You run the small-business division of a fictional regional bank. Growth has stalled and your team proposes FlexLine, an unsecured working-capital line for small firms. You must decide whether to launch, at what price, and how big. We'll then interview each form of AI on this same decision.
Proposed product
FlexLine working-capital line
Decision due today
CustomerSmall firms · 2–20 staff
Credit limitup to $50,000
Expected applications4,800 / year
Expected approval rate38%
Avg utilized balance$22,000
Fixed launch cost$1.2M
Competitor APR10.9%–13.9%
The numbers are estimates. Losses, fairness, competitor moves, and customer behaviour can all shift after launch.
Your unaided call
Decide before the machines do.
Pick the scale, then the APR you'd test.
55%
Why first? Every form of AI below will try to shape this call. Anchoring your own answer now is the only way to see what each one adds — and where it quietly moves you.
Candidate 1 · 1950s · Simulation
Model it as a formula.
Early computing let managers ask "what if" — write down the relationships you believe hold, then turn the dials. Find the price that maximizes expected first-year contribution, then switch to the recession preset and watch the answer flip.
Portfolio simulator
Assumptions
inputs → outcome
11.9%
4,800
2.8%
Set the dials, then calculate.
Interview card · Simulation
The recession preset flips a "profitable" launch to a loss. The model only knows the relationships you hand-coded — it can't tell you your assumption is wrong.
Candidate 2 · 1970s–80s · Expert system
Encode the credit officer as rules.
Lending policy captured as if–then logic, applied the same way every time. Test three applicants and move the coverage threshold to see how consistency becomes brittleness at the margin.
Rulebook
Eligibility rules
facts + rules → conclusion
1.25×
Applicant
Pick an applicant and run.
Interview card · Expert system
North Repair, 9 years and clean, is declined for missing the ratio by 0.01×. The rule is explainable — and still wrong at the margin.
Candidate 3 · 2012–2020 · Predictive AI
Estimate the risk from data.
Instead of hand-written rules, a model learns default patterns from 145,000 past facilities. Explore a borderline case — then switch on the new digital-only applicants the model never trained on.
Applicant features · North Repair
Inputs
like cases → likely outcome
684
32%
55
Digital-only firms now applying
underrepresented in training data
6.0%
Model output
Default probability · 12 months
within population
6.4%
A prediction — not an approval or a price.
0.79validation AUC
145ktraining facilities
8 mosince validation
Interview card · Predictive AI
Candidate 4 · 2022 → · Generative AI
Write the committee memo.
A generative model can turn calculations, rules, model output, and research into one executive memo in seconds. Choose what it may see, generate it, then turn on the evidence lens to see where the evidence actually ends.
Context supplied
What the model may use
context → plausible prose
Require citations & uncertainty labels
separate evidence from inference
Committee memo
Draft
SupportedInferenceUnsupportedOmission
Choose sources, then generate.
Interview card · Generative AI
Candidate 5 · 2023 → · Agentic AI
Give it tools and let it run.
The agent gets an objective — launch a pilot in 14 days — and access to bank systems. Decide what it may do alone and what needs a human gate, then run it and read the action trace.
Agent configuration
Objective: launch in 14 days
objective + tools → actions
Read reports (research, policy, model)
Configure a pilot product
Set the customer APR
Select & email 1,200 prospects
File the product disclosure
Committee approval before launch
required by Riverbend policy
Set the authority, then run.
Interview card · Agentic AI
Leave the gate off and let it price, email, and file on its own — it completes the objective and breaches authority it never held. Some of those actions can't be recalled.
Synthesis · the composable stack
Compose the layers, then stress-test.
No single form makes the decision. Assign each layer to the part of the FlexLine decision it should do — then pick a shock and see which layer notices and who must decide.
Which layer should do each job?
Stress test · which layer notices?
Pick a shock.
The deeper point
The history isn't a ladder from primitive to superior. Each era adds a different operation; contemporary systems are strongest when those operations are deliberately composed and each one's distinct claim stays visible.
Practice · a light self-check
Interview a mystery system.
Four quick calls. For each, use the six-question interview to decide. You'll see the answer and why immediately.
Your decision card
What would you authorize now?
Your final answer names the product decision and the evidence, the human checkpoint, and the authority boundary behind it.
Riverbend Bank · FlexLine
Product launch decision
classroom simulation
Fill the card from your choices.
Your first callnot recorded
Confidence then → now55% → 55%
Strongest quantitative insightnot tested
Most dangerous blind spotnot identified
Authority boundary for the agentnot configured
Stress testnot run
55%
The principle: new AI expands what an organization can calculate, infer, express, and do. It does not decide what the organization should value, or who holds authority.
Keep this. It's the first artifact of your A4 delegation portfolio — the tested, governed design you'll build across the term.