An ROI Framework for AI Automation That Survives the CFO

The short version

Credible AI automation ROI is calculated as annual realised benefit divided by total cost of ownership, where benefit is limited to money that actually changes a budget line and cost includes integration, evaluation, monitoring and change management — not just licences and tokens. Most business cases fail review because they count hours saved that were never removed from the payroll.

Why most AI business cases fail the CFO

Finance teams are not sceptical about AI; they are sceptical about unfalsifiable benefit claims. Three patterns get cases rejected:

• Phantom savings. "Saves 3 hours per person per week" with no headcount, overtime or contractor reduction attached.
• Cost tunnel vision. Licence and inference cost only; nothing for integration, prompt maintenance, evaluation or support.
• No counterfactual. The comparison is "do nothing," when the honest comparison is often "hire one person" or "fix the process first."

The framework

Step 1 — Define the unit of work

Pick something countable: a ticket, an invoice, a lead, a document, a report. Establish current volume per month and the fully loaded cost per unit today (labour minutes × loaded hourly rate, plus error and rework cost).

Step 2 — Classify the benefit

Assign every claimed benefit to one of four tiers, and discount accordingly:

1. Cash out (100% credit). Contractor spend removed, licence retired, vacancy not filled.
2. Capacity redeployed (50% credit). Freed hours reassigned to work that has its own measurable output.
3. Revenue influenced (30% credit). Faster response time, higher conversion — credit only what an experiment demonstrates.
4. Risk and quality (qualitative). Fewer errors, better auditability. Report it, do not monetise it unless you have loss data.

Step 3 — Build total cost of ownership

Include: build effort (engineering days × rate), integration and security review, model inference at forecast volume plus 40% headroom, evaluation and monitoring tooling, ongoing maintenance at roughly 20% of build per year, and change management (training, comms, hypercare).

Step 4 — Model three volume scenarios

Base, downside at 60% of forecast automation rate, upside at 120%. If the downside case is negative, the project needs a smaller scope, not a bigger deck.

Step 5 — Set the measurement contract up front

Agree with finance, before build, which system of record will report the metric, who reads it, and on what date the benefit is confirmed or the project is stopped.

A worked example

A support team handles 12,000 tickets a month. Average handling time is 9 minutes; loaded cost is £38 per hour, so £5.70 per ticket, £68,400 per month. An assistive AI deflects 18% of tickets fully and cuts handling time by 2 minutes on another 40%.

• Deflected: 2,160 tickets × £5.70 = £12,312/month
• Accelerated: 4,800 tickets × 2 min × £0.63/min = £6,048/month
• Gross benefit: £18,360/month → £220,320/year

Now apply the tiers. The team does not shrink; two open vacancies are cancelled (£90,000 cash out, 100% credit) and the rest is redeployed capacity (£130,320 at 50% = £65,160). Credited benefit: £155,160.

Costs: build £60,000, integration and security £15,000, inference £18,000/year, monitoring and evaluation £9,000/year, maintenance £12,000/year, change management £10,000. Year one TCO: £124,000.

Year-one ROI: (155,160 − 124,000) / 124,000 = 25%. Year two, with build cost gone, ROI exceeds 200%. That is a case a CFO can sign — and, crucially, it survives being wrong by 30%.

Payback, not just ROI

Boards care about payback period as much as ratio. Divide year-one TCO by monthly credited benefit: £124,000 / £12,930 ≈ 9.6 months. Anything under 12 months for an operational automation is generally fundable; beyond 18 months, expect scrutiny about whether the process should be redesigned instead.

Common adjustments

• Quality floor. Add a clause: benefit only counts while CSAT or error rate stays within an agreed band. This is what protects the case from a "we automated it badly" outcome.
• Ramp curve. Automation rate rarely starts at target. Model a three-to-six month ramp.
• Model price drift. Inference prices fall, but volume grows faster. Forecast on today's prices.

What to put on one page

Unit of work and volume; today's cost per unit; expected automation rate with ramp; credited benefit by tier; TCO by line; ROI and payback under three scenarios; the quality floor; the stop condition and review date. If your business case does not fit on one page, the scope is too large for a first project.