How to measure AI ROI without guessing
Most discussions about AI stop at technology. The important discussion is economics. If you can't show what the initiative costs, what it saves and what it earns, you're left with opinion. AI then becomes a project that survives on enthusiasm, not results.
A good ROI assessment starts from a simple principle: you measure against today's operations, not against an ideal state. Many overstate the gain because they compare against a process they should have improved anyway. Others understate the gain because they only look at license cost and ignore time, errors and lost revenue.
The standard formula
Net gain is the sum of documented gains minus running costs. Total investment is the upfront costs, including implementation, integration, training and internal time. The formula is familiar. The challenge is what you put into it.
The cost base in four categories
Start with the cost base. Split it into four categories so nothing falls between the chairs.
Development, setup, data cleanup and testing. Everything that has to happen once before the initiative can run in production.
Models, API usage, monitoring and support. Cost that scales with volume month over month.
The team that owns the process, improves rules and handles exceptions. Often undervalued.
Training, documentation and temporary slowdown during rollout.
Three gain tracks
Gain gets measured in three parallel tracks. One initiative can show up in all three, or only in one. Know which before you start.
How many hours per week are freed from repetitive tasks? Multiply by realistic hourly cost: salary plus social costs and overhead.
Fewer errors, fewer credits, fewer complaints and less rework. Calculate in concrete kroner, not in feeling.
Faster response in sales, higher conversion, or more capacity without hiring.
Example: a Norwegian service firm
A simplified example from a service firm with 45 employees. Three customer coordinators handle manual document processing and email routing on 90 hours a week. Internal hourly cost is 720 kroner. Error cost from misregistration is estimated at 55,000 kroner per month.
The upfront investment is 620,000 kroner. Running cost is 78,000 kroner per month. That's what we measure the gain against.
The calculation
The calculation isn't perfect, but it's better than gut feel. You see what drives the result and how sensitive the model is to changes in volume, hourly cost and error rate.
KPIs per initiative
A common mistake in Norwegian boardrooms is to use one combined ROI number for the whole AI portfolio. That hides what's working. Measure per process or use case. One initiative can show negative return for six months before it turns, while another pays back from quarter one. With separate tracking, you make better priority calls.
For 2026, follow at minimum these KPIs per initiative:
- Cycle time per case from start to finish
- Share of cases resolved without manual intervention
- Error rate and rework rate
- Volume per week and seasonal variation
- Cost per processed unit
- Effect on customer experience, for example response time or NPS delta
Set fixed measurement points. A good rhythm is baseline for 4 weeks, then measurement at 30, 90 and 180 days. Without baseline, every improvement becomes an argument, but not evidence.
Risk, compliance and strategic value
Not every gain is equally certain. Use a probability weight on each gain item. If the expected conversion lift is uncertain, multiply the value by the probability. That gives a more conservative and steerable picture when you prioritise across projects.
In Norway, regulatory factors come in. For initiatives processing personal data, privacy requirements, data storage and internal control have to be in the cost picture. If compliance work comes late, ROI becomes artificially high in the planning phase and disappoints in operations. Bake it in from the start.
Initiatives with strategic value but low direct savings (shorter time to market, better decision quality, less vulnerability on turnover) shouldn't be discarded. But they should have their own criteria. Don't squeeze every project into the same spreadsheet logic.
Practical working method
Require a one-page investment template before every AI initiative. It should contain problem, baseline, expected gain, cost profile, risk factors, owner and evaluation timepoint. If an initiative can't be described concretely on one page, it's usually not clear enough to put into production.
The goal of ROI isn't to prove that AI always pays. The goal is to find where it pays, where it doesn't, and what needs adjustment. Companies that work this way cut faster on initiatives that aren't working and scale what actually delivers. That's where the competitive edge sits.
- 2026-04-29v6rewritten in English; voice + tokens preserved
- 2026-04-20v5restructured with sections and visualisations: comparison bars, cost categories, gain tracks and animated calculations
- 2026-04-19v4added Norwegian service-firm example and updated error cost
- 2026-04-12v3clarified that internal hourly cost includes social costs and overhead
- 2026-04-05v2split the cost base into four categories, added growth-gain track
- 2026-03-29v1first edition
These notes are updated weekly.
Get the next edition in your inbox — short, concrete, no noise.
Transaction monitoring: rules engine, agent, and audit trail
Rules alone produce too much noise. Pure ML alone gives no explanation. How to combine them with an agent that keeps the auditor's shoulders relaxed.
CONTINUE