Sooner or later a proposal or pilot report lands on your desk asking you to approve, fund or extend an AI project. It may come from your own staff, a consultant or a supplier. It will usually be positive. Your decision commits public money and, often, affects citizens. This module gives you a structured way to read such documents, drawing on everything in the earlier modules.

Start with the problem

Before looking at the technology, check that the proposal is clear about the problem (module 2):

• What problem does it solve and for whom?
• How big is the problem today? How was that measured?
• Was the root cause examined? Or was AI assumed to be the answer?
• Were simpler options considered, such as a process change or better guidance?

A proposal that cannot describe the problem precisely is not ready for a decision.

Examine the evidence

Look closely at what the evidence actually shows.

• What was compared? Results should be measured against current practice, not against doing nothing.
• On what material? Was the test run on real work like yours, including difficult cases, Urdu and regional languages and poor-quality documents?
• How many cases? A handful of examples is an anecdote, not evidence.
• Who measured? Results reported only by the supplier or by the team that wants the project to continue, need independent checking.
• What failed? A report that mentions no failures has probably not looked hard enough.

Be wary of benchmark scores and demonstration results presented as proof. They tell you how a system performs on someone else's test, not on your work.

Check the measures

Good measures were agreed before the pilot started, not chosen afterwards. Look for a balance of:

• Quality: accuracy, error rates and the kinds of errors made
• Efficiency: time saved, after counting the time staff spend checking
• Service: effect on citizens, such as waiting times or complaints
• Fairness: whether results differ across groups, districts or languages
• Staff experience: whether users found it helpful and trustworthy

Time saved is the most common headline and the most commonly overstated. Ask whether checking time was included and whether savings were real or only estimated.

Test the costs

Ask for the full cost over several years (module 1), including data preparation, integration, training, checking time, usage charges, monitoring and exit. Compare it with the value of the benefits that were actually measured, not those that are hoped for. Ask what happens to cost if use grows and what the price is likely to be at renewal.

Review the risks

Check that the proposal has:

• classified the risk honestly and, for higher risks, completed an impact assessment (module 3)
• named the senior owner, operational owner and decision makers
• shown how human oversight will work in practice, with realistic caseloads
• addressed data location, protection and ownership (modules 4 and 5)
• planned how citizens will be told and how complaints will be handled (module 6)
• set triggers for suspension and a plan for monitoring model drift

Insist on an exit plan

Every AI project should be able to stop. Ask:

• How would we stop using this system and what would replace it?
• Can we take our data and records with us in a usable form?
• How long would exit take and what would it cost?
• Who decides to stop and on what evidence?

A project with no exit plan is a commitment you may not be able to undo.

Making the decision

You have four options. Each is legitimate:

1. Proceed as proposed, with conditions you set.
2. Extend the pilot to answer specific open questions.
3. Change course, for example to a lower-risk use or a different approach.
4. Stop, recording what was learned.

Write down your reasons. Stopping a project that did not prove its value is good stewardship, not failure. The how-to guide on running a small, safe pilot may help your teams design the next stage.

Try it

Write a one-page checklist for reviewing an AI proposal, using the headings in this module: problem, evidence, measures, costs, risks and exit. Under each heading, write the two questions you consider most important. Keep it for use when the next proposal arrives.

Check yourself

1. Why should results be compared with current practice instead of with doing nothing?
2. Why is "time saved" often overstated?
3. What four decisions can you make on a proposal or pilot report?

Answers: (1) Doing nothing is not the real alternative; only a comparison with current practice shows whether AI adds value. (2) Checking time is often left out and savings may be estimated instead of measured. (3) Proceed with conditions, extend the pilot, change course or stop.