Write This Before Any AI Project Starts

Before any AI project starts, founders need one page that names the metric, baseline, target, and timeline. Without it, you are not managing an investment.

A founder I know approved a $180,000 AI contract to reduce customer support ticket resolution time. Eight months later, the vendor presented a slide deck full of upward-trending lines. Nobody in the room could agree on whether the project had worked. The founder had never written down what "worked" meant.

What RAND found when AI projects collapsed

The RAND Corporation interviewed industry stakeholders across real AI deployments and identified the root causes of failure. The findings did not point at model performance or compute budgets. They pointed upstream: unclear problem framing and misunderstood leadership roles in setting direction. The Standish Group's decades of project data confirm the same pattern across technology investments broadly. Unclear requirements and misaligned stakeholders predict failure. The AI context does not change this. It amplifies it, because AI outputs are ambiguous enough to support almost any post-hoc interpretation.

That is the trap. When you approve an AI project without a written success definition, you are not leaving room for flexibility. You are creating the conditions for a story to be assembled after the fact, shaped by whoever has the most to gain from a favorable reading.

The four things that belong on one page

The success definition document does not need to be long. It needs four things written down before any build begins.

The metric is the specific number the project is supposed to move. Not "improve customer support" but "reduce average ticket resolution time measured in hours per ticket." One metric. If your team argues for two, pick one and put the second in a footnote.

The baseline is the current measured value of that metric. Not an estimate. Not a number from a comparable company. The actual current value from your own data. If you do not have it, measuring it is the first task of the project, not a detail to sort out later.

The target is the specific value you expect to reach. The Workmate ROI framework is explicit here: net present value and internal rate of return calculations require a projected target established before deployment. A target written after the results are in is not a target. It is a caption.

The timeline is the date by which you will evaluate whether the target was reached. Not a range. A date.

When no baseline exists before deployment

A reasonable objection to this structure is that many founders approving AI projects for the first time are deploying into processes they have never measured. A sales team that has never tracked follow-up rates within 24 hours has no baseline follow-up rate. Asking that founder to write a baseline into the document produces a number invented for the occasion, which is arguably worse than nothing.

This objection is real for a narrow category of work. It does not hold for the operational and product AI investments most founders are actually approving. If you cannot state the current value of the metric the AI is supposed to move, that is evidence the problem is not framed clearly enough to approve yet. The RAND findings name unclear problem framing as the upstream failure, not the absence of a document. The document is just the forcing function that surfaces the framing problem before you spend money on it.

The shared mental models research adds a second failure mode for the exploratory approach. Teams that begin without shared knowledge structures about goals and roles produce lower process quality and worse outcomes than teams that begin with them. Deferring the success definition to post-deployment learning means your team builds for weeks without a shared understanding of what they are building toward.

What the document actually does

I have seen founders treat this as a bureaucratic exercise and produce documents nobody reads. That is a real failure mode, and I will not pretend the document is self-executing.

The document's function is not filing. It is alignment. When the founder, the engineering lead, and the vendor all sign the same page before work begins, the shared mental models research suggests that team process quality improves and performance outcomes follow. When they do not, the post-project conversation becomes a negotiation over whose version of events is correct.

Write the metric, the baseline, the target, and the date. Get three signatures. Keep it to one page. If you cannot fill in all four fields before approving the project, the project is not ready to approve.