Build it around one process, a measured baseline and benefits that only arrive as the agent earns depth. Show the full cost, including the people who review its work. Name the owner. And include the conditions under which you would stop, because a case that cannot fail is not a case a board can test.
Why do most agent business cases mislead?
They assume the agent works at full authority from the first month. The benefit line shows every case handled end to end, while in reality the agent spends its first weeks reading, then drafting for a person to send, and only later acts on its own for a slice of the work. If the case does not model that sequence, the first quarterly review will look like failure when it is simply the plan working.
They also count the agent's output, not the work saved. A draft that a reviewer rewrites saves little. A business case that ignores reviewer time and rework will overstate the return by a wide margin.
What sections should the business case contain?
| Section | What it says | Who writes it |
|---|---|---|
| Problem and baseline | Volumes, handling time, error and rework costs today | Process owner and analyst |
| Scope | One process, the capabilities involved, what is out of scope | Process owner |
| Depth path | Expected depth at each stage, and the evidence needed to move | Process owner with risk |
| Benefits by stage | Savings at depth 1, 2 and 3, not just the end state | Analyst |
| Full costs | Platform, model, integration, reviewers, certification, fallback staffing | Analyst with finance |
| Risks and controls | What could go wrong, the bound that prevents it, the record it leaves | Risk |
| Ownership | The named owner and their authority to stop the agent | Sponsor |
| Decision points | When the case will be reviewed and what would stop it | Sponsor |
How do you build it, step by step?
- Sponsor: names the process and the owner. Evidence: a signed one-page scope. Our guide on who should own an AI agent in a business helps here.
- Analyst: measures the baseline for at least a month. Evidence: a baseline report.
- Process owner: sets a realistic depth path, assuming the agent starts at depth 1. Evidence: a staged plan.
- Analyst: models benefits at each stage, including a scenario where the agent stays at depth 2 indefinitely. Evidence: a staged benefit table.
- Finance: adds full costs, with reviewer time per item taken from a pilot or shadow run if one exists. Evidence: the cost model.
- Risk: maps risks to controls. Evidence: a risk-to-control table.
- Sponsor: writes stop criteria and the review dates. Evidence: signed decision points.
What does a sound case look like?
Say a credit union wants an agent to handle member requests to change direct debit dates and update contact details. The baseline shows around 3,000 such requests a month, at six minutes each.
The case models three stages. In months one and two, the agent shadows and saves nothing. In months three to five, it drafts changes at depth 2 and reviewers spend about two minutes each, so the saving is two-thirds of handling time. From month six, if certification supports it, it makes date changes within the member's limits at depth 3 and only a sample is reviewed. The case shows the depth 2 scenario alone still pays back within two years. Costs include a part-time reviewer role, certification every 90 days and a fallback plan for peak weeks. Stop criteria: if the edit rate stays high after four months, or any member is harmed by a wrong change, the board reviews whether to continue.
The board approves it partly because the downside is visible.
Which controls belong in the case itself?
- Benefits tied to evidence: no benefit from a depth until certification supports it.
- A no-widening scenario that still makes sense on its own.
- A named owner with authority to stop.
- Review dates with the metrics that will be reported.
- Links to the operating documents that will follow: the agent's brief, described in how to write a brief for an AI agent, and its runbooks for agent operations.
The sequence that the case funds is in our guide to deploying an AI agent.
Questions people ask
Should the case cover several processes at once?
Start with one. Multi-process cases blend strong and weak returns and make it hard to see which one is working. Add processes in later cases, using evidence from the first.
How do we put a value on fewer errors?
Use your own cost of errors today: refunds paid, complaints handled, regulatory remediation. If you do not track it, say so and leave it out of the headline rather than estimating.
Who should approve the business case?
The sponsor's normal investment committee, with risk sign-off on the controls section. The approval should cover the staged plan, not just the end state.