NEWMake Viral Reels turns any link into a ready-to-post reel.Try it free →
AI for Enterprise · Playbook

How to get an ROI number for an AI project you can defend

Sooner or later someone in finance asks what your AI project is worth. You give a number. They ask where it came from. If your answer is "the vendor's estimate" or "we assumed a productivity gain," the conversation is over, and so is your budget.

The problem is rarely that the project has no value. The number was built from the top down: a big percentage applied to a big salary bill. Nobody can check a number like that, so nobody trusts it.

A number you can defend is built from the bottom up. It starts with one piece of work, measured before and after, with every cost counted. Here is how to build it.

Step 1: Pick one unit of work

A unit of work is one thing that gets done many times and can be counted. One invoice matched. One customer reply sent. One contract summarized.

"Improve the productivity of the support team" is not a unit. You cannot count it, so you cannot measure it. "One customer complaint answered" is a unit. You know how many you handle a month. You can time one.

A good unit passes three tests:

  • It happens often. Hundreds or thousands of times a month, so you have enough cases to measure.
  • It has a clear start and end. You can say when the clock starts and when it stops.
  • Someone already tracks it. A ticket count, a queue report, a case log. Finance trusts numbers that come from a system they already know.

Your whole ROI case is the cost of one unit today, the cost of one unit with AI, and the number of units. Keep it that small.

Step 2: Baseline it before you build

Measure how the work is done today, before anyone writes a prompt. Once the tool is live, you cannot go back and measure the old way. People forget how long things took, and their guesses lean toward whatever makes the project look good.

For two to four weeks, record four things:

  • Volume. How many units a month.
  • Handling time. Minutes of a person's hands-on time per unit. Time a real sample. Do not ask people to estimate.
  • Error rate. How many units out of 100 have to be redone, corrected, or escalated.
  • Cost of an hour. Ask finance for the loaded hourly cost they use for that team. Use their figure, not yours.

Sample across the whole month, including the hard cases and the busy days. A baseline taken on easy cases will make your result look worse than it is. One taken on hard cases will make it look better. Both will be found out.

Step 3: Count the review time

This is the line most ROI cases leave out, and it is the first thing a careful reviewer will look for.

The model produces a draft in seconds. That is not the time per unit. A person still has to read the draft, check it against the source, fix what is wrong, and sometimes throw it away and start again. All of that is handling time.

So measure the new process the same way you measured the old one: hands-on minutes from start to finish, per unit, on real cases. Include the cases where the draft was discarded. Include the time people spend re-prompting. For an illustrative case: if the old way took 12 minutes and the new way takes 6 minutes of review, your saving is 6 minutes. It is not 12 minutes minus the few seconds the model took.

Watch for review that gets faster over time. Sometimes people have learned the tool. Sometimes they have stopped checking. If review time drops, look at the error rate before you count the gain.

Step 4: Add the run costs and the cost of errors

Savings are half the sum. Now count what the system costs to run every month:

  • Model usage. What you pay per request, at real volume, with real document lengths.
  • Licences and hosting. The platform, the storage, the monitoring tools.
  • Upkeep. The hours someone spends each month on prompts, test cases, and fixes when a model or a source system changes. Leaving this out is the usual mistake.
  • The build. A one-time cost, kept separate so you can work out payback.

Then price the errors. Compare the error rate with AI against your baseline rate. If it went up, each extra error has a cost: rework time, a credit to a customer, a senior person pulled in. Multiply and subtract it. If the error rate went down, that is a real gain and you can count it. Either way, show the line.

Step 5: Compare against the simple alternative

Finance will not compare your project with doing nothing. They will compare it with the cheapest thing that might work. Do that comparison yourself, first.

Ask what a plain fix would deliver. Better templates. A rule that routes cases to the right person. A cleaner form that stops bad inputs. A fix to the one step where people retype data. Estimate that saving and that cost with the same method.

The honest ROI of the AI project is the gain over that alternative, not the gain over today's process. Sometimes the simple fix gets you most of the way.

A worked example

Illustrative numbers: everything below is made up to show the method. None of it is a benchmark or a result from a real project.

The unit is one customer complaint reply. The team sends 4,000 a month. The baseline shows 12 minutes of handling time per reply. Finance gives a loaded cost of $40 an hour.

With AI drafting the reply, measured review and editing time is 6 minutes. So:

  • Time saved: 6 minutes x 4,000 replies = 24,000 minutes, or 400 hours a month.
  • Gross saving: 400 hours x $40 = $16,000 a month.
  • Run costs: $1,200 model usage, $800 licences and hosting, $1,200 upkeep. Total $3,200 a month.
  • Cost of errors: replies needing rework rise from 2 in 100 to 3 in 100. That is 40 extra cases a month at about $45 each in rework and credits, or $1,800.
  • Net saving: $16,000 - $3,200 - $1,800 = $11,000 a month.
  • Payback: the build cost $90,000, so a little over eight months.

Now look at what the demo version of this case would have said. The model writes a draft in about 30 seconds, so the saving is 11.5 minutes per reply, or about $30,700 a month, with no costs against it. That is nearly three times the number you can defend.

Then run the alternative. Say better templates would cut handling time from 12 minutes to 9, for a one-time cost of $10,000. That saves 200 hours, or $8,000 a month. The AI project now adds $3,000 a month over the templates, for $80,000 more. Payback on that extra spend is over two years.

In this made-up case, the right move may be to ship the templates first and re-measure. The method gave you that answer before finance did.

How to present the number so it holds up

  • Show the arithmetic on one page. Units, minutes, rate, costs, net. Anyone should be able to redo it with a calculator.
  • Label what is measured and what is assumed. Next to each figure, write where it came from: the queue report, the timing sample, the vendor quote.
  • Give a range. Show a low case, an expected case, and a high case. Say which one input moves the result most. It is usually review time.
  • Say what happens to the hours. Saved hours are not cash by themselves. Tell finance what they turn into: a backlog cleared, overtime cut, a hire you no longer need, faster replies. If the answer is capacity and not cash, call it capacity.
  • Commit to a re-measure date. Promise to run the same measurement 90 days after go-live and report it, whatever it shows.

Before the meeting, give the page to someone in finance and ask them to find the holes. Every hole they find in private is one you will not be asked about in the room.

Before you quote a number

Check that you have one unit of work, a baseline taken before the build, review time measured on real cases, run costs and error costs subtracted, and a comparison with the simple alternative. If one of those is missing, that is your next piece of work.

A defended number is usually smaller than the one in the pitch deck. It is also the one that survives the meeting, and the one that will still be true a year from now.

If you want more writing like this, the Hyperautomation AI Report is a free newsletter, three times a week, with practical notes on building and using AI. You can subscribe on the newsletter page.

Talk to me about AI in your company

If you're working through an AI rollout, tell me what you're trying to do. I read every message and reply by email.

Or connect with me on LinkedIn.