Half-day workshops with 45 people used to need to accomplish everything in the room. They don't anymore - and that changes how you should design them. This post makes the case for a new division of labour: the room surfaces contested decisions, social accountability, and tacit knowledge that AI can't manufacture. AI handles the sequencing logic, dependency modelling, and delivery planning that used to consume the best hours of the day. But the handoff only works if you design the session - before anyone walks in - to produce artefacts AI can reason over, not just summarise.
Marcus Hahnheuser
··12 min read
The Room Doesn't Resolve It Anymore
"Run the creative tooling track independently - it has no platform dependencies and can deliver value while foundation work continues. It is one of the few streams that can genuinely run in parallel with everything else."
That rule saved three months of sequential waiting. One entire work stream that could be built and adopted completely independently while the main platform was still under construction, rather than getting held up waiting for platform readiness that was still 60 days away.
How did 45 people in a full-day workshop produce a rule that specific?
The honest answer is: they didn't.
The room produced something AI can't manufacture: contested priorities, social accountability, and the tacit knowledge that only surfaces when delivery leads have to defend what had to be built first to their peers in real time. But the sequencing logic, the dependency chains, the classification of which work could genuinely run in parallel versus which was silently blocked? That came two days later, from feeding structured workshop outputs into AI and iterating through scenarios that would have taken three to five follow-up meetings to even surface, let alone resolve.
That's the new division of labour. And it only works if you design the session strategically, understanding both human and AI strengths, to produce artefacts AI can reason over, not just summarise.
The proof point first
Last week I ran a full-day planning workshop for a large-scale enterprise technology programme: 30-45 people, 19 teams each building different parts of the system, a web of interdependencies that nobody had fully mapped, and a 90-day window to prove the foundations worked and that we were prioritising deliverables that actually pulled the levers toward our objectives.
By the end of the day we had product proof scenarios, a foundational delivery skeleton, commitment cards where each team documented what they'd deliver, who owned it, and what they were waiting on from other teams, plus a full list of internal and external blockers and dependencies. Standard workshop output.
Two days later, using the structured markdown I'd built from those outputs, I had something very different:
A three-tier dependency model classifying every programme item by its prerequisite chain, distinguishing what's genuinely foundational, what requires integration before it can proceed, and what needs live stable output before it can even be instrumented
Four critical dependency chains with hard sequencing constraints across the content, creative, campaign, and reporting workstreams
Five critical path protection rules governing all future scope addition and re-prioritisation decisions
Delivery leader, entrepreneur, and dad based in Brisbane. Writing about what I'm learning across digital delivery, AI, business acquisition, and trying to be present while building for the future.
A delivery plan with explicit entry criteria and a 90-day milestone checklist
A parallel track classification, identifying the handful of work streams that could genuinely run alongside foundation work without dependencies, rather than getting held up waiting for platform readiness that was still months away
A risk register with specific mitigations and named owners
None of that came from the room. The room gave me the raw material. AI gave me the delivery plan. But, and this is the part most people miss, that only worked because of decisions I made about session design before anyone walked through the door.
What the old model got wrong
The traditional workshop assumed the session had to accomplish everything. Dependency map? Build it in the room. Trade-off decisions? Negotiate them live with 45 people watching. Sequenced delivery plan? Synthesise 19 teams and 40+ interdependencies in real time while keeping the energy up and the catering warm.
That model is cognitively brutal and produces lowest-common-denominator outcomes. You spend the best hours of the day on synthesis work, the stuff where being in a room full of people actively works against you. Social pressure pushes toward premature consensus. Live negotiation flattens nuanced trade-offs. The loudest rooms produce the bluntest plans.
And here's the part that rarely gets said: a lot of what workshop facilitators have been doing for the last two decades is genuinely AI work. Classifying items by dependency type, identifying what blocks what, modelling sequencing trade-offs across multiple streams simultaneously - these are pattern-matching and logical reasoning tasks. AI does them better, faster, and without the cognitive load of managing a room at the same time.
The shift isn't that workshops matter less. It's the opposite. Workshops matter more now, because what happens in the room is the only thing AI genuinely can't replicate: the creative tension when a delivery lead has to justify what had to be built first to a room full of peers who know exactly what it will cost them. The social contract that forms when 45 people hear the same commitment made publicly, with a name attached. The tacit knowledge that only emerges when accountability is visible and collective.
That stuff is irreplaceable. But for decades it's been buried under synthesis work. AI frees you to go get more of it.
The artefact design decision happens before the room opens
Most facilitators stop reading here and start thinking "fine, I'll just clean up my notes afterwards and feed them into AI." That's not what I'm describing.
The reason I could build a three-tier dependency model two days after the workshop isn't that I had good notes. It's that the session was designed to produce a specific kind of output - attributable, structured, and queryable - before a single person sat down.
Two decisions made the biggest difference:
Merging the dependency and blocker activity into the 90-day commitment cards. In the original session design, these were separate: teams build their commitment cards, then in a later session map blockers and dependencies onto a separate board. In practice, that produces two disconnected artefacts, a commitment register and a blocker list, that someone has to reconcile later, usually imperfectly, usually losing the thread between which blocker gates which commitment.
By combining them into one activity, I forced teams to attach blockers and dependencies inline with the commitment they belonged to. "We can't begin this item until [team] completes [thing] by [date]" sits next to the item it gates, with a name and a date, in the same card. That's not a time-saving facilitation trick. It's a structural choice that turns what would have been sticky-note snapshots into a queryable dependency graph. The same data that would have sat in a separate "blockers parking lot" became the raw material for a four-chain critical dependency model.
Making every commitment attributable. Not "the infrastructure team will set up log forwarding" but "[Name] owns log forwarding configuration and testing by Day 60, which unblocks AI visibility reporting in the next phase." The difference looks minor in the room. It's enormous afterwards. AI can reason about ownership chains, accountability gaps, and sequencing risks. It cannot reason about anonymous collective intentions.
These choices changed what I asked people to do. Instead of "map your dependencies on a separate board," the prompt was "for each 90-day item, name any team or decision you're waiting on and by when." The output looked roughly the same to participants. What changed was the structure underneath it, and that structure is what made two days of AI iteration possible instead of three more workshops.
There's a real tension here worth naming. Tight artefact design captures what you anticipated: the commitments, the named blockers, the sequencing logic. It doesn't automatically capture what you didn't anticipate. The most valuable signal from the workshop I described, the offhand comment from a subject matter expert about a fragile upstream dependency, didn't come from a structured template. It came from an open conversation during a playback session. Good artefact design leaves room for that. The structured outputs give AI something to reason over. The unstructured moments give you what the structure would have missed.
What only the room can give you
There's a version of this argument that sounds like "workshops are expensive theatre - just send a structured template and let AI do the rest." That's wrong, and it's worth being precise about why.
The dependency model and the sequencing I built after the workshop are more rigorous than anything the room would have produced live. But they're built on foundations that only came from the room:
Contested decisions with witnesses. When a delivery lead argues that a particular work stream belongs in the first phase and gets pushed back by peers who know the dependency chain, that's not just a decision, it's a socially validated one. The whole programme is more likely to hold to it because the room saw it happen. Forty-five people were present when the commitment was made publicly, with a name attached. AI can model trade-offs. It cannot manufacture the organisational buy-in that comes from accountability being collective and visible.
Surfaced tacit knowledge. The most useful signal from one session wasn't on any commitment card, it was a comment from a subject matter expert that a particular upstream dependency was far more fragile than the plan assumed. That kind of signal doesn't exist in any document. It exists in the heads of practitioners who've been working on this for two years, and it only comes out when they're sitting across the table from the people whose plans depend on it.
The texture AI needs to make good recommendations. The messy negotiation over scope, the defended priorities, the explicit concession of "we'll accept the next phase for this, but only if [condition] is met" - that's the organisational context AI needs to make sequencing recommendations that will actually hold. Limiting a particular stream's first-phase scope to access and mapping only, not activation, isn't a logical conclusion you can derive from first principles. It's a judgement call that emerged from the room, that I could then apply consistently across dozens of items. Without the room, you don't have the judgement. You just have the logic.
What this means for how you run the session
If you're redesigning your workshop approach for this kind of AI-assisted delivery, three things matter most:
Design for AI-legibility before the session opens. Decide what structured artefacts you need to come out of the room - not "notes" or "outputs" but specifically: decision logs with owners and dates, dependency relationships captured as "need | from who | by when," commitment cards with explicit blockers inline. Then reverse-engineer your activities to produce those things naturally, without asking participants to do extra documentation work. The session design and the artefact design are the same decision.
Protect the room for what only humans do. Kill any activity that's primarily synthesis or structuring work. That includes live dependency mapping on whiteboards, real-time conflict resolution across streams in a large group, and consensus-building on sequencing logic. Those are AI tasks now, and they're better AI tasks than they were facilitation tasks. AI doesn't lose the thread across 19 teams, doesn't succumb to social pressure, and doesn't run out of cognitive bandwidth at 3pm. Use the room time for contested decisions, expert input, and social accountability. That's what the room is actually for.
Treat the structured working document as the primary deliverable. Not the slide deck. Not the photos of sticky notes. Not the summary email. The structured, attributed, machine-readable record of what was decided, by whom, and with what dependencies, formatted so AI can reason over it, not just read it. That document is what makes everything afterwards possible: the delivery plans, the scenario modelling, the risk analysis, the iteration through different sequencing options. If you're leaving a workshop with a photo roll and a summary document, you've captured the conversation. You haven't captured the raw material for what comes next.
The line that changed how I think about this
The room gave you the raw material. AI gave you the delivery plan.
That's the new model. Not "AI replaces workshops." Not "workshops are for people stuff, AI is for process stuff." It's a designed handoff, a deliberate decision about where the room's job ends and AI's job begins, and what the room has to produce for that handoff to actually work.
What happens after the handoff isn't passive either. The valuable part isn't "feed structured outputs in, get a delivery plan back." It's the iteration: stress-testing sequencing assumptions, running scenarios where a dependency slips, surfacing contradictions between commitments different teams made in the room. Less "AI builds the plan" and more having a tireless analyst who never loses the thread across 19 teams and remembers every commitment every owner made. The human still drives. What changes is you're no longer doing the pattern-matching and synthesis work alone, in your head, at 11pm after the workshop.
The full-day workshop still matters. It matters more than it did when we were asking it to do everything. But what it needs to do, and how you have to design it to produce that, has fundamentally changed.
Most sessions haven't caught up yet. The facilitation is still running like it's 2019, trying to synthesise 19 teams and resolve every dependency before the afternoon break. The only thing that's changed is the notes are neater.
Design the artefacts first. Run the session to capture what only humans in a room can give you. Hand off something AI can actually reason over.
Then do the work together.
A note on when this breaks. This approach assumes workshop outputs are clean enough to structure after the fact, and sometimes they aren't. If the session produces genuine ambiguity that wasn't resolved in the room, no amount of AI iteration will manufacture clarity that doesn't exist in the raw material. It also assumes the organisation will act on outputs produced outside the room. In some cultures, if it wasn't decided in front of everyone with a pen in hand, it doesn't count, and the two days of post-workshop iteration will be invisible to the people who need to hold the plan. And tight artefact design can miss things; if the structure you designed pre-session didn't anticipate an important category of signal, that gap won't fix itself afterwards. None of these are reasons not to try. They're reasons to go in clear-eyed about what you're relying on.
Marcus Hahnheuser
