Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Run each part of the flow through a short test: is the input varied and messy, can the task tolerate occasional errors (or can they be caught), is it pattern recognition rather than policy, is there data to evaluate it, and is the value per use higher than the cost? In expense approval, reading receipts passes; applying spending policy and moving money fail and should stay as rules.
How to approach it
Ask what problem the stakeholder wants solved: slow approvals, policy violations, employee frustration, or finance workload. Then break the flow into steps and test each one, because "should this use AI" is almost never a yes or no for a whole product.
A strong answer
The expense flow has roughly five steps: an employee submits a receipt, the details are captured, the expense is checked against policy, a manager approves, and finance reimburses.
A five-question test, applied to each step:
- Is the input varied? Receipts come in every format, language and photo quality. Policies are fixed text.
- Is an occasional error tolerable or catchable? A misread amount can be shown to the employee to confirm. A wrong reimbursement is money out the door.
- Is it pattern or policy? Recognising "this is a hotel bill" is pattern. "Hotels above ₹8,000 a night need director approval" is policy, and policy belongs in rules that can be audited.
- Is there data to evaluate it? Years of past receipts with their correct categories and amounts exist. That makes extraction measurable.
- Does the value per use exceed the cost? Saving an employee three minutes per receipt across thousands of receipts a month easily covers a fraction of a cent of model cost.
| Step | Varied input | Errors catchable | Pattern or policy | Verdict |
|---|---|---|---|---|
| Read receipt: merchant, date, amount, category | Yes | Yes, employee confirms | Pattern | Model |
| Check against policy | No | Must be exact | Policy | Rules |
| Flag unusual expenses (duplicates, odd patterns) | Somewhat | Yes, human reviews | Pattern | Statistics first, model to explain |
| Manager approval | Not applicable | Not applicable | Judgment | Human, with a summary |
| Reimburse | No | Must be exact | Policy | Rules |
Read diagram description
An expense flow in five steps with a verdict for each: reading the receipt uses a model and the employee confirms the fields; checking policy uses rules because it must be exact; flagging unusual expenses uses statistics first and a model to explain, reviewed by a person; approval stays with the manager, helped by an AI summary; reimbursement uses rules.
So "add AI" becomes something concrete: receipt capture by model, with the employee confirming the extracted fields; policy checks as rules; an anomaly flag that a person reviews; and an approval screen that summarises each expense and any policy issues so the manager can approve in seconds.
The general lesson: rules beat models whenever the logic can be written down and must be exact. Models earn their place on messy input and pattern recognition, ideally where a person or a deterministic check can catch their errors. Google's Rules of Machine Learning open with the same advice: "Rule #1: Don't be afraid to launch a product without machine learning." Its People + AI Guidebook adds the cost test: AI is probably not better "if the cost of errors is very high and outweighs the benefits of a small increase in success rate" (user needs).
There is one place where the line moves, and a strong answer names it. Real expense policies are not all exact thresholds; they contain phrases like "reasonable", "client-facing" or "business purpose required". That part of policy is prose, and a model can read it. Ramp's Policy Agent applies a company's written expense policy to each charge while "reviewers always retain final authority", and the company reports about 85% fewer manual reviews at over 99% accuracy for early customers (company-reported, checked November 2026). So the split I would draw is: thresholds and limits as rules; judgment-laden policy text as a model recommendation that a person approves; money movement as rules.
What interviewers probe next
"The stakeholder still wants the model to approve expenses." Show them that approval is policy plus judgment. Automatic approval for small, in-policy, correctly extracted expenses is fine, and it is a rule applied to the model's extraction, not a model decision.
"How would you know the receipt extraction is good enough?" Field-level accuracy on a labelled set of past receipts, by receipt type and language, plus the share of extractions employees correct in production.
How it's scored
Weak: Says yes to AI everywhere, or rejects it without testing the steps.
Hire: Breaks the flow into steps and identifies receipt reading as the right AI use.
Strong Hire: Applies an explicit test per step, keeps policy and money movement in rules, and designs human confirmation where model errors can be caught.
Common mistakes
Asking a model to apply policy that could be encoded as exact rules, or, the opposite mistake, trying to hard-code judgment-laden policy text that a model could flag for a reviewer.
Evaluating "AI for the whole flow" instead of each step.
Ignoring whether errors can be caught before they matter.
See the AI fit test.