Frameworks & Templates5 min read

Separate Generation From Decision: A Two-Pass AI Template

Use AI to expand and challenge options, then make and record the accountable human choice in a separate pass.

Bokili EditorialPublished August 28, 2026· Verified August 28, 2026

AI expands one brief into several options before a human compares criteria, challenges assumptions and records a decision.

AI is very good at making options feel finished. Ask for a recommendation and it can produce a neat shortlist, a preferred choice and the reasons in one smooth answer. That convenience hides an important boundary: generating possibilities is not the same job as deciding which one the organisation should adopt.

When both jobs happen in one pass, the first plausible option can anchor the rest of the answer. Criteria appear after the recommendation, assumptions remain implicit and the human reviewer is left approving a conclusion rather than making a decision. A better workflow separates the work into two passes: use AI to widen and challenge the field, then make and record the accountable human choice.

The boundary to preserve

AI may help create, compare and challenge options. The accountable person still chooses the criteria, accepts the trade-off and owns the decision.

Why one-pass recommendations are hard to review

A recommendation combines several activities: understanding the brief, inventing options, selecting criteria, weighing evidence and accepting consequences. If the AI performs all of them inside one answer, a reviewer cannot easily see where a weak assumption entered or whether a different criterion would change the result.

NIST’s AI Risk Management Framework treats governance, context mapping, measurement and management as connected activities across the AI lifecycle. Its Generative AI Profile also notes that some uses warrant additional human review, tracking and documentation. For an everyday business decision, this can be lightweight: separate the exploratory artefact from the decision record.

	One-pass recommendation	Two-pass decision
Opening move	Ask for the best answer	Ask for materially different options
Criteria	Often inferred inside the response	Set and checked by the accountable person
Challenge	Applied mainly to the preferred option	Applied to every surviving option
Output	One polished recommendation	Option set plus a separate decision record
Ownership	Blurred by fluent language	Named at the choice point

Use two passes: EXPAND–CHALLENGE, then CHOOSE–RECORD

The four moves in a two-pass AI decision

EXPAND the field

Give the AI the outcome, constraints and evidence. Ask for three to five options that differ in mechanism, not merely wording. Require a ‘do nothing’ or manual alternative when it is genuinely available.

CHALLENGE each option

Test every option against the same criteria. Ask what must be true, what could fail, what evidence is missing and which stakeholder bears the downside. Do not let the AI silently eliminate an option.

CHOOSE outside the generation pass

The accountable person sets the weights, resolves value conflicts and selects, combines or rejects the options. If the evidence is insufficient, the decision is to gather more evidence—not to demand a more confident answer.

RECORD the decision

Capture the chosen option, decisive criteria, rejected alternatives, unresolved assumptions, owner and review trigger. Keep it short enough to be read when the decision is revisited.

The method does not assume that AI is neutral. It makes influence easier to inspect. The option set shows what the system surfaced. The decision record shows what the human accepted and why. Microsoft Research’s human–AI interaction guidelines similarly emphasise supporting correction, dismissal and appropriate human control when systems are uncertain or wrong.

Worked example: choose a manager practice format

Reading is a start. Practice makes it stick.

Start learning

Suppose an L&D lead must help 120 managers practise safe AI delegation. The session must fit 30 minutes, use no sensitive employee data and work without buying another tool. Instead of asking, “What is the best workshop?”, the lead runs two passes.

Pass 1 — build and challenge the option set

1
1. State the decision brief
Outcome: managers can define a task, an input boundary and a review step. Constraints: 30 minutes, existing meeting software, synthetic examples only.
2
2. Generate different mechanisms
The AI proposes a live scenario drill, a peer review clinic, a short demonstration followed by practice, and a manager-led team huddle kit. The options differ in who practises, who gives feedback and when transfer happens.
3
3. Apply one challenge grid
For each option, examine practice time, facilitation load, evidence of skill, access needs and the risk that managers watch rather than act.
4
4. Expose missing evidence
The option set flags two unknowns: available facilitators and whether every manager can join a live session. Those questions return to the owner before selection.

Pass 2 — make and record the choice

1
1. Set the decisive criteria
The L&D lead gives practice time and observable behaviour more weight than delivery convenience.
2
2. Choose with a named trade-off
The scenario drill is selected because every manager must make the three target decisions. It needs more facilitation, so groups are capped and a second session is planned.
3
3. Record the review trigger
After the first cohort, the lead will review completion, common mistakes and facilitator load. If the cap creates a long queue, the peer clinic remains the fallback.

A decision record that fits on one screen

Copy these seven fields

Decision and accountable owner
Outcome and non-negotiable constraints
Options considered, including the manual or no-change option when relevant
Criteria used and any explicit weighting
Chosen option and the trade-off accepted
Unresolved assumptions or evidence gaps
Review date or event that could reopen the decision

This record is not a transcript of every prompt. It is the minimum evidence needed to understand the choice later. Keep exploratory AI output as working material; keep the decision record as the organisational artefact.

Know when the template is too light

A two-pass template helps with ordinary, reversible business choices. It is not a substitute for formal controls when a decision affects safety, legal rights, employment, credit, health or other high-consequence outcomes. Those cases may require approved systems, specialist review, documented testing and governance beyond the person making the immediate choice.

Run a ten-minute two-pass decision

Pick one small, reversible decision you need to make this week.
Write the outcome, three constraints and the evidence already available.
Ask an approved AI tool for three materially different options plus one missing-evidence question for each.
Challenge every option against the same three criteria.
Close the AI response and choose, reject or pause as the accountable person.
Write the seven-field decision record and name the review trigger.

Better options do not remove the need to decide

The value of AI is not that it can spare people every choice. It can widen the field, expose assumptions and make trade-offs easier to inspect. The two-pass template protects the moment where assistance ends and accountability begins.

Bokili’s short workplace missions can help teams practise that boundary repeatedly: create options, challenge them with evidence, then make a visible human decision. The useful skill is not obtaining a recommendation. It is knowing how to use one without surrendering the choice.