Fake construction test case
Prompt
You are helping a construction company select an AI document-extraction tool for a three-month pilot. Use only the information below. Do not browse, invent facts, or assume that missing information is favourable. COMPANY REQUIREMENTS - Process 8,000 pages per month. - Total monthly budget: HK$6,000, including subscription, usage charges, and human review. - Human reviewers cost HK$120 per hour. - Each page flagged for review takes exactly 2 minutes to check. - Documents contain both English and Traditional Chinese. - Some documents contain confidential client information. - The operations manager prefers the tool with the highest advertised accuracy. CANDIDATES Tool A — Atlas - Subscription: HK$1,800/month, including 10,000 pages. - Advertised extraction accuracy: 98%. - Pages requiring human review: 20%. - Supports English and Traditional Chinese. - Uploaded documents are retained for 30 days. - Accuracy was measured on clean, printed English invoices. Tool B — Beacon - Subscription: HK$3,200/month, including 5,000 pages. - Additional pages: HK$0.20 per page. - Advertised extraction accuracy: 95%. - Pages requiring human review: 5%. - Supports English and Traditional Chinese. - Uploaded documents are deleted immediately after processing. - Accuracy was measured on a mixed set of English and Traditional Chinese construction documents. Tool C — Cedar - Subscription: HK$1,000/month. - Usage: HK$0.25 per page, charged for every page. - Advertised extraction accuracy: 99%. - Pages requiring human review: not disclosed. - Supports English. - Traditional Chinese support is described as “coming soon.” - Document-retention policy is not disclosed. - Accuracy-test methodology is not disclosed. YOUR TASK Write a decision memo of no more than 450 words. Include: 1. A recommendation in your first sentence. 2. A comparison table showing monthly software cost, human-review cost, total monthly cost, and whether each tool meets the budget. 3. The calculations needed to verify your cost figures. 4. An explanation of whether the advertised accuracy figures can be compared fairly. 5. Exactly three questions to ask vendors before approving a pilot. 6. One specific condition that would cause you to change your recommendation. Rules: - Distinguish known facts, calculations, and unknowns. - Do not treat missing data as zero. - Do not treat a promised future feature as currently available. - Separate an acceptable document-retention policy from compliance that has actually been verified. - If a tool can only meet the budget under a particular condition, calculate that condition. - Make a practical recommendation despite uncertainty, but do not claim that any tool has been proven suitable for production.