Design

Techniques for designing a new one.

The canvas describes an operating model. These are the moves you make when you have to design one rather than document it. Each borrows from a discipline that already solved a version of the problem, and each names where that discipline gets it wrong.

TechniquesSix, each with its failure mode

Borrowed fromWork design · sense-making · cybernetics · the shop floor · the studio · foresight

The decisionMatch the executor to the kind of problem

Technique, and where it comes from

What it does, and where it goes wrong

Deconstruction

from work design

Break the job into tasks before deciding anything. Jesuthasan and Boudreau's point is that automation types are not interchangeable, so the unit of decision has to be the task rather than the role. Almost no job is uniformly automatable, and the average across a job hides both the parts that should have gone and the parts that never should.

In practice

Do it on a workflow the room has actually run. Deconstructing a job description produces a job description.

Placement

from sense-making

Put each task in a problem domain before choosing an executor. This is the single highest-value hour in the method, because it is where the Klarna mistake gets made or avoided. The table below is the whole technique.

In practice

Watch for everything being classified complicated. Complex feels like an admission of ignorance, so rooms avoid it, which is exactly how a complex problem ends up agent-led.

Loop boundaries

from cybernetics

An agent is a loop. Deciding what it owns is deciding where the loop closes and what escapes it. Ashby's Law gives the test: an agent can own a loop only where its variety matches the loop's. Residual variety has to land on a person deliberately, or it lands on one by accident.

In practice

Draw the boundary before choosing the tool. Boundaries drawn around a product's capabilities are marketing, not design.

Escalation design

from the shop floor

The andon cord worked because pulling it was correct behavior rather than a cost, and because a human could see the problem. Neither holds now. The design question is what machine-detectable condition surfaces the exception, and whether a person is present and empowered when it does.

In practice

Test it by trying to trigger it. An escalation path that has never fired is a hypothesis about an escalation path.

Human-centered design

from the design studio

Empathize, define, ideate, prototype, test, and the double diamond behind it. Genuinely useful for the Purpose and Coordination cells, where the failure is almost always that nobody asked the people doing the work.

In practice

Read the critiques before you run it. Natasha Iskander's argument in Harvard Business Review is that it is fundamentally conservative, because even in empathy mode the designer decides which parts of the user's experience count. Lee Vinsel's polemic goes further and is worth reading precisely because it overstates.

Scenarios

from strategic foresight

Fill the canvas twice, for two futures that differ on one variable you cannot control. Regulatory timing is the obvious one right now, since high-risk obligations have already moved once. The comparison shows which cells are robust and which are bets.

In practice

Two scenarios, not four. Four produces a matrix nobody uses.

The composition decision, in one table

Match the executor to the kind of problem

Work type across the top, following Snowden's sense-making domains. Executor down the side. Read it as a placement rule rather than a maturity model: there is no column where agent-led is always better, and no row that is always safer.

	Clearknown procedure	Complicatedexpert analysis	Complexdiscovered by probing	Chaoticact first, novel
Agent-led	Automate fullyMistake-proof it, sample the output, keep a trail. This is where the return actually is.	Agent drafts, expert signsThe analysis is machine work. The signature is not.	Do notThe agent optimizes a problem nobody has framed yet. This is the Klarna cell.	Do notNo trained distribution to act from.
Hybrid	WastefulA human in the loop here is cost without cover.	Best fitAgent handles volume, person handles the exception and owns the call.	Best fitProbe with agents, sense with people. Cheap experiments, human interpretation.	Support onlyAgents supply situational awareness, never the decision.
Human-led	WasteYou are paying people to be a slow machine.	BottleneckDefensible only where regulation names a person.	RightFraming, negotiation, and anything with a victim.	Only optionAct, sense, respond. Command, then re-establish order.

Placement that holdsWhere the documented failures happened

The hatched cells are where the documented failures happened. Not because the technology was inadequate, but because somebody put an agent where the problem was still being discovered.

Test one workflow's fit