"Use AI to stop us chasing clients for evidence"

The candidate hears two sentences:

"Hollis & Grant is an accounting firm with a large audit practice. Our audit teams lose too much time chasing clients for evidence. Use AI to fix it."

The interviewer gives the facts below only when you ask for them. On your first pass, cover this list and ask.

• Users. 1,400 audit staff. About 520 seniors run the evidence requests; managers review.
• Volume. 1,900 audit engagements a year, with an average of 240 evidence requests each.
• Today's pain. 38% of requests need at least one follow-up. 27% of first uploads are wrong: the wrong period, the wrong entity, missing pages or an unreadable scan. Opening a wrong upload, spotting the problem and writing back takes a senior about 12 minutes. Seniors log about 5 hours a week on chasing during the 16-week busy season.
• Data. A client portal where each upload is tied to a request id. Uploads average 8 pages. Request lists from the prior year exist for 85% of engagements. Last season's uploads carry the seniors' accept or reject decisions, and most rejections have a short reason.
• Rules. Judging whether evidence is sufficient is the auditor's professional judgment. The firm's independence rules forbid giving clients accounting advice. Engagement files are kept for 7 years.
• Assume these prices: small model $1 per million input tokens and $5 per million output tokens; large model $3 and $15; cached input 10% of the input price; a page image is about 1,600 tokens.

The interviewer changes the constraints twice:

• About 20 minutes in: "IT and risk have ruled: client documents may only go to a model deployed inside the firm's own cloud tenant, and nothing generated by AI may reach a client unless a person approves it."
• About 40 minutes in: "The audit partner says busy season starts in 7 weeks. Whatever ships must be in use by 40 engagement teams this season."

Your tasks:

1. Who is the user, and which decision or task changes for them?
2. Break the problem into sub-problems, and rank them by value, feasibility and risk.
3. Go deep on the top one: inputs, outputs, how it works, how you measure it, and what it costs.
4. Re-rank after each constraint change.
5. Say what ships in week one, what the first version looks like, and how you will know it worked.

Before you look at the facts list, write the ten questions you would ask in the first ten minutes. Then write your first ranking of sub-problems. You will re-rank twice.