An AI business case lands on your desk looking polished. Upward-pointing charts. Bold promises. Whether it's come from a vendor or your own IT team, the problem is usually the same: it sells the technology, not the business case for it.
Your job isn't to evaluate the technology. It's to stress-test the promise behind it.
The Soft Savings Fallacy
The most common trick is presenting time savings as if they're the same as cost savings. They're not. A typical example:
"AI tool saves 2 hours per employee weekly. 100 employees. £40/hour rate = £352,000 annual savings."
Unless someone is leaving or working fewer hours, those savings don't hit the P&L. They're capacity — and capacity is only worth something if you fill it with work that generates revenue.
Strategy: Separate efficiency claims into two categories:
Cashable Savings — Explicit spending elimination (cancelled recruitment, reduced vendor fees)
Capacity Creation — Recovered time requiring reinvestment
If the whole case rests on people using their freed-up time productively, you need a much higher bar of evidence.
The Pilot Trap
Pilots look good because the conditions are ideal — keen participants, clean data, plenty of support. Production is different. The users are less enthusiastic, the data is messier, and there's no one standing over it.
When you scale, the unusual scenarios start showing up. Performance tends to dip before it stabilises.
Strategy: Apply conservative adjustments to pilot metrics. If pilots showed 90% accuracy, model at 70%. Projects remaining viable at reduced performance levels are robust.
The Cost of Verification
Generative AI produces text that reads well but may not be accurate. Someone still has to check it. That doesn't eliminate the role — it changes what the role involves.
Example workflow:
- Manual drafting: 45 minutes
- AI generation: 1 minute
- Review/correction: 30 minutes
- Net savings: 14 minutes (not 44)
Verification requires concentration and catches subtle errors humans easily overlook.
Strategy: Demand workflow diagrams showing verification steps and associated human costs deducted from savings projections.
The Integration Black Hole
"Seamless integration" typically means API availability, not functional operational integration. Legacy systems (Sage 200, older Dynamics versions, custom databases) weren't designed for modern API calls.
Implementation budgets are frequently severely underestimated. Real integration requires:
- Mapping — Ensuring consistent definitions across systems
- Security — Preventing unauthorised data access
- Maintenance — Managing updates and compatibility
Strategy: Examine Software License-to-Professional Services ratios. Complex integrations should heavily weight services costs in year one. Disproportionate software costs signal underestimated complexity.
The Exit Strategy
Once AI is handling a significant chunk of the work, you're dependent on it. If you've also reduced headcount, a vendor price hike or a model failure leaves you with no fallback.
Strategy: All proposals require Business Continuity & Exit sections addressing:
- Manual fallback procedures if AI fails
- Data exportation capabilities upon vendor change
- Cash reserves for staff re-hiring if automation fails
The FD's Decision Matrix
Evaluate proposals through five commercial questions:
Revenue Logic — Are savings demonstrably bankable or theoretical?
Pilot Validity — Does data reflect realistic or idealised scenarios?
Net Time — Are verification costs deducted from efficiency claims?
Integration Reality — Is implementation budgeting realistic for legacy environments?
Reversibility — Can the organisation survive reverting this system?
If a proposal can answer all five honestly and still show a positive return, it's worth backing. If it can't, don't.
This isn't about blocking innovation. It's about making sure the money goes where it actually delivers.