The playbook at a glance
From the problem to ten field-tested patterns.
AI-first systems often optimize for short-term automation while draining the professional expertise, trust and differentiation that took years to build. The Playbook starts from that problem and a shared vision: interfaces that grow expertise and create durable advantage over AI-first approaches. It’s the work of the HI Manifesto consortium, roughly 75 researchers and senior practitioners spanning AI, HCI, cognitive science, psychology, learning sciences, creativity and organizational science, who translated a broad range of cutting-edge research into one scalable method: identify and synthesize, design and prototype, test in real work, learn and refine, then scale. Within six weeks that method, run together with 22 practitioner organizations, produced the ten-pattern design language below and the thirteen working prototypes that put it into practice.
The ten Hybrid Intelligence design patterns, listed along the bottom of the graphic, are what this process produced. Click to open full size.
Who’s involved
Explore the ten patterns
Pick a pattern to see what it asks of an interface.
Contextual Prediction-Judgment Complementarity
“AI predicts. Humans judge.”
Design human-AI systems by systematically identifying prediction opportunities, preserving essential human judgments, and refining the allocation with domain experts in context.
AI systems increasingly absorb both prediction and judgment tasks, often without considering how expertise, responsibility, information and context are distributed across a workflow. The result can be deskilling, opaque decisions and misplaced accountability.
There is no fixed division of labor. Workflows are decomposed into prediction, judgment, coordination and execution, and each part goes to the actor best placed to contribute, weighing information, expertise, accountability, reversibility, risk and time pressure.
The ED interface only advances once the clinician has recorded their own reasoning and made a judgment, such as ordering an ECG or escalating triage.
Descriptions follow the HI Design Playbook pattern summary; guiding phrases and the five themes come from the manuscript (Table 2). The ten patterns are version 1, with a reformulation agenda for version 2 already underway.
From workflow to HI interface
How a team goes from ten patterns to a working prototype in a day.
The patterns are not applied to a blank screen. Teams follow a structured prompt-library pipeline, a variant of a workshop format run roughly fifty times before. It is Pattern 1 made operational: keep the workflow’s steps, separate prediction from judgment inside each one, then recombine them into a single human-AI workflow.
→
1
Choose a workflow
Pick an extended workflow, not an isolated task, and give an LLM thread the organizational context it runs in.
Output · Context-loaded thread
→
2
Break it down
Select and iterate the five to ten steps the work actually consists of, keeping each step intact, under explicit FERC discipline.
Output · 5 to 10 step workflow
3
Separate prediction and judgment
A fixed prompt produces a per-step decomposition table: what the AI can predict, what a human must judge, iterated critically against domain expertise.
Output · Prediction-judgment table
→
4
Apply the patterns
Turn the table into candidate interface “buttons” in three cost classes and co-engineered human-AI exchanges, guided by the ten Playbook patterns.
Output · Interface concepts
→
5
Prototype
Vibe-code a multi-step workflow demo, then narrow to the smallest implementable slice with a working front end and live LLM access.
Output · Working prototype
6
Test
Teams must point to the specific interface elements that build domain-expert learning, in a two-minute live demo, before fidelity is rated.
Output · Demo and fidelity rating
The demo criterion
Business relevance is not enough. Every prototype has to show where, in the interface, the domain expert learns.
Tested on 8 June 2026 with 38 external participants from 22 organizations. Thirteen working prototypes came out of the day.
Why we built it
AI alone beat the doctor with AI.
In a 2025 study, a diagnostic language model (AMIE) reached 59.1% top-10 accuracy on 302 real emergency cases, against 33.6% for unassisted physicians. Physicians assisted by the model scored only 51.7%, worse than the model working alone. Together with a lineage of similar findings, this has become a standard exhibit in the case for full automation: if the human in the loop makes the output worse, remove the human.
Look at what was actually tested, though. The “assistance” was a one-shot consultation, deliberately without onboarding or training. The study didn’t pit AI against hybrid intelligence. It pit AI against AI with an unscaffolded human attached at the end. For each failure mode the authors named, a decade of research already has a tested counter. None of it was in the interface.
Differential diagnosis performance
Top-10 accuracy on 302 emergency cases · higher is better
AI only
59.1%
AI + doctors
51.7%
Doctors only
33.6%
01
Anchoring
A fluent answer appeared before the clinician recorded a view of their own: textbook anchoring, with no judge-before-you-see sequencing to counter it.
02
Miscalibrated trust
Uncertainty, assumptions and limits stayed hidden. Seamful design, surfacing what the system does and doesn’t know, was never in the interface.
03
No skill-building
The study explicitly omitted training and gave clinicians no structured feedback on their own reasoning. Nothing in the loop built expertise.
Build systems where the human learns more from the interaction than the system does: perpetuating, not eroding, the need for human expertise.
Two flagships, two jobs
The DeepMind result, answered. And the route to automation, mapped.
The AMIE study is read as a verdict: the model beats the doctor with the model, who beats the doctor alone, so take the human out. Our first flagship tests that claim on the same task, emergency differential diagnosis, with the interface redesigned around the failures the study itself named.
The techno-optimist conclusion
Pure LLM > Doctor + LLM > Doctors
The human in the loop makes the output worse, so the human should be removed. Technology is ready to replace domain experts.
59.1%
51.7%
33.6%
AMIE study, Google DeepMind (2025), 302 emergency cases
VS
Flagship 01 · Systematic ED interface
The problem was the interface, not the doctor.
Judgment before prediction. The clinician records their own reasoning before the AI speaks, which defeats anchoring instead of inducing it.
Visible seams. Known and hypothesized diagnoses are kept apart, so the doctor knows what the system knows and what it only suspects.
A learning record. Each case ends by comparing AI and clinician judgments, building a record of when the doctor was right and when the machine was.
Flagship 02 · Standardized zone
Where automation is the goal, HI is the fastest responsible route to it.
JUAL’s frame-line quality checks are stable and repeatable, so the aim is to automate them. Starting as HI gets there sooner: an earlier supervised launch, corrections that carry their context, and fewer errors generated at source, while reviewers train the agent and move on to exceptions and improvement.
Read together, the two are one argument. The capacity that deliberate automation frees is precisely what funds the judgment work that deliberate interface design preserves.
The organizational complexity landscape
Two jobs for hybrid intelligence.
Management research maps work by two questions: how much certainty exists about how a task is done, and how much agreement exists about what should be done. That gives four zones, from simple to chaotic. Standardized, repeatable tasks sit where both are high. Judgment-dependent work, where cause and effect become clear only in hindsight, sits in the complex zone.
The boundary moves. Whenever tacit knowledge is formalized and made checkable, yesterday’s judgment becomes tomorrow’s rule, and work migrates from the complex zone toward the standardized one.
In the complex zone, the job is interfaces that make human judgment compound, because that zone does not empty out and it is where professional value lives. In the standardized zone, the job is to migrate deliberately, with domain experts holding the pen, so the capacity released is reinvested in judgment work.
Most AI strategy failures apply one zone’s logic to the other: agentic autonomy pushed into judgment work it cannot verify, or human review rituals kept around checks a machine now does better.
Figure 1 from the manuscript, after Stacey (1996) and Snowden & Boone (2007). Click to enlarge.
Simple
Stable, repeatable problems
High certainty, high agreement.
AI: Excellent today
Human role: Implementation and validation
e.g. Invoice processing, inventory restocking
Complicated
Analytically tractable with expertise
Moderate certainty, goals mostly shared.
AI: Strong
Human role: Expert oversight and exceptions
e.g. Forecasting, fraud detection
Complex
Patterns emerge through interaction
Lower certainty, outcomes depend on context.
AI: Support and mediation
Human role: Permanent judgment and interpretation
e.g. Market entry, organizational restructuring
Chaotic
Low certainty and low agreement
Contested, fast-shifting conditions.
AI: Framing and diagnostic support only
Human role: Leadership, ethics and crisis judgment
e.g. Crisis leadership, contested policy decisions
Flagship 02 · JUAL quality agent
Flagship 01 · ED diagnosis interface
The human role shifts from transitional implementation support to permanent judgment and relational intelligence.
Flagship 01 · Complex zone
An emergency-department interface that makes the clinician reason before the AI speaks.
Built by Systematic, a medical-software company
Blue is an AI suggestion, green is a doctor’s decision, and the workflow won’t advance until the clinician records their own reasoning first: judgment before prediction, which defeats the anchoring the AMIE study fell into. A closing Outcomes & Learnings view compares AI and clinician judgments in hindsight and builds a longitudinal “when was I right versus the AI” record.
Flagship 02 · Standardized zone
A manufacturing quality-assurance agent, trained by the employees it will relieve.
JUAL, a Danish manufacturing SME
Reviewers see every finding beside the original documents, confirm or reject it, and every decision is stored and retrieved for later cases. The team describes it as an “inexperienced colleague” they are actively training. The combined workflow ships earlier than full automation could, and reviewing the agent’s findings is already reducing errors at the source.
Earlier supervised launch
Corrections carry context
Fewer errors at source
Visualizations
ED diagnosis: techno-optimism vs Hybrid Intelligence
Systematic HI pilot · click to enlarge
JUAL: why start as Hybrid Intelligence
Frame-line order QA · click to enlarge
See it in action
Prototype walkthroughs.
Thirteen practitioner teams built working interfaces in a single five-hour sprint, and most were extended into narrated demonstration videos. Each walkthrough opens in a new tab. More are added as teams are ready to share.
Hybrid Intelligence in practice
Systematic
▶
Watch video →
Systematic · emergency department
An ED prototype that supports clinicians while keeping human reasoning and judgment central.
Xboarding.ai
▶
Watch video →
From offboarding knowledge to smarter onboarding
Xboarding.ai · knowledge transfer
Captures departing employees’ knowledge and turns it into structured onboarding through reports, AI assistance and reflective quizzes.
Knowit
▶
Watch video →
Co-create complex offers with AI
Knowit · professional services
AI analyses requirements and ranks options while people calibrate the reasoning and make the final decisions.
Involved
▶
Watch video →
Turn team signals into human-led action
Involved · leadership
A five-stage workflow helps leaders interpret employee signals, prioritize risks and act, with human judgment at every critical step.
Innovation Living
▶
Watch video →
Innovation Living · retail
An AI guide matches customers with the right sleeper sofa while staff stay in charge of reviewing, tuning and improving its advice.
+
JUAL quality-assurance agent
Walkthrough coming soon
+
Add a prototype demo
Space for the next walkthrough
Built at the Center for Hybrid Intelligence
Center for Hybrid Intelligence
▶
Watch video →
Center for Hybrid Intelligence
Helps users explore career paths, challenge assumptions and refine options without taking over the decision.
Center for Hybrid Intelligence
▶
Watch video →
Center for Hybrid Intelligence
A co-pilot for comparing quantum strategies, exploring sources and viewing policy problems from multiple perspectives.
+
Add an internal prototype
Space for the next walkthrough
All videos are hosted on hi-infinity.ai.
The research behind it
Grounded in a decade of human-centered AI research, and the HI Manifesto series.
This paper discharges the interface level’s burden of proof within the HI Manifesto series: that complementarity is buildable, that the patterns can be correctly enacted, and that fidelity can be measured rather than assumed.
The manuscript · early draft
The HI Design Playbook: Operationalizing a Decade of Human-Centered AI into Interfaces Where the Human Learns More Than the System
86 contributors across 15 institutions, led by Jacob Sherson, Center for Hybrid Intelligence.