A Human–AI Co-Production Workflow for Expert Reports That Actually Matter

Most think tank reports fail for a simple reason: they optimise for publication, not for decisions. More pages, more citations, more polish. Less impact.

AI does not fix this by making writing faster. It fixes it by forcing a rethink of how thinking itself is organised. The correct move is not automation. It is division of cognitive labour.

What follows is a general-purpose human–AI workflow for think tanks that want their work to shape policy, strategy, and institutional action rather than gather dust.

First Principle: Reports Exist to Enable Decisions

A report is not knowledge. It is a decision-support artefact.

If no identifiable actor can act differently after reading it, the report has failed regardless of technical quality.

This workflow is designed to align human judgement and machine capability around that invariant.

Role Clarity: No Overlap, No Confusion

AI excels at:

1. Scale
2. Recall
3. Synthesis
4. Consistency
5. Stress-testing logic

Humans excel at:

1. Judgement
2. Political and institutional context
3. Normative choice
4. Legitimacy and accountability

Blurring these roles produces noise. Clear separation produces leverage.

The Workflow

Phase 1: Decision Framing (Human-Led)

Objective: Define the real problem before analysis begins.

Humans:

1. Specify the decision the report must inform.
2. Identify who has authority to act.
3. Clarify constraints, incentives, and red lines.
4. Define what “useful” means in operational terms.

AI:

1. Translate the mandate into explicit sub-questions.
2. Surface comparable reports and prior failures.
3. Propose a decision-aligned outline.
4. Expose hidden assumptions in the brief.

Output: A decision-centred brief approved before research starts.

This phase prevents elegant irrelevance.

Phase 2: Evidence Expansion (AI-Led, Human-Governed)

Objective: Increase signal without increasing noise.

AI:

1. Aggregate academic, policy, and grey literature.
2. Map datasets, indicators, and methodologies.
3. Highlight contradictions and gaps.
4. Produce annotated evidence summaries.

Humans:

1. Judge relevance to context and political economy.
2. Exclude technically valid but misleading evidence.
3. Add tacit and experiential knowledge.
4. Set admissibility standards.

Output: A curated evidence base with provenance and confidence levels.

This phase replaces selective citation with disciplined inclusion.

Phase 3: Analysis and Sensemaking (Joint)

Objective: Turn information into defensible insight.

AI:

1. Identify patterns across geographies and time.
2. Simulate scenarios under explicit assumptions.
3. Generate alternative analytical framings.
4. Test internal coherence.

Humans:

1. Select the lens that fits institutional reality.
2. Distinguish correlation from causal relevance.
3. Decide which lines of inquiry to drop.
4. Interpret findings through lived constraints.

Output: A small set of insights that survive scrutiny.

This phase prioritises clarity over comprehensiveness.

Phase 4: Drafting the Argument (AI-Drafted, Human-Owned)

Objective: Produce clarity without surrendering authority.

AI:

1. Draft sections from approved insights.
2. Enforce logical structure and consistency.
3. Maintain traceability from claim to evidence.
4. Produce multiple audience-specific summaries.

Humans:

1. Rewrite for tone, legitimacy, and precision.
2. Insert contextual nuance and institutional voice.
3. Own every claim explicitly.
4. Remove false certainty.

Output: A report that reads as authored, not generated.

Authorship remains human. AI handles scaffolding.

Phase 5: Red-Teaming and Risk Control (Human-Led)

Objective: Avoid embarrassment, backlash, and strategic error.

AI:

1. Challenge arguments systematically.
2. Identify weak claims and missing counterpoints.
3. Surface alternative interpretations.
4. Check for contradictions and overreach.

Humans:

1. Decide which critiques matter in the real world.
2. Balance technical accuracy with feasibility.
3. Make final judgement calls where evidence is incomplete.

Output: A hardened report with explicit limitations.

This phase builds credibility by showing restraint.

Phase 6: Decision Translation (Human-Owned)

Objective: Ensure uptake, not applause.

AI:

1. Generate policy briefs, decks, and summaries.
2. Tailor outputs to different stakeholder profiles.
3. Adapt language and framing without changing substance.

Humans:

1. Sequence release and engagement strategically.
2. Brief decision-makers directly.
3. Defend conclusions in live settings.

Output: Actionable artefacts aligned to real timelines.

Reports do not move the needle. People do.

Governance Rules That Matter

1. Humans retain final authorship and accountability.
2. Every AI contribution is auditable.
3. No AI output enters final publication without human sign-off.
4. Normative and political judgements are human-only.
5. Feedback from real-world impact feeds the next cycle.

Without governance, AI amplifies mediocrity. With governance, it compounds judgement.

What This Changes for Think Tanks

1. Fewer reports. Higher leverage.
2. Less performative expertise. More decision relevance.
3. Reduced cognitive overload for senior researchers.
4. Faster synthesis without sacrificing rigour.
5. Institutional credibility strengthened, not diluted.

The Core Insight

AI is not a junior researcher.

Humans are not editors of machine text.

This workflow treats think tanks as decision engines, not publishing houses.