Responsible AI Practice4 min read

Before You Trust an AI Recommendation, Log Its Assumptions

A recommendation can be well written and still rest on a hidden premise. Use a five-field log to expose, test and own the assumptions that change action.

Bokili EditorialPublished September 5, 2026· Verified September 5, 2026

Four kinds of input—fact, assumption, unknown and test needed—separated before a human decision gate

An AI recommendation can sound careful, cite evidence and still rest on an assumption nobody has named. A sales forecast assumes the pipeline will convert as usual. A staffing plan assumes demand will stay flat. A supplier summary assumes missing detail means no exception. The output may be fluent; the decision beneath it is fragile.

Before a manager approves the next step, the important move is not another rewrite. It is to separate what is known from what has been assumed, decide which assumption could change the outcome, and assign a check. That turns hidden reasoning into a small, reviewable record.

Why a plausible answer can hide a weak premise

NIST’s AI Risk Management Framework asks organisations to document intended uses, context, assumptions and related limitations. It also says knowledge limits and the way people oversee AI output should be documented well enough to support subsequent decisions. These are system-level outcomes, but the behaviour scales down to everyday work: make the premise visible before it becomes a commitment.

The Generative AI Profile adds a practical warning. Generative systems can produce confabulated content that reads as if it were true. A reviewer therefore needs to check both the answer and the ground beneath it. A correct calculation built on the wrong time period is still the wrong basis for action.

	Ordinary answer review	Assumption-aware review
Main question	Does the output look supported?	What must be true for this advice to hold?
Evidence	Check claims and calculations	Check claims, premises and missing context
Uncertainty	Leave it inside polished prose	Name it and assign a test
Decision	Approve or reject the output	Approve, test, narrow or stop

Use a five-field assumption log

One card for each decision-changing assumption

Statement

Write the premise in one testable sentence: “Demand will remain within ten per cent of the current run rate.”

Basis

Mark whether it comes from a verified fact, an interpretation, an assumption or an unknown. Do not use “AI says” as a basis.

Consequence

State what changes if the premise is wrong: cost, deadline, safety, customer promise, rights or quality.

Check

Name the smallest useful check: open the source, ask the owner, run a sensitivity range or collect a new sample.

Owner and time

Assign a person and a date. An important assumption with no owner is only a warning, not a control.

The log is not a transcript of every thought. Record only premises that could change the recommendation or the next action. Low-impact drafting may need none. A pricing decision, policy change or safety-related recommendation may need several.

Worked example: a support staffing recommendation

Imagine an operations team asks an AI tool to recommend next month’s support staffing. The approved data shows that ticket volume rose for six weeks, average handling time fell, and two people are due to leave. The system recommends adding three temporary agents.

Reading is a start. Practice makes it stick.

Start learning

The arithmetic may be reasonable, but the recommendation rests on unspoken premises: the rise will continue, the handling-time improvement is stable, the departures will happen on schedule, temporary agents can reach useful proficiency quickly, and the existing queue definition has not changed.

The team creates three cards. The highest-consequence assumption is that volume will stay elevated. Its basis is a six-week trend, not a confirmed forecast. If wrong, the team may over-hire. The check owner compares the trend with the product launch calendar and runs low, middle and high demand scenarios before Friday. A second card covers ramp-up time; HR owns the check. The recommendation changes from “hire three” to “approve a two-week option, then trigger it only if the middle scenario is reached.”

Turn the answer into a decision-ready recommendation

1
1. Restate the decision
Write what the recommendation will authorise, spend, promise or change.
2
2. Ask what must be true
Extract every premise that materially affects the recommendation. Keep the list short.
3
3. Classify the basis
Separate verified facts, interpretations, assumptions and unknowns. Link facts to their source.
4
4. Rank by consequence
Check the premise whose failure would most change cost, safety, rights, timing or quality.
5
5. Assign a check or boundary
Verify it, test a range, narrow the recommendation or stop. Record an owner and date.

Do not ask the AI to certify its own assumptions

The same model can help propose possible premises, but it cannot turn an unsupported premise into evidence. Use it to widen the review question: “List the conditions this recommendation appears to depend on.” Then a person must verify the important items against approved sources and domain knowledge.

Microsoft’s human–AI guidance recommends scoping the service when the system is uncertain, making its behaviour understandable and supporting efficient correction. An assumption log brings those ideas into the work record. It shows why the recommendation holds, where it may fail and how a reviewer can correct it.

Connect the log to the wider workflow

Use the assumption log after you verify material claims and before the work becomes a decision hand-off. Pair it with the two-pass generation and decision template and an omission check for AI summaries. Each tool protects a different failure point.

Run a ten-minute assumption check

Choose one low-risk AI recommendation you have not acted on.
Write the decision it would trigger.
Complete this sentence three times: “This advice holds only if…”
Classify each statement as fact, interpretation, assumption or unknown.
Pick the highest-consequence premise and assign one check, owner and date.
Rewrite the recommendation with the result, boundary or stop condition visible.

Trust grows when the premise travels with the answer

A polished recommendation should not outrun its grounds. Keep the few assumptions that change action beside the evidence, the reviewer and the decision. Bokili’s short workplace missions help teams practise this kind of visible judgement on realistic tasks, with feedback and a second attempt before the habit reaches live work.