Language and display
BE THE BOSS OF AI.ADRIAN DUNKLEY ↗
← Daily critical thinking

2026-11-16 · A CRITICAL THINKING LUNCHEON

The Bank Job

A million clients. An eight-month plan. A vendor promising two. What would you approve?

An origami bank beside copper ribbons, teal fibre strands and a paper clock.
A paper bank at the end of a hurried upgrade.

The chair pushed the schedule toward Maya Croft. Her bank served one million clients. Its copper circuits and analog phones needed replacing with fibre; the accounts were already digital.

The board paper showed 12 million transactions a month. Revenue averaged US$15 million monthly, or US$500,000 a day using a 30-day planning month. A day without service would reach well beyond the branches.

The upgrade would cost US$20 million. Finance estimated US$300,000 in monthly operating savings after completion. “We need those savings this year,” the chair said.

Maya’s team allowed eight months for branch surveys, banking tests, training and recovery rehearsals. The vendor promised two.

Its generative AI tool had drafted 90% of the specification. It was designed for air traffic control. “It’ll work here,” the vendor assured her.

The tool had never received the bank’s network map, peak payroll volumes or recovery logs. The demonstration used sample payments below payday demand.

Maya opened the risk worksheet. It assumed a 30% chance that, within 12 months, reviewers would find a bank-specific error or omission in the AI-drafted specification capable of disrupting payments or recovery. The note said “planning assumption, not measured.”

Her engineer opened the test plan. “Two months gets the links installed. The payroll test and recovery rehearsal are blank.” Nobody had shown whether recovery would lose or duplicate payments.

The AI had no access to live systems. Maya’s signature would authorize the rollout. The chair asked, “Can we announce the earlier date?”

Compare your decision with the recommendation

Test the promise before the rollout

Approve a limited pilot with payment and recovery checks before expansion. Ask both teams to show the work behind their schedules. Run tasks in parallel where the evidence permits it. Keep a tested way back, and give a named bank officer authority to stop each release.

Maya owns the release decision. The bank owns the operational risk, even when the vendor supplies the system.

Read the clues

  • 90% describes how much the AI drafted. It tells Maya nothing about accuracy.
  • The missing network map, peak payroll volumes and recovery logs leave bank-specific requirements unchecked.
  • Installing fibre in two months does not establish that payments will survive payday demand or a failed connection.
  • Eight months also needs evidence: staffing, branch dependencies and test duration. The internal estimate is open to challenge.
Four possible outcomes over 12 months from approval
Maya’s choiceNo observed material impactMaterial impact
Tested pilot

Complete month 8. Original eight-month plan. Added cost: US$1.5m. Year-one savings: US$1.2m. Net outlay: US$20.3m.

Tests pass. No material service incident occurs during the first year. Later faults remain possible.

Complete month 10. 2 months later. Added cost: US$3.5m. Year-one savings: US$0.6m. Net outlay: US$22.9m.

The pilot exposes a payment fault. Maya stops expansion; repairs add cost and delay. Customer payments remain protected in this example.

Two-month rollout

Complete month 2. 6 months saved. Added cost: US$0.25m. Year-one savings: US$3m. Net outlay: US$17.25m.

Ordinary traffic runs smoothly. Payday and recovery cases remain untested. A quiet year does not validate the approval.

Complete month 2. 6 months saved. Added cost: US$9m. Year-one savings: US$3m. Net outlay: US$26m.

Payday overload causes a 24-hour outage affecting 180,000 clients. Payments fail or arrive late; recovery and remediation add cost.

All figures and outcomes are illustrative teaching assumptions. Net outlay = US$20m base + added cost − savings. Savings start after completion at US$300,000 a month. Material impact means at least US$2m additional cost, over 10,000 clients disrupted for more than four hours, or completion after month 9.

What the numbers can tell you

The fast plan offers US$1.8m more first-year operating savings than month 8 completion. Each cell describes a possible outcome. The table supplies no outcome probabilities. The 30% assumption concerns a bank-specific specification error. It is not a 30% outage probability and cannot be assigned to every downside cell.

To calculate expected cost, first estimate the probability of each outcome under each plan using bank-specific tests and recovery evidence. Until then, examine how the decision changes across plausible probabilities. A missed payroll payment can hurt a customer even when the bank’s net outlay looks attractive.

12 million monthly transactions average 400,000 a day. Revenue averages US$500,000 a day; neither figure is the value lost in an outage. The US$9m downside allowance covers recovery, reconciliation and customer remediation, including lost net fee income. Transfer values, fines and future lost business are excluded.

Apply the four Ds

Delegation
Let AI draft. Keep release approval with accountable bank staff.
Description
Supply the network map and peak payment loads. Set recovery requirements and limits on system access.
Discernment
Check the specification against banking tests. Ask for the evidence behind both schedules.
Diligence
Record approvals, rehearse recovery and reconcile payments after each stage. Stop expansion when a release check fails.

4Ds: Rick Dakan, Joseph Feller and Anthropic. Framework and attribution.

Bring this to your next meeting

Choose one pending upgrade. Ask the team to show a peak-load test, a timed recovery rehearsal and the name of the person who can stop release. Agree the pass criteria before discussing a launch date.