Decoding the AI Business Case
How to Read Between the Lines of an AI Vendor Proposal
Here is the scene: An AI business case lands in your inbox. It is glossy. The charts point upward. The promises are bold. Your operations will be revolutionised.
Whether the proposal comes from an external software vendor or your own IT Director, it shares the same blind spot.
It is written to sell a technology. You need to buy a business outcome.
As Finance Director, your role is not to assess the neural architecture of the model. Your role is to assess the validity of the promise. You are the stress-test.
Here is how to deconstruct an AI proposal and find the reality buried beneath the hype.
The Soft Savings Fallacy
The most common lie in an AI proposal is the Efficiency Calculation. It usually looks like this:
This AI tool saves 2 hours per employee, per week.
We have 100 employees.
Average hourly rate is £40.
Total Savings: £352,000 per year.
This number is fiction.
Unless you plan to reduce headcount or reduce working hours, you are not really saving £352,000. You are simply creating capacity. Capacity is only valuable if it is filled with profit-generating work.
If a credit controller saves two hours a week, do they use that time to chase harder debt (cashable benefit)? Or do they simply have a more relaxed Tuesday afternoon (soft benefit)?
The Red Pen Rule: Circle every line item marked efficiency savings. Demand that the proposer recategorise them into two columns:
Cashable Savings: Money we will explicitly not spend (e.g., cancelling a recruitment requisition, reducing agency spend, reducing software licences).
Capacity Creation: Time saved that must be reinvested.
If the proposal relies entirely on Capacity Creation, it is not a cost-saving project. It is a productivity bet. The burden of proof is higher: ask them exactly what that new capacity will produce in revenue or profit terms.
The Pilot Trap
Proposals often use data from a successful pilot to justify a full rollout.
"We tested this with the sales team for a month, and conversions went up 20%."
Be extremely sceptical of pilot data.
Pilots are run under ideal conditions. They typically involve your most enthusiastic staff, your cleanest data, and significant management attention.
Production is The Wild. It involves your most cynical staff, your messiest data, and limited hand-holding.
When AI moves from pilot to production, performance usually drops before it recovers. This is often due to Edge Cases.
In a pilot, you might test the AI on 50 standard customer contracts. It works perfectly. In production, it encounters a contract from 2017 scanned upside down with coffee stains. The AI fails. The process halts. A human has to intervene.
The Red Pen Rule: Apply a reality discount to pilot metrics. If the pilot showed 90% accuracy, build your P&L model on 70%. If the business case still works at 70%, it is robust. If it breaks, it is likely too fragile to fund.
The Cost of Verification (The Human-in-the-Loop Tax)
Generative AI is not deterministic software. It is creative based on what it knows. It makes things up. This means its output must be verified.
If you deploy an AI to draft legal responses, you have not removed the lawyer. You have changed the lawyer's job from drafter to reviewer.
Time to draft: 45 minutes.
Time to generate AI draft: 1 minute.
Time to review and correct AI draft: 30 minutes.
The saving is 14 minutes, not 44.
Reviewing is often harder than drafting. It requires a different type of concentration. It is easy for a bored human to glance at a plausible-looking AI document and sign it off without spotting the subtle error in clause 4.2.
The Red Pen Rule: Ask for the Workflow Diagram. If the diagram shows the AI output going directly to a customer or stakeholder without a verification step, reject it immediately for risk reasons. If it shows a verification step, ask if the cost of that human time has been netted off the projected savings.
The Integration Black Hole
Vendors love to say their tool "integrates seamlessly with your ERP."
To a salesperson, "integrates" means "we have an API."
To a Finance Director, "integrates" should mean "data flows bi-directionally without breaking the month-end reporting."
Most mid-market businesses run on legacy systems. Sage 200, older versions of Dynamics, or bespoke industry SQL databases. These systems were not built for modern API calls.
If the proposal allocates £10,000 for "Implementation," it is likely underfunded by a factor of five.
Real integration involves:
Mapping: Ensuring Gross Profit in the AI means the same as Gross Profit in the P&L.
Security: Ensuring the AI does not accidentally show payroll data to the warehouse manager.
Maintenance: Who fixes the link when the ERP gets a patch update?
The Red Pen Rule: Look at the ratio of Software Licence to Professional Services. In year one, for a complex AI integration, this ratio should be heavily weighted towards services (labour). If the software cost dwarfs the implementation cost, they are underestimating the complexity of your environment.
The Exit Strategy
Finally, the question nobody asks until it is too late: How do we turn it off?
If you deploy an AI agent to handle 50% of your customer service volume, and you subsequently reduce your headcount, you have created a dependency.
If the vendor doubles their price in year two (a common tactic), or if the model starts hallucinating due to a bad update, can you revert to the old way of working?
If you have fired the humans, the answer is likely no. You are captured.
The Red Pen Rule: Every AI proposal must have a section on Business Continuity & Exit.
If the AI fails, what is the manual fallback?
If we leave this vendor, can we export our data and our conversation history?
Do we have the cash reserves to re-hire staff if the automation fails?
The FD's Decision Matrix
You do not need to be technical to evaluate these proposals. You just need to be commercially cynical.
Use this simple approach to grade the proposal:
Revenue Logic: Are savings Cashable (bankable) or Theoretical (capacity)?
Pilot Validity: Does the data assume a perfect world or a messy one?
Net Time: Has the cost of human verification been deducted from the efficiency claims?
Integration Reality: Is there a realistic budget for connecting this to our legacy systems?
Reversibility: If this goes wrong, can we survive the rollback?
If a proposal can answer these five questions honestly and still show a positive return, sign the cheque. You have found a winner.
If it relies on "soft savings," "seamless integration," and "guaranteed accuracy," send it back. You just saved the company a fortune in failed implementation costs.
This is not about being the "Department of No." It is about being the "Department of Sustainable Growth." In the age of AI, that is the most strategic role you can play.