An AI project that cannot be written as one line on a P&L is not a project yet. It is an idea. The fastest way to turn it into a project is to put four numbers you already have into one formula, test the weakest of them on real work for an afternoon, and only then decide whether to fund it. That is the whole method. The rest of this note is how to do it without inventing a single statistic.
The short answer
Scope the project to hours a year, then dollars a year, using your own headcount, your own time, your own cost rates and a share you have measured rather than assumed. Agree that number with whoever owns the budget before anyone writes code. Then build to it, measure against it, and stop if it does not show up.
People × hours a week each spends on the task × 48 weeks × share AI takes on = hours a year
Hours a year × fully loaded cost per hour = the annual number. Twenty people, five hours a week, 48 weeks and a 30% share is 1,440 hours a year; at $75 an hour that is $108,000 a year. The calculator on our home page runs exactly this with your inputs and sends nothing anywhere.
Why most AI business cases fail the CFO
Having sat on both sides of the board table, I have watched the same three mistakes sink otherwise sensible proposals.
- Borrowed numbers. A vendor's benchmark or an analyst's industry average tells a CFO nothing about this company's cost base. The moment a slide says "companies like yours save 40%", the finance team stops listening, and they are right to.
- No baseline. If nobody measured how long the task takes today, nobody can prove it got faster. "Faster" without a before is a feeling, not a result.
- Scope that grows with the demo. The pilot was one workflow; the proposal is a platform. Operating partners see this constantly in portfolio companies: the number was never fixed, so it keeps moving.
The fix for all three is the same: a number built only from inputs the company owns, agreed before the work starts, and measured the same way after.
The four inputs, and where each one comes from
- People doing the task Count the people, not the department. If eight analysts pull financials and four of them do it weekly, the number is four for that task. Your org chart and a ten-minute conversation with the manager give you this.
- Hours a week each spends on it Ask the people, then check one week of calendars or timesheets. Round down. A conservative input here is what makes the final number believable in the room.
- Fully loaded cost per hour Salary plus benefits plus overhead, divided by hours worked. Finance already has this figure for every role; use theirs, not a guess.
- Share of the work AI can take on This is the only input you cannot look up. It is also the one everyone wants to assume. Do not assume it. Measure it with a test run on real work (section 6), and set it below what the test shows.
Forty-eight working weeks covers holidays and sickness without pretending people work every week of the year. If your business runs on a different rhythm, use your real figure.
The share AI can take on is the only input you cannot look up. It is also the only one you can test for the price of an afternoon.
Worked example: a public venture lender
Horizon Technology Finance is a publicly traded venture lender. Its task was prospect research: for each batch of target companies, an analyst pulled the financials (market cap, debt, cash), scored the company against the firm's lending criteria, found the CEO and CFO, and drafted a first-touch letter. The baseline, measured before anything was built, was six to nine analyst hours per 100 companies.
That baseline is the whole business case in one line. Hours per 100 companies, times the number of batches a year, times the analyst's loaded rate, is the ceiling on what the project can be worth. No industry figure needed. No assumption about "productivity" needed. The firm knew its own number before we wrote a line of code.
The build was an AI prospect analyzer that does the same four steps and delivers Excel and Word output the team already used. After it shipped, the same work took about 15 minutes per 100 companies. The result is stated in the same unit as the baseline, which is the point: a CFO can put the before and the after on one line and see the difference without a translation layer.
Two things to notice. First, the number was analyst time, not "revenue impact", because time was what could be measured honestly on day one. Second, nobody claimed the analysts were eliminated. The hours moved to the part of the job that needs judgment: deciding which of the scored companies to call. The P&L line is freed hours; what the firm does with them is a management decision, and it should be written down as one.
When the number is a count, not hours
Not every AI project saves time. Some are meant to bring in customers, and "revenue uplift" is the easiest number in the world to invent. The discipline is the same: pick a count you can measure before and after, with the same method, on a date.
For Chris David Salon in Delray Beach the count was simple. Ask ten AI assistants who to book for hair color in the area and count how many name the salon. Before the work: 0 of 10. After the site was rebuilt as an AI Front Door: 9 of 10. Alongside it, two numbers from systems the salon does not control: website visitors went from 110 to 651 a month, and the salon ranked #4 of 245 local salons for its core search. All three were measured in June 2026 against Google Analytics, Google ranking data and AI-citation tracking.
Notice what is not in that paragraph: a dollar figure for new bookings. The salon's owner can attach one, because he knows his average ticket and how many new clients walked in. We do not publish it, because we did not measure it. A scoped number is one you can stand behind in front of an auditor, and a count from a third-party system is easier to stand behind than a revenue estimate that depends on attribution.
The test run before the build
The fourth input, the share of work AI can take on, is where business cases go to die. Here is how to measure it instead of assuming it. Take a real batch of the work, run it through a quick version of the system, and time it. For the venture lender, the test scored 10 real target companies and drafted five personalized letters in 46 seconds. That single run told the firm three things: the output was usable, the time saving was not marginal, and the full build was worth funding.
A test run costs an afternoon. It replaces the slide that says "we estimate 60% automation" with a sentence that says "we ran 10 and here is what came out". Operating partners who want to roll one playbook across a portfolio should insist on this step in every company, because it is the only part of the case that transfers: the headcount and rates differ by company, but a test on real work is a test on real work anywhere.
- Use real inputs, not sample data. The point is to find out what breaks.
- Have the person who does the job today judge the output. If the analyst would not send the letter, the share is lower than it looks.
- Set the share below what the test shows. If the test suggests 80%, scope at 50% and let the result beat the case rather than the case beat the result.
What to do this week
- Pick one task, not a department. The one with the most repeated, well-defined hours. Prospect research, intake, reconciliation, first-draft reporting.
- Measure the baseline. People, hours a week, loaded rate. Write the three numbers down with a date.
- Run a test on ten real items. Time it, have the owner of the task grade the output, and set the share from what you saw.
- Agree the number with finance. Hours a year and dollars a year, from the formula, in writing, before the build starts.
- Build to the number and measure the same way after. Our engagements are scoped on one 30-minute call, built with your own team, and handed over as a working system in about eight weeks on average. The after-measurement uses the same unit as the before.
If you want the number tested before you commit to anything, book 30 minutes. You leave with the formula filled in for your task and three plays you can run this quarter, whether or not you hire us. If you would rather start with the arithmetic alone, the calculator runs in your browser and sends nothing to us.
Questions CFOs and operating partners ask
What is the formula for scoping an AI project to a P&L number?
People doing the work, times hours a week each spends on it, times 48 working weeks, times the share of that work AI can take on, gives hours a year. Hours a year times a fully loaded hourly cost gives the annual value. Twenty people, five hours a week, 48 weeks and a 30% share is 1,440 hours a year; at $75 an hour that is $108,000 a year. Every input is yours, not an industry average.
How do you know the share of work AI can take on before building anything?
You measure it with a small test run on real work before you fund the build. For Horizon Technology Finance the test scored 10 real target companies and drafted five personalized letters in 46 seconds, which is what justified the full build that took prospect research from 6 to 9 analyst hours per 100 companies to about 15 minutes.
What if the value is revenue rather than hours?
Then the number is a count you can measure before and after, not an estimate. For Chris David Salon in Delray Beach the number was how many of 10 AI assistants named the salon when asked who to book: 0 of 10 before, 9 of 10 after, with website visitors going from 110 to 651 a month and a #4 of 245 local ranking, measured in June 2026.
How long does it take to get from the number to a working system?
ISO Vision AI engagements are scoped to the number on one 30-minute call, built with the client's own team, and handed over as a working system in about eight weeks on average.
Stuart Kerr
Founder & CEO of ISO Vision AI. Two IPOs, $14B+ in value created; operating roles at Oracle, Apple & NeXT; five board seats including Azuga's $400M sale to Bridgestone.