Skip to main content
A process is the durable unit of the operating model. It is a piece of recurring work that a unit owns, such as overdue receivables collection, quote to order, or new-hire onboarding. Opening a process shows you how that work runs today and where it is heading.

​
The step sequence

The heart of a process is its step sequence, laid out like a tube map: a handful of coarse steps in order, each shown as a row. Processes are deliberately kept to roughly 5 to 15 steps. If a process needs 40 steps to describe, it is really several processes. Each step row shows:
  • The step name, a short verb phrase such as “Prioritise accounts to chase.”
  • The current disposition as a coloured chip: Human, AI-assisted, Automated, or AI-autonomous.
  • The target disposition, where one is set. Steps with a gap between current and target are highlighted, because they are the redesign opportunities.
  • The platform that does the work, where relevant, and the owner accountable for the step.
A process can be linear (a sequence with a start and end) or cyclical (a loop that repeats, like a monthly close). The shape is shown on the process so the sequence reads correctly.

​
Understanding a step

Open a step to see the detail behind it. Alongside its name and description, a step can carry:
When a step’s target is AI-autonomous, it carries an escalation trigger: the condition under which the agent hands back to a person. Autonomy always has a defined human fallback.

​
What goes in, what comes out

A step is easiest to reason about as a black box: something enters it, work happens, something leaves. Recording the input and the output in plain language, “approved PO plus supplier invoice PDF” in and “coded invoice posted to the ERP” out, does three things a step description on its own cannot:
  • It gives quality a definition. “Is the AI doing this step well?” has no answer until you have said what the step is supposed to produce.
  • It makes handoffs checkable. One step’s output should be the next step’s input. Where the two do not line up, you have found a coordination problem hiding between steps.
  • It gives an AI implementation its brief. Whatever ends up running the step, an accelerator, an agent, or a person following a prompt, works to the output you declared.
Acceptance criteria is the definition of done for that output. Write it as separate checkable lines, “correct GL code; correct cost centre; VAT treatment right; PO matched within tolerance”, rather than a general statement like “the data looks right”. Explicit lines are what a reviewer can mark against, and what an agent can be graded against.
All three fields are optional, and most maps fill them in over time. A step without them is still a valid step. But complete them on any step you intend to hand to AI, because the acceptance criteria you write become the brief for whatever implements it.

​
Why constraint and coordination matter

The operating model ranks your redesign backlog by constraint and coordination burden, not by whatever looks quickest. A step that is both a bottleneck and heavy on coordination is where AI creates the most value, because easing coordination work, the “work behind the work” of scheduling, chasing, and reconciling, is often what moves the numbers. Flagging these on a step is how the highest-value opportunities rise to the top.

​
What implements a step

Once a step is more than Human, you can see what implements it. A step can link to one or more:
  • Accelerators from your accelerator library.
  • Hub agents that run the step autonomously.
  • External tools, recorded with a name and link.
Each implementation has a status of planned, live, or retired. This is also how the map stays honest: retire an agent, and the step visibly loses its implementation, which is a prompt to revisit its disposition.

​
Quality checks on a step

An escalation trigger says when a step should stop and hand back to a person. A quality check says whether the output was right in the first place. Those are different questions, and a step heading for autonomy needs an answer to both. This matters more than it first appears. A step done by a person carries its own quality loop invisibly, because the person doing the work notices when something looks wrong. Move that step to AI and the loop disappears silently unless you replace it. A quality check is that replacement, written down and owned. Each check on a step records:
Keep it to one or two active checks per step. A quality check is a contract, not a dashboard. If a step targets AI-autonomous and has no live check, the step says so, because an escalation trigger on its own does not tell you whether the work is any good.

​
How a step earns more autonomy

The choice between gate and monitor, and the sample rate underneath it, is the mechanism by which a step earns autonomy.
1

Start as a gate

A step new to AI is checked every time. A person approves each output before it moves on. This is the human in the loop, made explicit and visible on the map rather than assumed.
2

Move to a monitor at a high sample rate

Once the outputs hold up, switch to sampling. You are no longer inspecting everything, but you are still looking often enough to catch a drift early.
3

Ratchet the sample rate down

As the score stays at target, review less. Each reduction is a decision backed by evidence you can point to.
4

Move the disposition

A sustained score at a low sample rate is the case for moving the step from AI-assisted to AI-autonomous. The score changes a real decision, rather than sitting on a chart.
The sample rate should never reach zero. Sampled review is quality assurance, and it is also how your organization keeps the judgement that an escalation trigger hands work back to. If nobody has looked at a step’s output for a year, there is nobody left who can tell a good one from a bad one when the agent finally escalates.

​
Measuring value with KPIs

Coverage shows where AI is. KPIs show whether it is working. A process can carry a small number of outcome KPIs, the numbers a leader actually recognises, such as days sales outstanding, cycle time, or recovery rate. Each KPI tracks three values:
  • Baseline — where you started, signed off up front.
  • Current — where it is now.
  • Target — where you are aiming.
KPIs know their direction (whether up or down is good), so progress reads correctly at a glance. Keeping KPIs to a handful per process is deliberate. This is a scoreboard that proves a redesign worked, not a business-intelligence tool. Your existing analytics stack still does the deep analysis.
Quality checks and KPIs answer different questions. A quality check is a leading, in-process measure on a step: was this output right? A KPI is a trailing outcome measure on the process: did days sales outstanding actually come down? You want both. A redesign where the quality checks pass but the KPI does not move is worth questioning.

​
How processes connect

Real work crosses process boundaries. A process can have connections to others:
  • A handoff, where one process passes work to another.
  • A feed, where one process supplies information another depends on.
Connections can be pinned to the specific steps where the handoff happens, and the map highlights when a connection crosses a unit boundary, which is often where coordination cost and delay hide.

​
Linked initiatives

A process shows the delivery work attached to it: the Opportunities, Deliverables, and projects that are changing it. This is what makes an initiative legible as a change to a specific process, with an address in the business and a value hypothesis, rather than a free-floating project.

​
What’s next

Owning and improving a process

Edit steps, set targets, and suggest changes

From map to delivery

Turn a redesign gap into delivered work