Thirty ideas, one budget

The situation

A simulated mid-sized bank has thirty AI proposals from across the business and delivery capacity for four in the next twelve months. Every proposal has a sponsor, and each sponsor can show a vendor demo. The executive team needs to choose without turning it into a contest of who argues best.

The evidence

In two working sessions the proposals were mapped to the processes they would change, and each process was given a baseline: monthly volumes, handling time, error and rework rates, and who owns the data. 81% of the evidence needed was available. Where a figure was missing, it was recorded as missing rather than estimated.

What the assessment showed

Eleven proposals overlapped and were merged, leaving nineteen candidates in nine process areas. Four process areas had both the evidence and the readiness to proceed. Two of the most popular proposals did not. Mortgage advice through a customer-facing assistant carries a high oversight requirement, and the bank has no baseline of why customers call. AI-drafted credit memos depend on source data that was only 54% complete.

The decision

• Invest now payment exception handling
• Governed pilot KYC document review, internal service desk triage, supplier invoice matching
• Defer mortgage advice assistant, credit memo drafting, and nine others, each with a written rationale
• Keep human-led four process areas where the oversight requirement outweighs the value

Estimated value of the funded four: €0.6–1.1m in annual effort, with every assumption stated in the record.

What would change it

The mortgage assistant is re-scored once eight weeks of call-reason data exist. Credit memo drafting is re-scored when the data owner confirms the source data. If the payment exception volume falls by more than a third, the Invest now decision is reviewed.