AI business cases fail in both directions. Some overstate the benefit and collapse at the first review; others omit the cost of the change and quietly cost more than they save. A case that survives is one that names its assumptions, prices the work honestly, and defines the measure before the spend.

This guide covers how to build that case: the costs to include, the benefits to quantify, how to structure the numbers, and how to present a result that can be checked. It is the companion to How to Adopt AI in a Small Business.

A business case with visible assumptions can be argued with and improved. One with a headline saving cannot.

Start with the baseline

A financial case rests on a baseline, and the baseline is usually the missing element. Capture four numbers before anything changes.

  • Hours per cycle for the task you intend to change, end to end, including review.
  • Loaded hourly cost of the people involved, as finance accounts for it.
  • Cycles per year — the cadence matters more than the single-cycle number.
  • Correction rate — how often the output is wrong, and at what stage it is caught.

Without these, any claimed benefit is an assertion. With them, the case is arithmetic, and the arithmetic can be verified after the fact. The baseline should be captured by the person who does the work, not estimated by the person who wants the project.

The four cost lines

Most business cases include one cost line. Honest cases include four.

Line What it covers Commonly omitted?
Tool cost Subscriptions, per-seat licenses, usage No
Setup cost Workflow design, prompt development, integration, policy and governance work Frequently
Training cost Design, delivery, practice time, refreshers Frequently
Run cost Verification time each cycle, maintenance, refreshers, incident handling Almost always

The fourth line is where marginal use cases die, and it is the line that makes a case credible. Verification time is real time, every cycle, and a use case whose savings are consumed by its checks is not a saving.

The benefit lines

Quantify four categories, and label anything you cannot quantify.

  • Time returned. Hours per cycle saved, times cycles per year, times loaded rate. Reduce this by an honest adoption factor — not everyone uses a workflow at full speed from day one.
  • Work avoided. Tasks that no longer need to happen, or that can be deferred rather than done.
  • Error and rework reduction. Fewer corrections, and fewer corrections caught late.
  • Capacity reallocated. Senior time moved from assembly to judgement or client work — valuable, but only if the time is genuinely redeployed rather than absorbed.

Anything that cannot be quantified — morale, retention, competitive position — should be listed as an unquantified benefit, not assigned an invented number. Doing the latter is the fastest way to lose an argument to a sceptical finance lead.

Structuring the case

A structure that survives review has four steps.

  • Cost per cycle today, from the baseline.
  • Cost per cycle after, including the run cost line.
  • The difference, annualized using the cycle count.
  • The payback period, against the setup and training investment.

Then state the assumptions in a short list: whose time, at what rate, at what adoption rate, and what has been excluded. Finally, state the measure that will determine whether the case was right — cycle time, correction rate, usage — so the review is evidence-based rather than retrospective.

A worked example

A firm spends fourteen hours per monthly client report, five of which are narrative drafting. Loaded cost is $65 an hour, and there are twelve reports a year.

  • Baseline cost: 14 hours × $65 × 12 = $10,920 a year per report type.
  • Task time saved: 4 hours per cycle, at an 80% adoption factor = 3.2 hours.
  • Verification added: 0.5 hours per cycle.
  • Net time saved: 2.7 hours per cycle × $65 × 12 = $2,106 a year.
  • Setup and training: 20 hours at $65 = $1,300, plus $600 in tool seats.
  • Payback: roughly 11 months.

The number is modest, and that is the honest output of a single report type. The case becomes compelling when the same workflow is applied across three or four recurring documents, because the setup is largely shared. Presenting the single-use-case number first is what makes the scaled number believable.

No figure in this example is a benchmark; it is arithmetic on stated inputs. Use the AI Adoption ROI Calculator to run your own.

When the case does not stack up

Sometimes it does not, and saying so is part of the discipline.

  • The task is too rare. The setup cost never repays.
  • Verification is expensive. The check consumes the saving; narrow the output instead.
  • The data is unreliable. The workflow will spend its savings on corrections.
  • The volume is too low. The same workflow applied to more cycles changes the answer.

In each case the response is to change the scope rather than the arithmetic. Adjusting an assumption to make a case pass is how programs acquire the reputation for overpromising.

How to present the case

The presentation matters as much as the arithmetic, because a sceptical reader looks for the weakest assumption first.

  • Lead with the baseline, not the saving. A measured starting point is the most persuasive line in the document.
  • Show the four cost lines in the same table, so nothing appears to have been hidden.
  • State the adoption factor explicitly, and use a conservative one.
  • Include the use case that did not stack up, if there is one. A case that reports only wins reads as advocacy.
  • Name the measure that will decide the question afterward.

A one-page case structured this way survives review because it invites scrutiny. A case built on a headline number invites a search for the assumption that produced it, and that search usually succeeds.

Common mistakes

  • No baseline. The benefit is asserted rather than measured.
  • Omitting the run cost. Verification time is the line that decides marginal use cases.
  • Assuming 100% adoption. Real adoption ramps, and the case should reflect it.
  • Assigning numbers to feelings. Unquantifiable benefits should be labeled, not priced.
  • Ignoring the shared setup. The second use case reuses most of the first one’s design.
  • No defined measure. The review becomes a debate about impressions.

Frequently asked questions

How do you calculate the ROI of AI?

Capture a baseline (hours per cycle, loaded rate, cycles per year), subtract the post-change cost including verification time, annualize the difference, and divide the setup investment by the annual benefit to get a payback period.

What costs are usually left out of an AI business case?

Verification time each cycle, the internal effort to design and document the workflow, training time, and ongoing maintenance. Those are the lines that decide whether a use case pays back.

How much can AI save a small business?

It depends entirely on the task and the verification cost, which is why the number should come from your own baseline. Where the burden is documented — 51% of the finance week on manual work (Intuit, May 2026, n=2,000) — the prize exists; the size of your share of it is a measurement, not a benchmark.

Should we include unquantifiable benefits?

List them, clearly labeled as unquantified. Inventing a number for morale or retention undermines the parts of the case that are real.

What payback period is reasonable?

Contained use cases with mostly shared setup should pay back within a year. Where the payback stretches beyond that for a single task, apply the same workflow to a second document type and re-run the numbers.

How do we prove the case afterwards?

Define the measure before the pilot — cycle time, correction rate, usage — and report it monthly against the baseline. A case that cannot be tested afterwards was not a case.

Should the case be built by the person who does the work?

Yes, with finance’s help on the loaded rate. The person doing the task knows the real cycle time and where corrections appear; a case built by the project’s sponsor tends to underestimate both.

What if two use cases share the same setup?

Allocate the shared setup to the first and record it as reusable. The second use case then shows a payback well under a year, which is the honest way to present a portfolio rather than inflating a single case.

Next step

Capture the baseline, price all four cost lines, and state the assumptions. See Measuring AI Adoption and Return for the measurement side, or run your own numbers with the AI Adoption ROI Calculator. To have the case built with your finance lead, book an AI adoption call.

Sources

  • Intuit Enterprise Suite, Future of Finance 2026 Report (survey of 2,000 CFOs, controllers and VPs of Finance at US businesses over $2.5M revenue, May 2026): 51% of the finance week spent on manual work such as reconciliation and report stitching.

Figures are cited from their sources and dated. The worked example above is illustrative arithmetic on stated inputs, not a benchmark.