The written ROI commitment: our template, annotated

The commitment we ask for, ourselves included, is one page with six lines: the metric, the baseline with its dates and method, the target with its date, the method after go-live, the review dates, and what happens if the target misses. This guide is that page, line by line, with what each line is for.

What is a written ROI commitment?

A page that turns "this will pay for itself" into a measurement with a date and a consequence. It names one number, says what that number is today and how it was counted, says what it must become and by when, says how it will be counted after go-live, and says what the vendor does if it misses. Both parties sign it before money changes hands.

It exists because AI projects fail quietly. Nobody cancels them; they just stop being measured, and after a year nobody can say whether they worked. A page with a baseline and a review date cannot fade. On the review date, someone reads the number, and the number is either above the line or below it.

The template is short on purpose. Six lines fit on one page, and one page gets read. Once it is published, the annotated template as a file will be at Link coming soon; what follows is the same template as prose.

What goes on the metric line?

Line 1. One number, in a unit someone in the office would recognize, from a system that already records it. Missed calls answered per month, from the phone log. Quote turnaround in hours, from the request timestamp to the sent timestamp in ServiceTitan or Jobber. Board build time in minutes per day, from the dispatcher's own clock. Dormant customers rebooked per month, from HubSpot or the scheduling software.

The annotation: one metric. A project that moves two numbers gets two commitments on two pages, because a single page with two numbers invites averaging, and an average of a hit and a miss is a story. The metric is also chosen for what it is not: not revenue, which moves for a hundred reasons the project does not control, and not satisfaction, which cannot be read from a system. Revenue is what the metric is worth; the worksheet that turns the metric into dollars is a separate page, and it is an estimate, and it says so.

Where people get it wrong: the metric is chosen for how good it will sound at the review rather than for how directly the project moves it. A voice agent is measured on calls answered and jobs booked from those calls, because that is what it does; measuring it on monthly revenue is measuring the weather.

How is the baseline written?

Line 2. The number the metric shows today, the period it was counted over with both dates, and the method used to count it, in enough detail that a stranger could repeat it. Where we have a signed commitment on a named deployment, its baseline line reads Number coming soon; yours reads the same way with your numbers.

The annotation: the baseline is taken before anything is built, from the system the metric lives in, over a period long enough to be typical. If the metric is seasonal, the period says which season, and the target is set against the same season. If the system cannot produce the number, the first deliverable of the project is the log, field or report that can, and the commitment is signed once the baseline exists. A baseline from memory, from a sales call, or from after go-live is not a baseline, and a commitment written on one is void from the start.

Where people get it wrong: the baseline is taken over the vendor's favorite week, or over the owner's worst month, and the review date reads a number that was never comparable.

How is the target set?

Line 3. The number the metric must reach, in the same unit, counted by the same method, and the date by which it must hold. Where we have a signed commitment on a named deployment, its target line reads Number coming soon.

The annotation: the target is proposed by the vendor from two things, the baseline and what comparable deployments measured, and it is argued over by the owner until both believe it. Where the proposed number came from is written under it, so the argument at signing is about evidence. The target is a number in the metric's unit, not a percentage of a feeling; a percentage of a baseline nobody trusts is a percentage of nothing. The miss margin on line 6 can be a share of that number, as ours is, but only once the number itself is written. And the target has a date, because a number that must hold "eventually" holds never.

Where people get it wrong: the target is the number that closes the sale. A vendor who proposes a target without asking for the baseline first has proposed a number from a brochure, and the miss is already written.

What does the method line say?

Line 4. Exactly how the metric will be read after go-live: the system, the report or the query, the period, who runs it, and who sees the result. Word for word the same method as the baseline, or the two numbers cannot be compared.

The annotation: the method is the least interesting line and the one that decides whether the page means anything. If the baseline counted missed calls from the phone system's log including after-hours, the method counts them from the same log including after-hours. If the method needs something that does not exist yet, a call log, a timestamp on requests, a field on the job, building it is the first deliverable, and the page says so. Where we have a signed commitment on a named deployment, its method line reads Agreed in the build plan.

Where people get it wrong: the baseline came from one system and the method reads another, and at the review the two sides bring two numbers and argue about which one is real.

When is the review?

Line 5. The reading dates. On our own projects they are the checkpoints the how we work page states, thirty, sixty and ninety days after go-live. The earlier readings come soon enough that a miss is seen while there is time to act on it; the last is the review, far enough out that a full measurement period has passed, when the number is read against the target. Every date names who is in the room, and all of them go in both calendars the day the page is signed.

The annotation: the earlier readings are not the review. Their job is to catch a method problem, a missing field, a log that stopped, before the review date makes the miss official. The review is a reading, not a meeting: the person named on line 4 runs the method, the number is written next to the target, and the page says what happens next.

Where people get it wrong: there is one date, it is far away, and by the time it arrives the people who signed have moved on and nobody runs the method.

What happens if it misses?

Line 6. The consequence, in plain words: what the vendor does, at whose cost, for how long, and what the owner can do if the remedy fails too. Where we have a signed commitment on a named deployment, its consequence line reads Agreed in the build plan.

The annotation: this is the line that makes the page a commitment instead of a forecast, and it is the one vendors most want to leave out. Our standing consequence is on the how we work page, under the measure step: a workflow that underperforms its projection by the margin stated there is redesigned at our cost. A usable consequence is one the vendor can deliver and the owner would accept: more build time at the vendor's cost until a second review date, a reduced monthly fee until the target holds, an exit with the data and the accounts handed over. A consequence the vendor cannot deliver is a lawsuit, and nobody signs up for a lawsuit. Write the remedy, its cost, its duration, and the exit.

Where people get it wrong: the consequence is "we will work with you to make it right," which is a sentence, not a consequence.

What is the worksheet for?

The seven lines of the worksheet are the six lines of the template plus the signatures, in the order they are written. Fill it in before the deposit, with the vendor, in one sitting. If a line cannot be filled, the project is not ready to be signed; the line that cannot be filled is the first thing to build. Bring the sheet to a working session and we will say what our side of each line would be for the project you have in mind, or which line stops it.