Assumptions, uncertainty, and AI as a thought partner
Separate evidence from assumptions, check AI-generated claims, and choose a limited test before committing.
Challenge a decision's assumptions, using AI when helpful, while independently verifying claims, protecting confidential information, and choosing a reversible next test.
Overview
A manager at established retailer Elm Furniture wants to offer Saturday delivery. Fifteen customers have asked about it. An AI-generated draft recommends launching immediately because “customer interest proves demand.”
The manager needs evidence that customers will pay.
All businesses, figures, and sample AI responses in this lesson are fictional.
Recall Lesson 016: Why did Orchard Photography's $900 advertising payment fail to settle whether it should run the workshop?
Facts and assumptions
A fact is supported by evidence you can inspect. An assumption is something you treat as true while reasoning, even though it has not been established for this decision. Uncertainty means you do not know exactly what will happen or which explanation is correct.
Elm's 15 recorded inquiries show interest. The claim that 12 customers will pay $60 for delivery is an assumption.
Write assumptions so someone can test them. “Customers will like it” is vague. “At least ten customers will book a $60 delivery for the trial Saturday” specifies behavior, price, quantity, and timing.
Prioritize uncertain assumptions that could change your decision. Ask what result would make you reject or revise the proposal.
Use AI to challenge the proposal
You can ask a generative AI tool, which produces content from instructions, to identify missing assumptions, suggest alternatives, or challenge a calculation.
A useful request using fictional data is: “Separate supplied facts from assumptions. Identify three assumptions that could reverse this recommendation. Suggest evidence for each and one small test. Do not invent customer findings or probabilities.”
The instruction does not ensure an accurate answer. NIST's July 2024 Generative AI Profile identifies risks including confidently stated false content, data privacy problems, and human overreliance. NIST source
Verify important claims independently. Open cited sources and check whether they support the claim. Recalculate numbers with a calculator. Confirm costs with the supplier and demand with customers. Asking the same AI to confirm its answer provides no independent evidence.
Use tools approved for the information involved. Share only necessary details; replace confidential records with synthetic examples where possible. Removing names alone may leave identifying details. Check the tool's current data-handling terms and your organization's rules before sharing sensitive information.
Worked example
Elm can test one Saturday using a trained delivery crew for $240, plus $30 per completed delivery. Each customer pays $60. The crew can handle up to 12 deliveries. Elm has $600 available after reserving cash for other obligations.
Assume no other additional costs, no displaced work, and full collection for completed deliveries. Ignore taxes. The crew requires no commitment beyond the trial.
Let n mean completed, paid deliveries. Additional revenue is $60 × n. Additional cost is $240 + ($30 × n). The amount left is therefore ($60 − $30) × n − $240, or $30 × n − $240.
| Completed deliveries | Revenue | Additional costs | Amount left |
|---|---|---|---|
| 6 | $360 | $240 + $180 = $420 | −$60 |
| 8 | $480 | $240 + $240 = $480 | $0 |
| 12 | $720 | $240 + $360 = $600 | $120 |
Elm needs $240 ÷ $30 = eight completed deliveries to cover the trial's additional costs. These are scenarios: possible outcomes under stated inputs. They are not forecasts or probabilities. Fifteen inquiries do not establish which row will occur.
A reversible test limits the commitment so the manager can stop or change course after learning. It still consumes money and time. The Institute for Healthcare Improvement recommends small tests with predictions, data collection, and review before broader implementation. IHI source
Elm could seek ten confirmed bookings before reserving the crew, leaving a two-delivery cushion above the eight-delivery threshold. Confirm cancellation arrangements first. Record bookings, cancellations, completed deliveries, actual costs, and damage complaints; review on Monday before booking another Saturday.
Ten bookings are a proposed management threshold, not a guarantee. The $600 covers the stated maximum costs. Damage or overtime could change that. One Saturday cannot establish year-round demand.
Exercise
Allow 7–9 minutes. You own Birch Craft Studio and are considering an evening class. Twenty people asked for information, but nobody has booked. The price would be $50 per person, with $20 of materials per attendee and $120 of instructor cost for one evening. Capacity is eight people.
You have $280 available after other obligations. Assume no other additional costs, displaced work, or taxes, and full payment. You can arrange one evening without a continuing commitment.
A sample AI response says: “A standard 30% inquiry-to-booking rate means six attendees and $60 left after costs, so commit to a full month.” It supplies no source for the 30%. A month would require a separate $900 nonrefundable room commitment, excluded from the one-evening figures.
- Separate a verified calculation from an unsupported assumption in the AI response. Calculate the one-evening result with six attendees and with three.
- Calculate the attendance needed to cover the evening's additional costs. Explain why the monthly recommendation is premature.
- Choose a reversible next step. Name what to measure, when to review, and what would make you stop or revise.
- Write one follow-up AI instruction and name information you would keep out of an unapproved tool.
Attempt the exercise before reading on.
Answer
Six attendees would leave (6 × $50) − (6 × $20) − $120 = $60. Three would leave $150 − $60 − $120 = −$30. The arithmetic is correct under the supplied costs; the 30% booking rate is unsupported.
Each attendee supplies $50 − $20 = $30 toward the instructor cost. Covering $120 requires $120 ÷ $30 = four attendees. The eight-person trial costs at most $120 + (8 × $20) = $280 under the stated assumptions.
The $900 monthly commitment exceeds available funds, and sustained demand is unverified.
Offer one class and seek at least five confirmed bookings before committing to the instructor, with cancellation terms checked. Five provides a small cushion above four. Measure bookings, attendance, payments, costs, and participant feedback. Review after the class before expanding; revise or postpone if bookings fall below the threshold, and investigate if actual costs erase the cushion.
A useful instruction is: “Treat 30% as unverified. Compare three, four, and six attendees using only supplied costs. Identify evidence needed before a monthly commitment.” Keep customer contact details and private financial records out of an unapproved tool. Verify the calculation yourself and test demand through actual bookings.
Recall answer: The $900 was unrecoverable under either choice. Orchard needed to compare future costs, benefits, and the alternative use of its Saturday.
Applying this lesson
Name the assumption most likely to change your decision, identify evidence that would challenge it, and choose a limited next test. AI can help generate questions; you remain responsible for checking the evidence and deciding what to do.
Sources
Comments
Loading comments…