AI by Design — A Manifesto in 5 Acts

Where it comes from

This framework grew out of a question I kept encountering while teaching Designing Organizational Culture at Stanford d.school: why do high-performing teams tend to fall apart when AI gets introduced? The tools get better, the output speeds up, but something in the team dynamics quietly breaks.

The answer, I think, is that teams treat AI as a tool upgrade when it actually functions more like a new team member — one with distinct capabilities, clear limitations, and a presence that changes how everyone else relates to each other. Richard Hackman's research on real teams, Amy Edmondson's work on psychological safety, and Google's Project Aristotle all point to the same thing: effective teams depend on purpose clarity, boundary clarity, and trust. Add AI without re-establishing all three and you get drift — ambiguous ownership, eroding standards, quiet loss of accountability.

I've tested this framework across d.school workshops, a two-part virtual series co-facilitated with researchers from Meta, DeepMind, and Google, and a career expo experiment at Stanford co-designed with Monchu Chen. Each iteration sharpened what the framework actually needs to do.

The framework

Three layers, in sequence. You can't do layer two without layer one.

Re-establish purpose

What is this team trying to accomplish, and does AI serving as a team member advance that or complicate it? This sounds obvious but teams skip it constantly. The question isn't "how do we use AI" — it's "does AI move us toward what we're actually here to do?" If the answer is unclear, start here before anything else.

Define role clarity

Who leads which decisions? What can AI initiate without asking? What must come back to a human? Where are the non-negotiable checkpoints? Role clarity in a human-AI team isn't just about efficiency — it's about accountability. If nobody knows whose call it is, nobody owns the outcome.

In workshops, teams define an AI teammate persona: its voice, its boundaries, what it takes initiative on, and what it defers. This isn't anthropomorphism — it's a design decision about how the team structures its collaboration. A team that has named these things is harder to surprise by an unexpected AI output.

Choreograph at the task level

This is where Task Split Design lives — the micro-level choreography of who leads, who critiques, and when the handoff happens within a specific piece of work. The moves: map the task into 3–5 micro-steps before involving AI, assign each step a lead — human or AI — and a critic, insert at least one explicit human-decides checkpoint, and reflect afterward on whether the split was right.

Keep human judgment active throughout. Saving it for a final sign-off, over time, turns oversight into atrophy — and atrophy into cognitive surrender.

In the workshop

The workshop opens with a paper clip exercise — not because paper clips are interesting but because divergent thinking needs a warmup. The deeper point, which becomes clear in the debrief: your brain just did something AI finds genuinely hard. It invented categories. It broke rules deliberately. That's the capability the team is about to design around.

The core activity is building an AI teammate canvas: participants define their teammate's role, name, voice, constraints, and the tasks it handles — then swap canvases with a partner. The partner's job is to find the one task that's too vague — the one where a real teammate would say "I'm not sure what you mean." They hand it back with one question, and the original designer has two minutes to sharpen it.

That exchange — build alone, test on each other — is the heart of the workshop. It surfaces quality issues without needing a formal evaluation structure, and it makes the team's implicit assumptions visible before they become collaboration breakdowns.

In the career expo experiment with Monchu Chen, we ran a hidden-condition version: different teams worked with AI teammates configured with different personalities — Baseline, Craftsperson, Surrealist — without knowing which they had. The share-out produced wildly different team stories despite an identical brief. The debrief question — "what did you notice about your teammate's style?" — surfaced more insight about human-AI collaboration than any structured analysis could have.

What participants consistently discoverThe boundary-setting step is harder than anyone expects. Teams can easily name what they want AI to do. They struggle to name what AI should never do — the decisions where human presence is the point, not a checkpoint. Getting clear on that is the most valuable thing the workshop produces.

The artifact

A team charter that names AI as a member alongside humans. It covers: purpose (what the team is here to do), AI's role and scope, explicit boundaries (what AI doesn't decide), human checkpoints, and a quarterly review commitment. The charter is a working document, not a ceremonial one — it should be revisited every time the team's AI configuration changes, which will be often.

The individual AI teammate canvas lives inside the charter — each person's design for how their specific AI collaboration will work, reviewed and agreed on by the team.

Grounding

Hackman (2002) on real teams: effective teams need clear purpose, real boundaries, and enabling conditions. Edmondson (2018) on psychological safety: people need to feel safe enough to flag AI errors, raise concerns about AI outputs, and say "I don't think our AI teammate should have the final say here" without social cost. Google's Project Aristotle: psychological safety predicts team performance more than any other factor — and that finding applies when AI is on the team too.

When AI joins a team, the team needs to re-form.

Sections

Submit

Colophon