Tool-usage counts alone do not establish business return.
“80% of employees tried ChatGPT.” Interesting, but not ROI.
“We ran three AI workshops.” Useful, but not business impact.
“People say they save time.” A signal, but not enough to guide investment.
A defensible AI ROI analysis should identify the affected workflow or other agreed unit. Which task changed? How often does it happen? How much time changed? Did quality improve or decline? What risk was introduced? Is the new behavior sustained after the initial rollout period?
This article gives a practical measurement model for SMEs and teams.
Measure changed workflows with a baseline, quality and risk controls, sustained use, and the full cost to serve. Do not invent time savings or adoption periods.
Start with the unit of value
The unit is not “AI use.” The unit is a workflow:
- Draft customer proposal.
- Triage support ticket.
- Summarize meeting and assign actions.
- Extract invoice fields.
- Prepare sales research.
- Review contract clauses.
- Generate product description.
- Answer internal policy question.
For each workflow, measure before and after.
Define net benefit and ROI separately
A common financial structure is:
Net benefit = quantified benefits - total relevant costs
ROI (%) = net benefit / total relevant costs × 100
The organization’s finance reviewer must define the period, cash versus non-cash benefits, labor valuation, attribution, tax/accounting treatment, discounting, and which implementation or shared costs belong in the denominator. If benefits cannot be monetized credibly, report them as operational metrics rather than forcing them into ROI.
Potential value measures include:
- Time saved.
- Higher throughput.
- Faster response time.
- Better quality.
- Fewer errors.
- More complete records.
- Higher conversion.
- Lower support load.
Use one non-overlapping cost ledger: implementation/training, tool/inference, infrastructure, review/rework, and ongoing operations/control work. Allocate shared security, legal, monitoring and incident costs once. If review is already included in measured handling time, do not subtract it again as a separate labour charge.
Distinguish capacity released from money saved. Fewer staff minutes can create useful capacity without reducing payroll. Revenue change also needs an attribution design; a simultaneous pricing or sales change can explain it.
The baseline
Before changing the workflow, capture a real baseline. The entries below are examples of fields, not reported results:
| Metric | Example |
|---|---|
| Volume | 120 support tickets/week |
| Current time | 6 minutes per ticket triage |
| Current quality | 8% misrouted |
| Current delay | Median first routing in 2 hours |
| Current cost | Staff time and tools |
| Current risk | Sensitive customer data, escalation errors |
Then run the AI workflow on a comparable pilot and record differences in task mix, staffing and other changes that could explain the result.
Without a comparable baseline, label the observed figures as pilot performance rather than an attributable improvement.
Measure quality, not just speed
AI can make bad work faster. Measure quality in parallel:
| Workflow | Quality metric |
|---|---|
| Support triage | Correct category, correct priority, correct escalation |
| Meeting summaries | Action item accuracy, owner/date correctness |
| Sales research | Source quality, relevance, no unsupported claims |
| Contract review | Correct clause identification, missed-risk rate |
| Invoice extraction | Field accuracy, exception rate |
| Knowledge RAG | Citation correctness, refusal correctness |
For customer-facing work, add trust metrics: complaint rate, correction rate, opt-out rate, human escalation satisfaction.
Measure adoption beyond usage counts
Usage is not enough. Track:
- Repeat usage after a defined evaluation window; four weeks is one example, not a universal adoption threshold.
- Workflow completion rate.
- Manual override rate.
- User edits after AI output.
- Rework caused by AI output.
- Cases where users avoid the workflow.
- Reasons for avoidance.
If use appears only during observation or a mandated trial, do not infer sustained voluntary adoption. Investigate policy, task frequency, measurement effects, and reasons for use or avoidance.
Maturity levels
The six levels below are an AI Expert OÜ house rubric, not an externally validated maturity standard. Adapt it and document why each level matters:
| Level | State | Evidence |
|---|---|---|
| 0 | No managed AI | Ad hoc personal tool use |
| 1 | Individual productivity | People use approved tools for drafts and analysis |
| 2 | Repeatable workflows | Named workflows with owners, prompts, and checks |
| 3 | Governed automation | Logs, evals, review gates, fallback, data rules |
| 4 | Integrated systems | AI connected to systems of record with monitoring |
| 5 | Optimized portfolio | ROI, risk, cost, and quality managed across workflows |
The goal is not to maximize the level. Advance only where evidence shows that added integration and governance are justified.
A completed synthetic comparison
These invented figures demonstrate the accounting, not an observed return. Assume 1,000 comparable tasks per month and a finance-approved labour value of €30/hour.
| Component | Baseline | Candidate |
|---|---|---|
| Preparation | 5 minutes/task | 2 minutes/task |
| Review | 1 minute/task | 2 minutes/task |
| Rework | 1 minute/task | 1 minute/task |
| Total labour | 7,000 minutes | 5,000 minutes |
The candidate releases 2,000 minutes, or 33⅓ hours, valued at €1,000. Review and rework are already included. Suppose the month also carries €300 in incremental tools, €200 in operations and €100 allocated implementation cost: total incremental costs are €600. Under this declared incremental-value convention, net benefit is €400 and ROI is 400/600×100≈66.7%.
That is a capacity-valued return, not automatically €400 of cash saved. If the released time cannot be productively used, the assumed €1,000 benefit needs revision. If quality or a mandatory safety slice worsens, positive arithmetic does not authorise scaling.
A dashboard can show both the financial assumption and the operational observation: “2,000 fewer minutes at this task mix; quality unchanged within the tested uncertainty; capacity valued at €30/hour.” Only claim the quality clause when an appropriate comparison actually supports it. Compare a holdout or phased/matched rollout where feasible, and record confounders before attributing the difference to AI.
Leading and lagging indicators
Candidate leading indicators:
- Number of workflows with owners.
- Number of workflows with baseline metrics.
- Percentage with data rules.
- Percentage with fallback paths.
- Eval pass rate.
- Human review queue volume.
Candidate lagging indicators:
- Hours saved.
- Cost reduced.
- Revenue influenced.
- Error rate changed.
- Cycle time changed.
- Customer satisfaction changed.
- Incident count.
Classify indicators according to the causal model and measurement window. Leading indicators may show whether enabling controls and behaviors exist; lagging indicators may show later operational or financial outcomes. Neither proves causality without an attribution design.
A staged measurement plan
Stage 1: Baseline.
- Pick a manageable set of candidate workflows.
- Capture volume, time, quality, and risk.
- Choose the pilots using value, feasibility, and risk evidence.
Stage 2: Pilot.
- Run AI-assisted workflow with human review.
- Measure time, quality, override rate, and user feedback.
- Stop or revise weak pilots.
Stage 3: Scale decision.
- Compare baseline vs pilot.
- Decide: scale, keep small, revise, or cancel.
- Add governance controls for scaled workflows.
Do not call a pilot successful because people liked it. Call it successful when the workflow metrics justify continuing.
Record the decision and its limits
Use the ROI and maturity scorecard to preserve baseline, task mix, total handling time, quality slices and the finance-approved valuation. The FinOps Foundation’s unit-economics guidance connects technology cost to an organisational value unit; it does not supply your labour value or prove attribution.
Decide whether to scale, keep small, revise or stop. A higher house-rubric maturity level is not a return, and more integrations are not a goal by themselves.



