Most AI business cases are written to win approval, not to survive it. They describe the value in confident terms, cost the model and the licences, and leave out the two things that decide whether the money comes back: what it costs to run the system every month, and how many people have to use it before it pays.
Our position is simple. An AI business case is not a forecast of benefits. It is a statement of what has to be true for the investment to return more than it costs, written so a finance team can check it later. If it cannot be checked, it is a pitch.
Why AI business cases fail
They rarely fail because the value was imaginary. They fail for more ordinary reasons:
- No baseline. "Faster service" is not a number. Without today's cost, cycle time or error rate, nobody can say later whether anything changed.
- The cost stops at the model. Data remediation, integration, workflow redesign, security review and change management are left out, or folded into "implementation".
- Run cost is a guess. Usage-based pricing, monitoring and human review are treated as small, then turn out to scale with every request.
- Adoption is assumed. The case quietly assumes everyone uses the system, every time. That assumption carries most of the value.
Each one is fixable before the funding meeting. None of them is fixable after it.
Start from the outcome and its baseline
"We need an AI chatbot" is not a use case. "Lower the cost of routine service contacts without losing first-contact resolution" is close to one. The difference is that the second names a business result, implies a baseline and tells you what must not get worse.
Before anyone models value, write down:
- the outcome you are changing, in business terms;
- its baseline today, measured, with the source named;
- the user and the process the system lands in;
- the measure that must not get worse, the guardrail;
- one named owner who will be held to the result.
If the baseline cannot be measured, that is the first finding. Fund the measurement, not the model.
What it is worth
Value comes in more than one form: revenue, margin, cycle time, capacity, customer outcomes, risk reduction or avoided cost. Pick the one or two the outcome actually moves, and model them in three scenarios: a low case, an expected case and a high case. The low case matters most. It is the one the CFO will believe.
Be honest about capacity. Saving ten minutes per task is only worth money if the time is redeployed, the work absorbed or the hiring plan changed. Otherwise it is a benefit to morale, which is real but belongs in a different column.
What it costs to build
The model is usually the smallest line. A credible build estimate covers the full set:
| Cost line | What it covers |
|---|---|
| Data remediation | Cleaning, joining and governing the data the use case depends on |
| Integration | Connecting to the systems where the work happens, with access controls |
| Workflow redesign | Changing the process so the output is acted on, not ignored |
| Evaluation | Test sets and acceptance criteria that say when it is good enough |
| Security and review | Threat assessment, data protection and the governance sign-off |
| Change management | Training, communication and the time people spend learning it |
If the readiness work is large, say so. A use case with the data ready and the workflow not ready is a different investment from one with both ready, and it should be priced as one.
What it costs to run
This is where AI business cases differ most from ordinary software cases. Many AI systems carry a cost that rises with use: model calls, retrieval, compute. On top of that sit costs that do not go away after launch:
- monitoring for quality, drift and misuse;
- human review of outputs where the stakes require it;
- re-evaluation each time a model, prompt or vendor changes;
- support, incident response and access reviews;
- licences and platform capacity.
Model run cost per unit of work, not per month. "Cost per resolved contact" or "cost per reviewed document" can be compared with the baseline directly. A monthly total cannot.
The number that decides
Put the pieces together and one number usually decides the case: the adoption level at which the investment breaks even. How many people have to use it, how often, for the value to cover the cost to build and the cost to run, within the payback period the business accepts?
If break-even needs near-universal use from day one, the case is fragile. If it pays back at a level of use you would be disappointed by, it is worth funding.
This number does three useful things. It moves the argument from "is AI valuable?" to a question operations leaders can answer. It shows how much the case depends on adoption and change management, which is usually more than the model. And it gives the owner a target to manage after launch.
When several use cases compete for the same budget, rank them on value, readiness, time to value, dependencies, cost and risk together. A high-value idea with weak data and no owner should not outrank a modest one that is ready now.
After the money is approved
Keep the business case and value realization separate. The business case says what the investment should return. Value realization checks whether it did, against the same baseline, with the same definitions. Mixing them lets a project redefine success after the fact.
Set a review date when the case is approved, not when the project ends. At that review, compare actual adoption with the break-even level, actual run cost with the model, and the outcome with its baseline. Then decide: scale, fix or stop. Stopping is a legitimate result. It is cheaper at month three than at month eighteen.
This is also where most pilots stall. We set out the wider pattern in why AI projects fail after the pilot.
A checklist for the funding meeting
| Question | What a fundable answer looks like |
|---|---|
| What outcome changes? | A business result, with a measured baseline and a source |
| What is it worth? | Low, expected and high scenarios; the low case defended |
| What does it cost to build? | All six cost lines, including readiness work |
| What does it cost to run? | A cost per unit of work, compared with the baseline |
| Where does it break even? | An adoption level and a payback period, stated plainly |
| Who owns it? | One named person, accountable for the result |
| When is it reviewed? | A date, and what would cause it to scale, change or stop |
This article describes how we approach AI business cases. It is not financial or investment advice for any specific decision.
