An executive does not read a business case to learn. They read it to decide whether to bet a slice of budget and reputation on you being right. Everything in the document either reduces their uncertainty or wastes their attention. Most agent business cases fail because they spend pages on what the technology is and almost none on what it costs, what it risks, and whether it has already been proven on real work. This is the structure that does the opposite.

It builds on the stakeholder buy-in guide and uses the outputs of the ROI calculator and the TCO model. Those produce the numbers; this assembles them into a decision.

What the case has to do

The job of the document is to move one person from "maybe" to "yes" while leaving them able to defend the yes to their own boss. That reframes everything. You are not trying to be comprehensive; you are trying to be sufficient. You are not selling enthusiasm; you are removing reasons to say no. The most persuasive business case is often the one that volunteers the most about its own weaknesses, because that is what an experienced executive is scanning for anyway.

Keep it to one page or seven slides. The detail belongs in an appendix you carry into the room and open only if asked. Leading with the appendix is how good ideas get talked to death.

The seven-part structure

  1. The problem, in business terms. What is broken, who it hurts, and what it costs today in hours, errors, or delay. No technology yet.
  2. The proposed agent, as an outcome. What it will do, described the way a user would describe the result, not the architecture. "It triages every inbound ticket and routes it correctly within a minute," not "a multi-step reasoning pipeline."
  3. The numbers. Annual cost at projected volume, payback period, total cost of ownership. The gap between the cost of the problem and the cost of the agent is the headline.
  4. The risks, named first. Each one paired with how you mitigate it. Accuracy, security, vendor stability, adoption.
  5. The evidence. Results from a pilot on your own data. This is the slide that turns a proposal into a fact.
  6. The ask. Scoped and specific. Not "approve our AI strategy" but "fund this one agent for this one team for this period, with this success metric."
  7. The decision and timeline. What you need decided, by when, and what happens next if it is a yes.

The numbers that matter

Three numbers carry an agent business case, and a fourth quietly undermines it if you let it in. The three that matter: the annual cost of the agent at your real volume, the payback period, and the total cost of ownership including integration and maintenance, not just the subscription. The cost-versus-ROI framing matters more than the absolute figures, because a small, credible payback beats a large, fragile one every time.

The number to keep out is anything that does not translate to money or risk. "Processes 10,000 items a day" is a vanity metric unless you connect it to a cost saved or a revenue enabled. Executives discount optimistic projections by reflex, so bring a conservative estimate you can defend under questioning. A figure that survives the room is worth more than a bigger one that does not. For the underlying economics of why pay-per-use agents change this math, see the economics piece.

Name the risk first

The instinct is to hide the weaknesses and hope they do not come up. The opposite works better. When you list the risks before anyone asks, you do two things at once: you signal that you have thought the bet through, and you take away the executive's main job in the meeting, which is to find the hole you did not mention. A risk you raise is a sign of rigor. A risk they raise is a sign you were not ready.

Pair each risk with a concrete mitigation. Accuracy risk meets a human-in-the-loop gate on high-stakes actions. Security risk meets scoped access and audit trails. Vendor risk meets a pay-per-use model with no lock-in, so leaving is cheap. Adoption risk meets the pilot evidence showing one team already using it. Named and mitigated, each risk stops being a reason to wait.

Lead with pilot evidence

The strongest line in any agent business case is "we already ran this on one of our own teams, and here is what happened." A scoped proof of concept on your real data outweighs every external benchmark, because it answers the only question the executive truly has: will this work here, for us? Run the pilot before the business-case meeting, not after the approval. The order matters. Evidence-first cases get funded; projection-first cases get a follow-up meeting that never happens.

If you genuinely cannot pilot first, then the ask itself should be the pilot. "Fund a three-week experiment" is a business case an executive can approve in the room, and it sets up the real funding conversation with evidence in hand.

A one-page template

The whole case, compressed to what fits on a page:

If it does not fit on a page, you have not decided what matters yet. Before you walk in, sanity-check the platform side with the evaluation framework so the obvious "but which vendor?" question has an answer ready.

FAQ

What goes into an AI agent business case?
Seven parts: the problem in business terms, the proposed agent as an outcome, the numbers, the risks named first, the pilot evidence, a scoped ask, and a decision with a timeline. Lead with the problem and the numbers.
What numbers should it include?
Annual cost at projected volume, payback period, and total cost of ownership including integration and maintenance. Quantify the current cost of the problem so the agent has something to be measured against.
How do I present ROI to executives?
Show the cost of the problem today, the cost of the agent, and the gap as a payback period. Use a conservative figure you can defend; executives discount optimistic numbers automatically.
Should the case mention risks?
Yes, first. Listing risks before anyone asks signals rigor and removes the executive's job of finding the hole. Pair each risk with its mitigation.
How long should it be?
One page or seven slides. Executives decide on a small amount of well-chosen information. The full detail belongs in an appendix you bring but do not present unless asked.
What is the strongest evidence?
A completed pilot on your own data. Industry benchmarks frame the opportunity, but a result from one of your own teams persuades more than any external statistic.

Sources