Start with what is known.
Use relevant research on judgment, groups, creativity and goal pursuit. Read the task, method and limits behind the headline.
Science-led. AI-powered. Keeping it human.
Better thinking needs more than a better model.
The question is not only what AI can do. It is when it enters the work, whose knowledge it can see, and who remains responsible for the choice.
Our aim is to put stronger thinking tools within more people’s reach—while preserving the independent judgment a team needs. Research informs the sequence. Evidence must keep shaping it.
Why the interaction mattersWhat science-led means here
A reason for the design.
A responsibility to test it.
Make the translation visible.
Explain why a finding informs a workflow choice. Keep a design hypothesis distinct from a demonstrated effect.
Let the method evolve.
Test the choices we add, including their costs. Revise when better evidence warrants it, and report results that challenge the product.
Published studies support component mechanisms. The complete Sikizana sequence has not yet been validated in a published outcome study.
The new design problem
More capable individuals.
A more capable team?
That second step cannot be assumed.
Professional writing
In a randomized experiment with 453 professionals, ChatGPT reduced task time by 40% and improved evaluated quality by 18%.
Noy & Zhang · Science, 2023 ↗Customer support
A study of 5,172 agents found a 15% average gain, rising to 30% for less experienced and lower-skilled workers. The highest-skilled group had small quality declines.
Brynjolfsson, Li & Raymond · QJE, 2025 ↗Defined tasks, different study designs. These results support broader access to useful assistance; they do not mean everyone improves equally or that the gains transfer automatically to organizational decisions.
The interaction changes the result.
Two experiments make the design question concrete.
AI ideas for short stories
Better individual stories.
More similar stories.
Access to AI ideas improved assessed creativity, especially for writers with lower baseline creativity. But the resulting stories became more similar to one another.
Doshi & Hauser · Science Advances, 2024 ↗AI examples in evolving brainstorming
A wider collective range.
No individual creativity gain.
In a different experiment, high exposure to AI examples increased collective idea diversity without improving average individual creativity.
Ashkinaze et al. · ACM Collective Intelligence, 2025 ↗Different tasks and interaction designs—not conflicting replications. Similarity in stories is not the same as organizational groupthink. The contrast gives us a reason to test how assistance is introduced, rather than assume AI always broadens or narrows thinking.
Keeping it human, in practice
People contribute. People challenge.
People commit—and can change course.
Our sequence starts with a person’s own first words. Optional AI questions and expansion follow; colleagues’ unreleased reflections stay outside that private assistance. Shared synthesis follows release and remains open to correction.
Later brainstorming can draw on a wider AI contribution, clearly identified as such. The aim is to add possibilities without replacing the independent thinking the process needs.
Human approval is not a test of truth. A consulting experiment found gains on tasks within AI’s capabilities and worse correctness on a task outside them. Evidence checks, challenge and a route to decline advice still matter. Dell’Acqua et al. · Organization Science, 2026 ↗
Compare a direct AI prompt with the shared workflow →From the first view to the work that follows
Why the order matters.
The framework connects research with explicit design choices. Open each step to see the finding, the translation and its limits. The pre-mortem belongs when there is a plan to examine; it need not precede every first reflection.
01PrepareProtect the first view.
Published finding
In an estimation experiment, seeing other people’s answers narrowed the range of estimates and increased confidence without improving collective accuracy. Separate group-task research associated more balanced conversational turns with collective performance.
Our design response
Begin with each person’s own observations, before colleagues’ positions or an AI-generated starting answer. Offer a route for people whose knowledge might otherwise lose out to airtime.
Where the evidence ends
These studies do not prove that equal turns cause better decisions. Privacy, unnamed display and release timing are distinct; none guarantees psychological safety or freedom from shared assumptions.
02Prepare / ProposalsGive a possible failure a voice.
Published finding
Prospective-hindsight experiments found that treating an outcome as certain changed the explanations people generated. A later descriptive study used written pre-mortems to elicit implementation barriers.
Our design response
When a concrete plan exists, ask people independently how it might fail. Turn useful concerns into evidence needs, safeguards and conditions for proceeding.
Where the evidence ends
A pre-mortem is a way to examine a plan, not proof of fewer failures. The earlier research did not validate the full modern method; the later study had no comparison group.
Mitchell, Russo & Pennington, 1989 ↗Gilmartin et al., 2019 ↗
04ProposalsExpand before narrowing.
Published finding
An ideation experiment found that people working alone and then together generated more and better ideas, and identified stronger ideas, than groups working together throughout.
Our design response
Protect the first independent round. Then use optional AI expansion, brainstorming, commitment rounds or a focused meeting to develop and examine proposals. Not every question needs every tool.
Where the evidence ends
This is task-specific evidence. The precise timing of AI support, the quality of alternatives and the effort the process adds still need evaluation.
Girotra, Terwiesch & Ulrich, 2010 ↗Fan et al., 2025 — AI timing experiment ↗
05Decide → AcceptKeep three human acts distinct.
Design hypothesis
We propose a testable distinction between expressing a view, authorizing a choice and accepting responsibility. It is not a scientific scale.
Our design response
A position on a proposal, an authorized decision and an owner’s acceptance of defined work are different acts. Keep “yes, if,” reservations and missing evidence visible. One question can advance while another remains open.
Where the evidence ends
These are design and governance choices. Sikizana’s exact response options are not a validated commitment scale. Acceptance does not itself establish capacity, correctness or technical approval.
What to test: Test whether people understand scope and dependencies more accurately than with ordinary action lists—and whether the benefit justifies the effort.
06Follow throughMake action specific. Check progress.
Published finding
Meta-analyses support if–then action planning and progress monitoring as ways to improve goal pursuit across the settings studied.
Our design response
Name the next action, relevant prerequisites and what will count as completion. Keep progress and proportionate evidence with the accepted work.
Where the evidence ends
An if–then action cue is different from conditional acceptance. Much of this evidence concerns individual self-regulation; it does not establish this enterprise workflow’s productivity or ROI.
07Review & learnLet changed evidence change the work.
Published finding
Research syntheses associate team processes and shared understanding with adaptation. After-action reviews can improve learning and performance in studied training contexts.
Our design response
Compare intended and observed results. When a recorded condition changes, review the affected commitment while retaining its earlier rationale. Pausing or stopping can be responsible follow-through.
Where the evidence ends
This does not validate Sikizana’s dependency links or review rules. Satisfaction and decision counts are useful signals, but neither is decision quality by itself.
This is a map of the method and its research basis, including practices being developed for the expanded book. It is not a claim that every practice is a dedicated software feature. Explore the current illustrative workflow →
The questions are part of the product
Measure what changes.
Including what gets worse.
The business hypothesis is fewer missed constraints, less avoidable rework and more reliable delivery.
Saved time becomes financial value only when it reduces expense, enables valuable additional work or improves delivery. A favorable rating alone cannot establish that.
- Information
- Useful knowledge surfaced; important constraints retained; unsupported claims detected.
- Judgment
- Breadth and quality of options; assumptions challenged; meaning preserved through synthesis.
- Execution
- Scope understood; prerequisites met; rework, elapsed time and total effort tracked.
- Business value
- Expense avoided, valuable additional work enabled, or delivery improved—not minutes saved alone.
A research agenda, not a result.
Compare unassisted preparation, AI available from the start, and an independent first contribution followed by optional AI. Examine conditional responses and explicit acceptance against ordinary action lists. Track quality, diversity, effort and downstream outcomes separately.
Small pilots can establish usability and feasibility. Stronger causal claims require a suitable comparison, sufficient data and a study designed for the question.
Consent before research.
No automatic reuse of private reflections. Agree the purpose, data access and protections; use appropriate ethics review. De-identification reduces risk, but cannot promise that identity is never inferable.
Specify, record, disclose.
Pre-specify outcomes and analysis where appropriate. Record protocol, prompt and model versions. Commit to reporting null and adverse findings, as well as improvements.
Explore a research collaborationSeeking collaborators in decision science, organizational psychology and human–AI interaction. Study arrangements and any institutional partnerships remain to be agreed.
The Third Field · Edition 2 · In development
The deeper argument.
The worked examples.
The expanded hardcover is being developed to examine the evidence, tensions and design choices in greater depth—from private reflection and pre-mortems to proposals, commitments and follow-through.
Planned cases span life-science evaluations, manufacturing supplier qualification and customer onboarding. They will illustrate the method without presenting fictional outcomes as proof.
About the expanded second edition →Explore cases from the record →