AI implementation business case: what to measure before and after launch.

The decision in 60 seconds.

1. Define the outcome. Name one primary operating result and the exact work in scope.
2. Measure the baseline. Record where each number came from, who owns it, when it was observed, and how confident you are.
3. Model a range and the full cost. Keep released capacity, actual cash, service or revenue effects, and risk reduction separate.
4. Agree on the gate before launch. Set the evidence and thresholds that will lead to stop, continue, or expand.

Go directly to the worksheet, illustrative example, decision gate, or sources.

Why ROI alone is incomplete.

Capacity is not cash

Fewer hours of task work may create capacity for service, growth, resilience, or other priorities. It becomes a labor cash effect only when a paid outlay changes: overtime, contractor spend, a vacancy, or another cost finance can verify.

A forecast is not an actual

Before launch, adoption, task time, exception volume, and cost are assumptions. After launch, replace them with observed evidence and explain the variance. A forecast that is never reconciled is a sales artifact, not a decision record.

A model is not operating evidence

A calculator can make assumptions consistent and visible. It cannot establish demand, feasibility, quality, risk, adoption, or business value. Those require defined measures, accountable owners, source records, and an approved decision date.

The AI implementation business-case worksheet.

Complete these seven blocks with the operating owner, finance or evidence owner, and accountable sponsor. Keep the HTML open as the field guide; use the blank text version for a working session or printout.

Print / download blank worksheet

1. Decision and scope

2. Current baseline

Capture at least one volume measure, one time or cost measure, and one quality or service measure. Include seasonality, elapsed and touch time, rework or exceptions, service outcome, current cash costs, staff roles, review time, sample period, sample size, exclusions, and known gaps.

3. Benefit mechanics

4. Full investment and operating cost

List discovery, process mapping, data preparation, design, build or configuration, integration, testing, reviews, project management, migration, parallel operation, training, backfill, communications, temporary productivity loss, software, usage, infrastructure, monitoring, support, maintenance, vendor management, changes, human review, exception handling, quality assurance, audit, incident response, management oversight, exit, retention, and contingency where relevant.

Put per-task review in post-change task time. Use separate oversight only for work outside the task, so the same effort is not counted twice.

5. Forecast range and sensitivities

Use conservative, expected, and strong execution scenarios. These are scenarios, not confidence intervals. Give every changed input a documented basis rather than applying an arbitrary percentage.

Identify the two or three inputs that most affect the decision and the evidence that would reduce their uncertainty.

6. Adoption, controls, and measurement plan

Measure operating outcomes and controls alongside adoption. Define adoption as eligible work using the system divided by all eligible work, with exclusions recorded.

7. Decision record and actuals

Use the table in “After launch” to compare the primary outcome, adoption, post-change time, quality or control, recurring cost, net cash, and material non-financial effects. Select one decision and record the approver, date, reason, corrective actions, next review, and evidence location.

Print / download blank worksheet

A worked illustrative example.

This recurring-work example uses the same “Some cash realization” inputs and method as the interactive calculator.

Outputs after the ramp

The expected case only just recovers its modeled cash outlay inside the selected horizon. That is not a strong expansion case. It is a reason to verify the uncertain inputs, compare the non-AI alternative, and define a limited evidence-gathering scope.

Test the cash and capacity assumptions. Change adoption, post-change time, cash conversion, and rollout cost; any of them can reverse the conclusion.

Set the stop, continue, or expand gate before launch.

Real thresholds must be approved by the accountable leaders before a pilot. Record the threshold, evidence source, owner, correction period, decision date, and approver for each rule.

What the illustrative economics permit us to say

• Stop or reframe the cash-saving claim if no paid cost will change. The modeled result is capacity, not labor cash savings.
• Continue as a limited test while named owners validate 70% adoption, 6-minute post-change time, 10 monthly oversight hours, and 40% cash conversion by recorded dates.
• Do not call the forecast expansion-ready. Expected cumulative net cash is only $336 and payback occurs at the edge of the 24-month horizon.
• Consider expansion only after approved quality and control floors are met, operating evidence supports the reforecast, and finance verifies the cash mechanism and full costs.

Replace the forecast with observed actuals.

Update the record at month 1, the end of ramp, and the decision date. Explain each material variance and name the corrective action; do not overwrite the original forecast.

At the review, record exactly one decision: Stop, Continue within the current scope, or Expand. Add approver, date, reason, corrective actions, next review, and the retained evidence location.

How this worksheet was developed.

The worksheet combines operating-scope and measurement ideas from the NIST AI Risk Management Framework with documented cost-estimate and sensitivity practices from the U.S. GAO guide. Lasso adapted those principles into a decision record for an AI-enabled workflow, connected system, or managed operating process.

The illustrative calculator uses these rules:

• Task hours = volume × (minutes before − minutes after) ÷ 60 × adoption.
• Net capacity subtracts ongoing internal oversight.
• Labor cash effect = net capacity × loaded hourly cost × cash-conversion share.
• Implementation and rollout cash occur in month 0.
• Benefits ramp linearly; recurring software, support, and oversight start in full in month 1.
• Payback is reached only if cumulative cash recovers the upfront outlay and remains recovered through the modeled period.

The example excludes discounting, taxes, financing, and terminal value. If a buyer requires discounted cash flow, record the approved discount rate, timing convention, and source. The worksheet does not determine whether a project is feasible, safe, compliant, or valuable; accountable leaders must approve the evidence and thresholds for their context.