Skip to content
be-human.ai
← Back to the journal

Business workflows

AI Business Case Template: Gains, Costs, Risks and Approvals

A one-page AI business case template for a first company AI layer: expected gains, budget categories, risks and the approval points a person must keep.

Be Human typographic cover with the title: AI Business Case Template: Gains, Costs, Risks and Approvals
Conceptual illustration created with AI · Be Human

An AI business case template is a one-page document that states the decision you are asking for, the single role an agent will play, the gains you expect and how you will verify them, the budget in clear categories, the risks with named owners, and the points where a person must approve before the system acts. For a first company AI layer, the page works best as a decision tool rather than a forecast. It should let an operations leader answer yes, no, or yes with conditions in one meeting.

Key takeaways

  • Start with one role and one measurable outcome, not with a platform purchase.
  • Show budget in categories such as setup, integration, usage, review time and maintenance, so each line can be challenged on its own.
  • Write expected gains as ranges tied to a baseline you measured yourself, and count the time people spend reviewing the agent's output.
  • Pair every risk with one control and one owner, and name the approval points before the agent takes any action.
  • Set a review date and the evidence you will bring to it, so the case is tested against results instead of kept as a promise.

What belongs on one page?

A useful case has nine fields, and each one should fit in a few lines. Keep the whole document on one page because a longer document tends to be read as a plan to negotiate rather than a decision to make. Teams preparing for more capable AI systems often want to discuss the long term first, but the case should be decided on the work in front of you today.

  • The decision requested: approve a pilot role, approve a budget envelope, or approve a review period only.
  • The role: one job, the people who depend on its output, and the systems it reads from.
  • The baseline: how the work is done now, with the time, volume and correction patterns you can observe.
  • Expected gains: stated as ranges, each tied to a specific baseline figure.
  • Budget categories: setup, integration, usage, monitoring, human review time and an incident reserve.
  • Risks: each with an owner and a mitigation.
  • Approval points: the moments where a named person decides before the agent acts.
  • Success criterion: one observable measure and the threshold that counts as success.
  • Review date: when the case will be checked, and who brings the evidence.

If a field does not fit, that usually means the scope is too wide. Narrow the role until each field can be filled with a plain sentence.

How do you estimate gains without inflating them?

Start from the work, not from the technology. Choose one task that repeats often enough to measure, and record who performs it, how long it takes from start to finish, and how often it returns for correction. A gain is the difference between that baseline and the same work with the agent handling the routine part, minus the time people spend checking what the agent produced. Many inflated cases fail at this subtraction because review time was never counted.

Write each gain as a range with the assumption stated beside it. For example, a case might say that routine intake moves from a person reading every message to a person reviewing a proposed category, and that the reviewer's time per item is the figure to watch. Keep hours removed separate from hours redeployed. Redeployed hours are real, but they are not a cash saving unless someone changes what those hours produce. The approach described in measure AI workflow ROI without inflated claims is a good companion for this section.

Unit economics is a useful lens here. The FinOps Foundation frames unit economics as connecting spend to business output, which translates well to a single question: what does one completed task cost, including the human minutes it still needs? If that number is not visible in the case, the gain cannot be checked later.

Which budget categories should the case show?

Show categories rather than one total wrapped around guesses. Categories let finance and operations challenge each line separately, and they show where cost will rise if volume grows. A first AI layer usually needs these lines:

  • Design and setup: mapping the process, defining the role, and writing the approval rules.
  • Integration: connecting the systems the agent reads from and writes to, with permission scopes limited to the role.
  • Usage: model and tool consumption, measured per completed task so growth is visible.
  • Evaluation and monitoring: recurring checks that outputs still meet the success criterion.
  • Human review time: the hours people spend approving, correcting and handling exceptions.
  • Maintenance: updates when a connected service changes its interface or an internal tool is replaced.
  • Incident reserve: time and budget set aside for investigating and fixing something that goes wrong.

Human review time is the line most often left out, and it is the one that decides whether the gain is real. State the volume assumption beside the usage line, and say what changes if the volume is higher than planned. A case that shows its assumptions is easier to approve and easier to correct later.

What risks belong in the case?

Name risks in the terms your operation uses rather than as general fears about AI. The NIST AI Risk Management Framework organizes this work around governing, mapping, measuring and managing risk, and those four verbs make a practical checklist for a one-page case. For a first agent, the risks that usually matter most are these:

  • Wrong action: the agent sends, changes or books something it should only have proposed.
  • Exposure: client or employee information reaches a tool or context where it does not belong.
  • Injected instructions: text inside an email, document or web page tries to steer the agent away from its role.
  • Silent drift: output quality declines after a model, prompt or source changes, without an obvious error.
  • Single-person dependency: only one person understands how the agent is configured.
  • Vendor change: an integration breaks when a connected service alters its interface.

For each risk, write one control and one owner. The OWASP guidance on prompt injection recommends treating content the agent reads as untrusted input and limiting what the agent can do with it, so a control might be that the agent may summarize an inbound message but cannot forward or change records based on instructions inside it. For data exposure, the controls belong in the permission design. The checklist for isolating client data in AI agents gives a practical starting list.

Where should a person approve before the agent acts?

The approval section turns the case from a promise into a control design. For each action the agent could take, decide whether it runs alone, runs after a named person approves, or stays a proposal. Use this sequence as a decision guide:

  • Read and summarize: the agent may act alone, and every summary links to its source.
  • Tag or classify internal records: the agent may act alone within a fixed list of categories, with a sample reviewed on a set schedule.
  • Draft a message or document: the agent drafts, and a named person sends or publishes it.
  • Change a customer, financial or legal record: the agent proposes the change, and a named approver accepts or rejects it.
  • Spend money, grant access or issue a refund: a person approves each instance, and the limit is written into the case.
  • Anything outside the defined role: the agent stops and places the request in an exception queue with a reason.

The routing rules themselves can be documented separately, and the guide on routing each request to the right approver shows how to keep that logic readable.

Illustrative scenario: a small services firm with one shared inbox wants an agent to categorize incoming requests and draft first replies. The case states the baseline as the current time from arrival to a useful first reply, the gain as a shorter wait for routine requests, and the approval point as a person sending every reply during the first review period. The risk section names the inbox as untrusted input, and the success criterion is a sampled check of routing accuracy reviewed by the operations lead each week.

How should the case be reviewed after go-live?

A business case is a hypothesis until it is checked against operating records. Set the review date when you approve the case, and name the person who presents the evidence. That evidence should include the baseline comparison, the exceptions logged, permission changes, incidents and the result against the success criterion. Decide in advance what happens if the criterion is missed: narrow the role, return the agent to observation only, or stop. Anthropic's guidance on evaluations for AI agents describes building repeatable tasks to measure behavior over time, which is more useful for a review than a single impression of how the system performed.

Be Human handles this by treating the business case as part of the operating record. In a first engagement, the role, memory and tools, approval boundaries and observable success criterion are defined before anything runs. In managed operations, the team monitors the agent, runs recurring evaluations, reviews permissions, responds to incidents and proposes controlled changes, then brings the evidence back to the business owner at the agreed review point, so the case is revisited rather than filed.

If you are not yet sure which of your tasks is the right first one to hand to an agent, the free Be Human AI diagnostic can help you find it.

Frequently asked questions

How long should an AI business case be? One page is enough for most first decisions, and that limit keeps the reader focused on the choice in front of them. If the team needs more detail, put it in an appendix that the approver can request, while the summary page stays short enough to read in one sitting.

Should the case include a single return figure? Include ranges for expected gains and the assumptions behind them, but avoid relying on one total return figure, because it tends to hide the assumptions that matter. Show what you measured, what you assumed and what would change the answer. The approver can then judge the case on its evidence rather than on its headline.

Who should own the business case? A named business owner should own it, because that person makes the decision and is accountable for the result. Operations, finance and the people who run the systems should write their own sections, but one person signs the page and holds the review date.

Does the case need to change when the AI tools change? The role, the approval points and the success criterion should survive a tool change, because they describe the work and the controls rather than the product. The budget lines and usage assumptions will usually need updating, and the review should check whether a new tool changes the risk picture, especially around data access and permissions.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →