Why AI projects drift
AI work changes as it develops, often more than traditional IT projects. Data turns out to be weaker than expected. A supplier changes its model, its price or its terms. Users find shortcuts, or stop using the tool when it gets something wrong. Results in daily work differ from those in the demonstration.
None of this is a reason to avoid AI. It is a reason for oversight designed for change, rather than a single approval followed by silence.
A green status without a definition provides more reassurance than insight.
What the board actually needs to see
The board does not need technical detail. It needs a short, regular report that connects the initiative back to the decision it approved.
The objective as originally approved, in one sentence.
The baseline agreed at the start, so progress has a reference point.
The observed outcome so far, measured the same way as the baseline.
Costs to date against budget, including the ones that were not in the original case.
Quality and incidents: errors, complaints, near misses and what was done about them.
The next decision the board will be asked to take, and when.
That is usually one page. It is enough to make the conversation better, and short enough that it actually gets read.
Give the colours a meaning
Traffic lights are useful only when everyone agrees what triggers each colour. Before the first report, agree the thresholds.
Green means the agreed measure is on track for this period.
Amber means a named deviation, with an owner and a date for a decision.
Red means the board will be asked whether to continue, change course or stop.
State the measurement period and the source next to each figure. "Adoption: 80%" means little until you know 80% of whom, measured how, over which weeks.
Make stopping a real option
Good oversight includes the possibility of ending an initiative. If stopping is never discussed, sunk costs quietly become the reason to continue.
Ask for stop criteria at approval: which results, by which date, would lead the organisation to pause or end the work? Revisit them in each report. An initiative that is stopped on time, with lessons recorded, is a governance success, not a failure.
Fictional example. A pilot for AI-assisted case handling is approved with a target of 20% shorter handling time after three months, quality maintained. At three months, time is down 8% and complaints have risen slightly. Because the criteria were agreed, the conversation is not about whether the project "feels" successful. It is a clear decision: adjust the process and review again in six weeks, or stop.
A rhythm that works
For most boards, a short written update each quarter is enough, with a fuller discussion once or twice a year and immediate escalation for material incidents. The rhythm matters more than the format. Oversight that happens predictably catches drift early. Oversight that happens only after a problem catches it late.
Three questions for your next meeting
- For our most important AI initiative, what did we approve, and what is the result so far against the baseline?
- What exactly would turn the status amber or red?
- Which result, by which date, would lead us to change course or stop?