Technical Lead & BPMN Educator·8 min read
ROI is a difference, so you need a before
Return on investment is, by definition, a comparison: the value created minus what it cost, relative to what it cost. For AI, "value created" is the difference between how the process performed before and after. If you never measured the before, you cannot honestly measure the after.
That is why measurement starts long before go-live. Capture the baseline - cost per run, cycle time, error rate, volume - and freeze it. This is the single most skipped, and most important, step in the whole exercise.
A simple ROI formula
Annual value = (baseline cost per run − new cost per run) × annual volume
+ value of time saved, quality gained, capacity freed
ROI = (annual value − annual AI cost) ÷ annual AI cost
Worked example
An invoice-coding step costs €6.00 per invoice today and runs 40,000 times a year. After adding AI-assisted coding with a human check, it costs €2.40 per invoice. Annual value = (6.00 − 2.40) × 40,000 = €144,000. If the AI (usage + review + maintenance) costs €36,000 a year, ROI = (144,000 − 36,000) ÷ 36,000 = 300%. The number is only credible because the €6.00 baseline was measured, not assumed.
What to measure
| Metric | How to capture it | Why it matters |
|---|---|---|
| Cost per run | Fully-loaded cost to execute the process once (labour + tooling). | The headline lever most AI initiatives move. |
| Cycle time | Elapsed time from trigger to completion. | Speed often matters as much as cost - and is easy to measure. |
| Error / rework rate | Share of runs that need correction or rework. | AI can cut errors - or add new ones. Watch this both ways. |
| Volume handled | Runs completed per period without adding headcount. | Captures capacity gains that a cost-per-run figure hides. |
"Watch the trap of agent metrics. A dashboard showing 98% task-success and low latency feels like progress, but it is not ROI. If cost per run and cycle time have not moved against the baseline, there is no business value yet - however good the agent looks."
Outcome metrics vs agent diagnostics
Keep two layers separate. Outcome metrics - cost, cycle time, quality, volume - are the ROI, and they only mean something against the frozen baseline. Agent metrics - success rate, token cost, latency - explain the outcome but are not the outcome. Report the outcome to the business; use the agent metrics to diagnose and improve.
This is the final stage of a proper implementation method - capture, design, deploy, then measure against the baseline. See how to implement AI in business for the full sequence. Keeping the baseline numbers on the process model itself - which AI-native tools such as Crismo support - makes the before/after comparison straightforward instead of a spreadsheet archaeology project.
Related guides
Frequently asked questions
How do you calculate the ROI of AI?▼
ROI = (annual value created - annual cost of the AI) / annual cost of the AI. Value created is the difference between your process metrics before and after: (baseline cost per run - new cost per run) x annual volume, plus any gains from speed, quality, or freed capacity. The critical input is the baseline you captured before you started.
Why do you need a baseline to measure AI ROI?▼
Without a before-picture you have nothing to compare the after-picture to, so any ROI number is a guess. Capture cost per run, cycle time, error rate, and volume before you deploy, and freeze them. Every claim of impact is then a measured difference, not an assertion.
What is the difference between outcome metrics and agent metrics?▼
Outcome metrics are the business result: cost, cycle time, quality, volume. Agent metrics - token usage, task success rate, latency - explain how the AI behaved. Outcome metrics are the ROI; agent metrics are diagnostics that tell you why the outcome happened. Do not confuse a good agent score for business value.
How long before AI ROI shows up?▼
For high-volume, rule-clear tasks, weeks. For judgement-heavy or change-sensitive processes, a quarter or more, because adoption and refinement take time. Measure against the frozen baseline at set checkpoints rather than waiting for one big reveal.
What costs go into an AI ROI calculation?▼
Model/API usage, the build and integration effort, ongoing maintenance and monitoring, human review time for AI-assisted steps, and change-management. Leaving out review and maintenance is the most common way AI ROI gets overstated.