Most return models are worthless, and they fail the same way: a percentage arrives that nobody in the room can trace back to a number the business already keeps. Return on an AI project is more art than science, and the art sits nowhere near the arithmetic. It sits in choosing one process, measuring it as it actually runs before anything changes, and refusing to borrow a benchmark for the part nobody has measured yet.
Proving the Return
How Do You Build an
AI ROI Business Case?
Four lines, one process, your own numbers. The worksheet finance checks, the claims it rejects first, and where every figure comes from.
Built from real buyer questions in our sales meetings
Direct Answer
Baseline one process, then measure the same process the same way after the change. Put the current cost in hours and error rate, state the post-AI cost using the identical measurement method, and show the delta net of license, hardware and the human review that remains. A case built on a generic productivity percentage does not survive finance. Bring the numbers your own operation already keeps, never a benchmark borrowed from another business.
The limit: every return here is a projection, and the largest input is an assumption rather than a measurement. Buyers described projects that open optimistic and reach a look-back where nobody can track what was accomplished. One put it directly: the cost model assumes every user talks to the AI twenty times a day, which would be fantastic adoption. Usage drives the whole worksheet, and we hold no completed before-and-after to calibrate it against. Give that assumption its own line, make it testable at ninety days, and expect the number to move when you test it.
Three things to settle before the model leaves your desk. Define the metrics you agree to be measured against. Set a baseline per process type rather than one average for the department. Carry the scope of every borrowed figure alongside the figure. The targeted questions below turn each into something written down.
Proving the return and paying for it are separate jobs, taken in that order. The method and the worksheet come first; the price, the license structure and the funding route are seven separate decisions downstream of the four lines, each linked from the line it belongs to.
The Four Lines Finance Actually Checks
A finance team reads a business case in four cells and a subtraction; everything else in the deck is commentary. Only your own operation can fill two of the cells. The other two are settled by decisions that each have a page of their own, so the third column names where each number comes from instead of a figure.
| Worksheet line | What goes in the cell | Where the number comes from |
|---|---|---|
| 1. Current cost baseline | Hours a week on one named process, the loaded rate of the people doing it, and what the mistakes cost. Manual processing was described to us at roughly a one-in-five error rate, so rework sits here too. | Your own operations record. No published source can supply it, and a borrowed baseline collapses at the first question. |
| 2. Post-change cost | The same process, measured the same way, after the change. Count the queries a person still handles and the work the model hands back. | A bounded pilot on your own documents, measured at the end of it rather than promised at the start. |
| 3. One-time and recurring cost | The license, the machine it runs on, preparing the data, and whatever running it consumes each month. | Four separate decisions: the cost page, the licensing page, the own-or-rent comparison and the running-cost page. |
| 4. Residual human review | Minutes per output times volume, held for the life of the deployment. Review never reaches zero; on the tasks that pay, it shrinks to a spot check. | Your pilot again, plus the adoption page for what happens when review swallows the saving. |
Two questions sit beside the worksheet rather than inside it. Where the money comes from is a funding decision taken after the delta is proven; for more information visit the funding page. When the invoice may be raised is a paperwork decision, and buyers repeatedly told us policy pays only once a milestone is met; visit the procurement page.
The Three Claims Finance Rejects First
When a finance leader refuses a business case, the refusal is rarely about AI. It lands on one of three moves inside the model, and buyers described all three in their own words. Each fix costs a line of transparency rather than a point of return.
- An assumption wearing the clothes of a measurement. A buyer read the model back to us and stopped at the usage rate: it assumes every user talks to the AI twenty times a day, which would be fantastic adoption. The saving scales with a number nobody has observed yet. Move it onto its own visible line and show the return at half the assumed rate.
- A saving borrowed from somebody else’s workload. Another buyer rejected an example outright: it is geared toward large-scale systems and repetitive querying, which may not be the challenge we have. A benchmark taken on high-volume retrieval says nothing about a team of twelve doing varied work.
- Hours that never became money. One organization found that AI delivered time savings which did not correlate to money saved; another weighed the cost against the manual work team members still had to do. An hour returned to a salaried person is a capacity gain rather than a cash release. State which of the two you claim.
Behind all three sits the same absence: no look-back. Book the look-back date in the document that asks for the money, and name who runs it. A case that schedules its own audit is harder to refuse and harder to abandon quietly.
Every Figure Carries Its Scope
A number without its scope is a rumor, and rumors are what finance teams are trained to find. Any figure entering the worksheet from outside your operation needs three labels: what it measured, whose data, over what period. Strip those away and a precise-looking percentage becomes unfalsifiable.
What Iternal’s own return tooling produces, stated plainly. Iternal builds a business-case deliverable written for the CFO: it models labor and opportunity cost, computes net present value, and projects investment and returns across a three-year horizon, breaking the ask into good, better and best. The model compares freely available published benchmark sources against the client’s own internal costs. Read the output for what it is, a projection built partly on published benchmarks — which is why the internal-cost side has to be yours.
Pin it down: questions for your evaluation
-
Which inputs in this return model came from our own operation, and which came from published benchmark sources?Whether the percentage describes your business or an industry average dressed as yours.
-
What usage rate per person does the model assume, and what does the return look like at half of it?The size of the single largest assumption, before your finance team finds it for you.
-
For each stated multiple or percentage: what was measured, on whose data, and over what period?The scope behind a headline figure, so a large number never enters your model without its context.
For more information on the tooling behind that deliverable, visit the blueprint builder page.
The Arithmetic, Run End to End on One Process
Take the process buyers name most often: staff re-keying data between two systems. Every letter below is a number you hold already or can count inside two weeks, and no price appears anywhere in the sum.
The subtraction, in full
- Line 1, baseline. Hours a week on the named process (H) times the loaded hourly rate (R) times 52. Add rework: at the one-in-five error rate described to us for manual processing, part of every week goes on correcting the week before. Baseline = H × R × 52 + rework
- Line 2, after. Count the same hours at the end of a bounded pilot, never before it. A worked example on our own record recovered 12 hours a week on one process — observed hours, not projected ones. After = (H − hours recovered) × R × 52
- Line 3, what it costs to get there. License plus machine plus data preparation, divided across the life of the asset. Three-year return horizons and four-year total-cost horizons both appear in these models; pick the one matching what is being depreciated, and say which you picked. Cost per year = one-time total ÷ horizon + anything recurring
- Line 4, the review that remains. Minutes a person spends checking each output, times volume, times the rate. On an AI-drafted response to a request for proposal, that check was described as a few minutes against the hours drafting used to take. Review = minutes × volume × R ÷ 60
Net annual return = Baseline − After − Cost per year − Review.
Convert the weekly hours into a dollar figure and stop there. The percentage is an output
of the subtraction, never an input to it.
Two amendments the arithmetic needs. Where the work created a sale that would not otherwise have happened, the baseline is zero and the whole figure is new revenue. And Iternal sorts return into three kinds: cash, the operational side of workforce and capability, and competitive advantage. Only cash belongs in the subtraction; put the other two in the narrative beside it, labeled as what they are. For more information on choosing which process to run this on first, visit the use-case selection page.
What a Proven Return Settles, and What It Leaves Open
A worksheet proves a return and buys nothing. It quotes no price, argues no license structure and names no source of funds, because each is a separate decision with its own signatory. Work the four lines first: a delta the business believes turns each question below into a negotiation.
Answered elsewhere
- What each product and buying route charges — see the cost page.
- Whether the entitlement recurs, and what one license covers — see the licensing page.
- Owning capacity on a device against renting it by the seat — see the own-or-rent comparison.
- What running the model consumes, and how to cap it — see the running-cost page.
- Which approved budget a purchase can ride when nothing was set aside — see the funding page.
- Buying rules, payment triggers and the clauses legal pushes back on — see the procurement page.
- Putting seats you already bought back to work — see the adoption page.
- Choosing and ranking the process to run this on first — see the use-case selection page.
- The software that produces the business-case deliverable — see the blueprint builder page.
FAQ
FAQ: Proving the Return Before You Ask
Prove it on one process rather than on the category. Baseline that process in hours and error rate, run a bounded pilot, measure it the same way afterwards, then subtract the one-time cost and the review that remains. Money gets wasted where nobody defined success before the purchase, so write the metric and the look-back date into the document that asks for the funding.
Replace the borrowed percentage with your own arithmetic. Skepticism lands on one of three things: a usage rate nobody has observed, a saving measured on a workload that looks nothing like yours, or hours that were never converted into cash. Show the return at half the assumed usage, match the benchmark to the shape of your work, and say whether you claim released cash or recovered capacity. A model that exposes its weakest input is easier to believe.
Count the same thing before and after, on one named process, with the definition agreed in advance. Weekly hours on the task converts most cleanly into a dollar figure, and error rate captures the rework. Set a baseline per process type rather than one average for the department: a contract review and a status report improve at completely different rates.
No, and treating them as identical is where business cases lose credibility. One organization found that AI delivered time savings which did not correlate to money saved. An hour returned to a salaried employee becomes cash only when somebody stops backfilling a role, retires a contract or redeploys the time onto revenue work. Label the claim as released cash or recovered capacity, and name who converts the second into the first.
Hand finance a subtraction rather than a story: the baseline cost of one process, the measured cost afterwards, the one-time investment spread over the life of the asset, and the review effort that stays. Price the status quo too, because running the current process another six months carries a cost somebody should name. For more information visit the cost page.
Organizations described taking roughly 12 to 15 months to reach tangible return the old way, and six to nine months where a top-tier expert team ran the work. Iternal states its packaged deployment compresses that to a 30-day return and insight period. Treat those as targets to test at your ninety-day look-back, never figures to bank.
Start With One Process
One process, four lines, your own numbers, and a look-back date somebody signs for. Every decision that follows — what it costs, how it is licensed, what it consumes, who pays and on what terms — has a page above.