AI agent business cases: baselines, assumptions and staged investment decisions

A reviewable agent business case identifies the workflow to change, its current cost, the intended improvement and how the organization will decide whether to continue. Five sections keep evidence separate from assumptions and make the discussion more useful than a single efficiency percentage.

1. Specify the problem and outcome

Name the task, users, inputs, completion criteria and current constraints. For support order inquiries, the first scope might cover staff lookup, reply preparation and review rather than automatic handling of every support request.

Where quantities are unknown, identify the missing evidence and its owner. Keep forecast scenarios separate from measured baselines. Incomplete measurement can justify a bounded discovery phase rather than automatically disqualifying the opportunity.

2. Measure normal and exception work

Record task volume, handling time, tools, outsourcing, corrections and takeover. Compare the same period and task scope. First reply, issue resolution and valid completion are different endpoints; peak waiting time cannot be substituted for average waiting time.

Use this worksheet with values supplied by the responsible data owner:

Keep materially different task types visible rather than measuring only straightforward cases. Report unresolved work separately from established success or failure.

3. Explain benefit calculations and realization

Use share does not automatically equal saved-time share. Task mix, partial automation, review and rework affect the result. Record units, evidence, horizon and realization conditions for each benefit, then compare lower, central and higher scenarios.

Separate redeployable hours from actual spending reductions. Do not count the same time value as both capacity benefit and removed payroll. The UK Government Efficiency Framework, updated November 24, 2025, offers measurement guidance on non-cash efficiency, spending reductions and avoiding overlap.

Payback needs dated cash flows, rather than an annual-benefit figure alone. List initial payments, renewals, operations, review and training, alongside benefits that become cash and their timing. Keep recurring charges and potential update work in later years.

4. Assign detection, action and ownership

Define observable signals for insufficient data, unacceptable quality, access problems, integration failures and obstacles to use. Pair each risk with an action, owner and pause condition.

If use share triggers a pilot review, define the eligible user or task denominator and retain quality and safety conditions. Low use can reflect access, workflow, task fit or inadequate outcomes; investigate the cause. High use does not replace quality acceptance. The NIST AI RMF, released January 26, 2023, supports context-specific risk management, not budget approval.

5. Make each milestone a decision point

Confirm data, interfaces, permissions, baseline and acceptance criteria; run a controlled pilot; then decide whether to expand, adjust, hand over or stop. Specify deliverables, acceptance owner, spending limits and exit arrangements. Dates depend on scope and prerequisites.

The GAO Cost Estimating and Assessment Guide, published March 12, 2020, supports assumptions, alternatives, sensitivity analysis and actual-cost updates. Combine an agent investment worksheet with a workflow scope discussion to maintain a decision document. Clear presentation still requires investment review and evidence from actual acceptance.