Prove AI ROI on one workflow before committing to a large project. Measure its baseline first: volume, handling effort, error rate and backlog. Run AI on that workflow with a named person approving the output, measure the same numbers again, and build the payback case from the difference. A small approval process can sign off on that evidence.
#Key takeaways
- Start with one workflow, not a programme: one measured result is stronger evidence than many promised ones.
- Measure the baseline before anything changes, or there is nothing to compare against.
- Use the same four measures before and after: volume, handling effort, error rate and backlog.
- Keep a person approving the output during the test, so quality is measured, not assumed.
- Let your finance team turn the measured difference into a payback case using its own cost figures.
#Why start with one workflow instead of a programme?
Because most AI projects that fail do so on value, not on the technology. Gartner expected 30% of generative AI projects to be abandoned after proof of concept, with unclear business value among the reasons (Gartner, july 2024). Coverage of the MIT NANDA report found 95% of organizations saw no business return from generative AI (virtualization review, august 2025).
One workflow with a clear baseline gives an approval committee a result it can check. Pick one with enough volume to measure and a process owner who wants it. We explain how in how to pick the first process to automate with AI.
#What baseline should you measure before using AI?
- volume: how many items the team handles in a fixed measurement period, such as supplier queries, invoices or claims.
- handling effort: how much staff time each item takes, from opening it to closing it, sampled across people and item types.
- error rate: the share of items that need rework, correction or an escalation after they were closed.
- backlog: how many items are waiting, and how old the oldest ones are.
Measure over a normal period, not a quiet or unusual one, and write down how you measured so the after-measurement uses the same method.
#What does a before-and-after measurement look like?
Illustrative example, the numbers are assumptions. A team handles 2,000 supplier queries in a measurement period, with an error rate of 6% and a backlog of 300 items. With AI drafting each reply and a person approving it, the next period of the same length shows 2,000 handled, an error rate of 3% and a backlog of 90 items.
In that example, handling effort is measured too: if the share of staff time spent on queries falls from 40% to 25%, that is the effort the team gets back. The figures are invented to show the method. Your own baseline decides whether the result is worth scaling.
#How do you turn the measurement into a payback case?
Give the measured differences to finance. It converts effort returned, errors avoided and backlog cleared into the cost lines it already tracks, then compares them with the cost of running and owning the workflow. Stennir supplies the measurements and the method; your finance team owns the conversion, so the case uses numbers it trusts.
#What should you not count as ROI?
Licence counts, prompts sent and survey answers about usage are activity, not value. The Work Trend Index found 75% of knowledge workers already use AI at work (Microsoft and LinkedIn, may 2024), yet that says nothing about whether any workflow got faster or more accurate. Count only what moved in the baseline.
We are
Stennir measures one workflow before and after, with a named person approving output, and hands you the numbers and the method behind them.
We aren't
Stennir does not sell a large programme on projected savings, or count usage statistics as a return.
#What should a finance or operations lead do next?
Pick the workflow, write down the four baseline measures and start counting before anything changes. AI consulting ranks candidate workflows with a payback case for each, and AI for finance shows common starting points. take the free readiness assessment or book a 30-minute call.
#Common questions
What do you need from us to measure ROI?
Access to the system where the work is tracked, a sample of real items, and the process owner's time to confirm how the work is done today. If volume or error data is not recorded yet, Stennir helps set up simple counting first, because without a baseline there is no return to show.
What if the workflow does not show a return?
Then the approval committee has learned it cheaply, on one workflow, with evidence. The measurements often show why: data was harder to reach, or the error rate was already low. That result is still useful, because it moves the effort to the next workflow on the ranked list.
Does the measurement work for Arabic workflows?
Yes. The same four measures apply, and the test set should include Arabic items in the proportion they really arrive, including Gulf dialect if customers write that way. Stennir works in Arabic and English, so quality is checked in both rather than only in English.
Who decides whether to scale up?
Your approval committee, using the measured result and the payback case finance built from it. Stennir recommends, but the decision and the numbers stay with you. If it goes ahead, the same measures continue in production, so the board can see whether the return holds.
