Part 3 — Do the Numbers

Four steps. Every AI ROI calculation, however dressed up, reduces to this: what does the process cost today, what will it cost with AI, what’s the benefit across three scenarios — never one — and what’s the return and payback.

Two rules matter more than the arithmetic. Be conservative on the AI scenario: if you think AI will handle 40% of the volume, model 30% and let reality beat the number, not the other way round. And never present a single benefit figure — three scenarios, conservative, expected, and optimistic, or the number won’t survive the first hard question it gets asked.

Step 1 — Baseline the current process

Establish what it actually costs today, before AI enters the picture.

• Unit of work — ticket, invoice, contract, call. Use whatever your process actually processes.
• Volume per month
• Average time per unit — the real time observed, not the time the procedure claims it should take
• Cost per hour, fully loaded — salary plus employer cost plus overhead
• Error or rework rate — share of cases needing correction or escalation
• Extra minutes per error, or a flat cost per error if you have one

From these: labor hours per month, labor cost per month, errors per month, rework cost per month. Sum them for the current cost per month and per year.

Step 2 — Define the “with AI” scenario

Model what changes, conservatively.

• Share of units handled by AI — be conservative here specifically; this is the number every optimistic pitch inflates
• AI time per unit, including any human check built into the process
• Human time per remaining unit — usually longer than the old average, since AI typically absorbs the simplest cases first and leaves the harder ones for people
• New error rate
• New monthly capacity with the same team — what the same headcount can now handle, which matters as much as the cost saved when a team is at capacity and facing growth

From these: hours saved per month, and the time saving per month and per year.

Step 3 — Benefits, in three scenarios

Never one number. Build conservative, expected, and optimistic cases side by side, driven by the same inputs at different assumption levels — the share of volume AI actually handles, the adoption rate among the team, the error rate achieved in practice versus in the pilot. The conservative case is the one that goes in the business case. The other two exist so you know the range, not so you quote the best of them.

Step 4 — ROI and payback

• Total implementation cost — direct and indirect, from the Canvas in Part 1. The line everyone forgets here is internal staff time; a project that “only” costs the vendor invoice has quietly ignored the hours your own people put in.
• Total annual benefit — from Step 3, conservative case
• Net annual benefit = benefit − ongoing annual cost
• ROI = net annual benefit ÷ implementation cost
• Payback period = implementation cost ÷ monthly net benefit

Worked example: AI-powered customer support

• 5,000 tickets/month, 25 minutes each, €35/hour fully loaded, 8% error rate at 20 extra minutes each
• Current cost: roughly €932,600/year
• With AI: 60% of tickets handled by AI at 2 minutes each, human time on the rest rises to 35 minutes, error rate drops to 3%
• Time saving: roughly €343,000/year
• Result: €300,000 net annual benefit, 214% ROI, payback in 5.6 months

Worked example: AP invoice matching

Continuing the back-office case from Part 2 — a 60-person distribution business, 3,000 supplier invoices a month.

• Current: 3,000 invoices/month, 12 minutes average handling time, €32/hour fully loaded, 16% needing a manual mismatch chase at 25 extra minutes each
• Current cost: roughly €19,200/month, €230,400/year
• With AI: 70% of invoices auto-matched at under a minute of review each, remaining 30% take slightly longer than today’s average since they’re the genuinely harder cases, mismatch rate on auto-matched invoices drops to 4%
• Conservative-case annual saving: in the region of €85,000–95,000/year, depending on how much of the freed clerk time converts to headcount avoidance versus simply less overtime at month-end
• The honest caveat: unlike the ticket example, the biggest lever here isn’t time saved per invoice — it’s whether freed capacity actually avoids a hire the business was about to make. Put that assumption in writing before you present the number, because it’s the one a CFO will test first.

Document your assumptions as you go

Every input above should have a one-line note on where it came from and who confirmed it — the current handling time from a timed sample, not a guess; the cost per hour from finance, not HR’s headline salary figure. A number without a source is the first thing a skeptical reader will pull on, and the fastest way to lose the room.

---