If You Can't Prove Your AI's ROI, You Don't Have ROI.
A forecast you can set your calendar by: at some future budget meeting, someone will ask what the company gets for its AI spend, and the quality of the answer will determine whether the budget survives. 'We think it saves time' does not survive. 'It saves €6,400 a month against €2,100 in costs, here's the one-pager' survives and compounds. Unmeasured AI is uniquely vulnerable because it's simultaneously expensive, fashionable to question, and — here's the tragedy — often genuinely valuable, just never proven to be. This page is the measurement framework: what to measure per use case, how to handle the missing-baseline problem honestly, and how to report it so the value becomes undeniable.
Why AI ROI stays invisible in most companies
Three causes, all structural rather than malicious. First, no baseline: nobody measured the old process — its time, its error rate, its cost — before the AI arrived, so there is precisely nothing to compare against, and improvement becomes unfalsifiable in both directions. Second, diffuse value: the benefit lands as twenty minutes saved here and forty there across a dozen people. Each fragment is below the threshold of anyone's attention; the aggregate might be a full-time salary's worth. Diffuse value is invisible by default and must be made visible by design. Third, absent ownership: IT deployed the system, the business uses it, finance questions it — and measuring it is, therefore, nobody's job. Until someone owns the number, there is no number. Notice that all three causes are fixable in an afternoon of decisions: pick the metric, assign the owner, take the baseline. The hard part was never the measurement; it's the prior decade of habit.
The baseline problem: measure first, or reconstruct honestly
The correct approach, for anything not yet deployed: before go-live, record how the current process actually performs — time per task measured with a stopwatch on real cases, not estimated; error rate from a sample; monthly cost in loaded hours. Two to four weeks of measurement buys you a permanently credible before-picture. If the AI is already live without a baseline — the common situation — reconstruct one honestly: time the manual process now on a sample of current work; mine your systems for before/after throughput (tickets closed, orders processed, quotes sent); interview the people who lived both eras. Then apply the credibility rule: round down, not up. A conservative number that survives skepticism is worth ten optimistic ones that collapse at the first probing question from a CFO. Your reconstruction method, documented on one page, is itself a defense — it shows the number came from a process, not from enthusiasm.
The right metric for each use case
Measure what the system was bought to change, in units that convert to money.
| Use case | Primary metrics | Conversion to money |
|---|---|---|
| Customer support AI | Cost per resolved conversation; resolution time; escalation rate | Deflected tickets × cost per manual ticket |
| Document processing | Minutes per document; error rate; throughput | Hours saved × loaded hourly cost |
| Sales & lead handling | Leads qualified; time-to-first-response; conversion rate | Margin on incremental converted leads |
| Internal copilots/assistants | Hours saved per person per week (diary sampling) | Pool of hours × loaded cost, discounted 50% for realism |
The honest cost side — where ROI claims usually cheat
An ROI figure is only as credible as its denominator, and AI costs are easy to understate. Count all of it: subscriptions and usage fees; the amortized build cost (a €24,000 build is €1,000/month over two years, and should appear as such); the internal hours spent managing, correcting, and feeding the system; and — the line item nobody includes — the cost of its residual errors, the human review time and the occasional mistake that gets through. Illustrative worked example: a document-processing system saves roughly 180 hours a month against €6,300 of loaded labor cost; its all-in costs are €1,800 monthly. Net €4,500 a month, honestly counted — payback inside five months. That figure survives a finance review precisely because the costs were overstated rather than hidden. The discipline generalizes: an ROI claim that includes its own costs is believed; one that omits them teaches your finance team to discount everything you say next.
The one-page monthly AI report
Everything above converges into a single artifact: one page, monthly, per system. Four numbers per system — baseline cost, current cost, hours saved, net euro value — plus a running ROI and one line of commentary. No dashboards requiring logins, no appendices; the format's austerity is the point, because the audience is a busy executive deciding whether this continues. This one page, maintained for six months, transforms AI from an article of faith into a managed investment with a track record. It changes budget conversations from 'should we keep paying?' to 'where's the next one of these?' — which is exactly the position you want to be arguing from. We build this reporting into every engagement as standard, because work we can't show the value of is work we don't expect to keep.
Related pages
- AI implementation ROI guide coming soon
- AI automation cost guide coming soon
- AI tool too expensive
- AI ROI calculator a guide for SME cf os
Frequently asked questions
What's a realistic ROI for SME AI projects?+
Our planning model for well-chosen automation assumes payback within 6–12 months and 200–400% annual ROI thereafter — an illustration to reason with, not a promise — but only measured projects ever find out which side of that they're on.
We deployed a year ago with no baseline. Is it too late?+
No. Reconstruct conservatively from system data and current manual timing samples. Imperfect, honest measurement beats perfect memories.
How do we count soft benefits like better decisions?+
Don't — keep the ROI math to hard time and cost, and mention soft benefits separately. One inflated line discredits ten solid ones.
Who should own AI ROI measurement?+
The business owner of the automated process, with finance reviewing the method — never the team that built the system grading its own work.