The Flow in Workflow

The word is a compound, and the compound is honest. Workflow is work and flow, and the second half is doing most of the work. Flow is the older idea, borrowed from the movement of fluids: a continuous passage of something from one place to another. A workflow, read against its own construction, is the flowing of work. The work moves. Something carries it from where it began to where it is wanted.

obviously generated by AI

This reading seems too obvious to dwell on, and yet a good deal of confusion in how teams use boards comes from quietly abandoning it. The field that has thought hardest about flow, the lineage that runs from lean manufacturing through queueing theory to modern Kanban, defines flow in item-centric terms without hesitation. Then, when the same field defines workflow, it often switches to something else: a series of activities, a sequence of things people do. The compound comes apart. The flow half describes work moving; the workflow whole describes work being acted upon. These are not the same picture, and the gap between them is where the trouble lives.

This essay is about closing that gap by taking the word at its construction. If flow is the movement of work, then a workflow is the structure through which work moves, and the right way to model that structure is as a space of states the work passes through, not as a chain of activities performed on it. An earlier essay argued that a Kanban board models the state space of work items rather than the action space of the people working on them. This one asks the prior question: what is a workflow, such that a board could model it that way, and why is that the reading the work itself demands.

Flow is the movement of work

Start with how the flow tradition defines its central term, because the definitions are unusually consistent and unusually item-centric.

A word about sources is needed first, because the name Kanban sits over two distinct bodies of knowledge, and each has a foundational document that people, in conversation, call “the Kanban Guide.” One is Kanban University, the lineage descending from David Anderson, whose document is The Official Guide to the Kanban Method. The other is ProKanban.org, whose document is titled, plainly, The Kanban Guide. The two agree on a great deal and diverge on the very point this essay turns on, so it matters at every step which is meant. Throughout, I name the body rather than the document: Kanban University for the first, ProKanban for the second.

Begin with ProKanban, whose guide states it plainly: flow is the movement of potential value through a system.¹ The same definition, in almost the same words, opens Daniel Vacanti’s book-length treatment of the metrics that make flow measurable: flow is the movement and delivery of customer value through a process.² Vacanti authored both, so this is one voice rather than two. In both statements the subject of the sentence is the value, and the verb is movement. The system holds still while the work passes through it. Neither definition mentions an activity, a role, or a person. The unit is the thing that moves.

The metrics inherit this and make it sharp. The four flow metrics ProKanban treats as mandatory are work in progress (WIP), throughput, work item age, and cycle time. Every one of them is a property of work items, not of workers. Work in progress is the number of work items started but not finished. Throughput is the number of work items finished per unit of time. Work item age is the elapsed time since an item started. Cycle time is the elapsed time between an item starting and finishing.³ Read the list and the worker is absent throughout. The measurements range over items and over time, and they are silent about who did what.

This silence is not an oversight. It is the point. A flow metric tells you how the work is moving, and to do that it must attach to the work, not to the hands the work passes through. The moment you tried to define cycle time in terms of activities performed, you would have to decide which activities counted, in what order, by whom, and the clean elapsed-time measurement would dissolve into a question about labour. The metrics stay coherent precisely because they refuse to look at the action space. They watch the item.

So the flow tradition, at the level of its definitions and its measurements, is item-centric without qualification. Flow is work moving; the metrics watch the work move. This is the half of workflow that holds together.

The other reading

The trouble begins when the same field defines the whole word.

Kanban University defines a workflow in its glossary as a series of activities, often performed within a service, that results in products or services being delivered, represented on a board by a set of sequential columns showing the activities the work item passes through.⁴ The standards bodies say much the same from a different direction. The workflow-management vocabulary maintained in the lineage of the Workflow Management Coalition treats a workflow as the automation of a business process, during which tasks are passed from one participant to another for action according to procedural rules.⁵ Two independent authorities, one from the Kanban community and one from business-process modelling, land on the same reading: a workflow is a series of activities.

This reading is respectable. It is held by serious people and codified in formal standards, and it is not the product of carelessness. There is a coherent story behind it: work items pass through activities, so one can name the segments of the passage by the activities that occur in them, and call the whole an activity series. The columns-as-activities convention has decades of practice behind it and a great deal of working software built on top of it.

But notice what has happened to the compound. The same tradition that defines flow as the movement of work defines workflow as a series of activities, and in Kanban University’s guide both definitions sit inside one document, the switch unremarked. There is no need to make a great deal of it. The point is only that the activity reading is in quiet tension with the field’s own concept of flow, and once that tension is visible the question becomes which reading to keep. The rest of this essay argues that the item-and-state reading is the one that coheres, and that the activity reading imports an assumption the work cannot bear.

Why the factory could leave it implicit

The activity reading is not arbitrary. It is a memory of where flow came from, and understanding that origin shows both why the reading was once natural and why it stopped being adequate.

Flow as a managed idea begins on the factory floor. The Machine That Changed the World, the study that named lean production for a general audience, locates the decisive move in Ford’s assembly line: the innovation was to make the work item itself move, in a continuous passage past stationary workers, with the operations arranged to align with that movement.⁶ The consequence for visibility was profound, though the book does not frame it in these terms. Once the item moved along a line, the question that told you the state of production changed. It was no longer “which operation is being performed right now,” the question a workshop foreman would ask. It became “how far along the line is the item,” a question about the item’s position. The item’s place in the line had become the readable index of its condition.

Lean refined this without abandoning it. The book’s account of Ohno’s reorganisation at Toyota is telling: the first step was to group workers into teams and give each team a stretch of the line, their piece of it, and let them work out among themselves how to perform the operations that stretch required.⁷ Even when the organising unit was a team of people, the principle around which they organised was a segment of the item’s journey, not a fixed allocation of who does what. The people arranged themselves around a band of the flow.

Here is the part worth dwelling on, because it is the bridge from the factory to knowledge work, and it is an inference the lean literature does not draw for us. The assembly line could leave the notion of state almost entirely implicit. The work was low in variability: the thousandth car needed substantially the same operations as the first, in substantially the same order. Because the operations were uniform and uniformly ordered, the flow could be made physically and continuously real, a line that moved at a steady pace. And because the flow was a continuous physical thing, the state of any item was simply its position along that physical line, visible to the eye, requiring no abstraction. The line was the model. Nobody needed to write down a list of named states, because the states were a continuous physical dimension you could point at.

Knowledge work removes both of the conditions that made this possible. The work is high in variability of path: at the level of the operations themselves, the steps, their order, the loops and branches and returns, no two items move quite alike. And the work has no physical extension, so there is no moving line whose positions could serve as a continuous, self-evident index of state. Neither of the things that let the factory keep state implicit survives the move to knowledge work. You cannot read state off a physical position, because there is no physical position. You cannot assume a uniform operation sequence either. At a coarse enough grain an order can sometimes still be named, every item discovered, then built, then validated, but a band that holds fixed across every item has stopped describing an activity and begun describing a state. Push the description down to those operations and the uniform sequence dissolves; raise it until the sequence holds and the activities have quietly turned into states.

This is why the state concept has to be made explicit, and why it has to be discrete. The continuous physical dimension of the factory line is replaced by a small set of named states, ordinals along an abstract passage rather than positions along a physical one. The named states do for knowledge work what position along the line did for the factory: they make the condition of each item visible at a glance. But they do it abstractly, because there is nothing physical left to do it concretely.

And the named states do something the activity chain cannot. A state says where an item is and stays silent about how it got there or what will move it on, and that silence is the flexibility high-variability work requires. An activity chain offers no such silence: it commits, in its very structure, to a sequence of operations, the one thing such work cannot accept. This is the deeper reason behind the earlier essay’s claim that a board is not concerned with the action space, and it makes a stronger claim available. For high-variability work, the board must not be concerned with the action space, because to encode the activity sequence into the workflow is to import the factory’s low-variability assumption into a domain that cannot support it.

What the formal tradition builds on

If the activity reading is a memory of the factory, the item-and-state reading is what the formal tradition reaches for the moment it tries to be precise.

A small but telling piece of evidence: the word workflow does not appear in the index of either of Donald Reinertsen’s books on product development flow, the works widely regarded as the sharpest economic treatment of the subject. When the tradition gets formal, it drops the compound and talks instead about flow, queues, and items. Reinertsen’s apparatus is queueing-theoretic throughout. The unit of analysis is the work product and the queue it waits in; cost of delay, the central economic quantity, is a property of the item, the rate at which that item’s value decays as its delivery is postponed.⁸ None of this needs the word “workflow,” and none of it is about activities. It is about items moving through a system, and the economics of their waiting.

Vacanti’s metrics rest on the same foundation, and he is explicit about it. Before he defines a single metric, he sets up a queueing system: arrivals, items in the system being worked on or waiting, and departures.⁹ This is preparatory, and what it prepares is the pair of boundaries on which everything else depends. To measure flow you must fix the point at which an item enters the system and the point at which it leaves. Without a defined entry and exit, the flow metrics are named but not defined; they have no values, because there is no bounded region over which to compute them. Fixing the entry and exit is a state-space act. It demarcates the subspace of states that count as “in the system,” and the metrics then range over that subspace. The commitment point and the delivery point are not activities. They are the boundaries of a state space.

On this base sits Little’s Law, the result that underwrites the whole edifice. In the form Vacanti gives it, the law relates three quantities: L, the average number of items in the system; W, the average time an item spends in the system; and λ, the average number of items arriving per unit of time.¹⁰ Read the three definitions. Items in the system, time an item spends, items arriving. There is no term for an activity and no term for a worker. The relationship is written entirely in terms of items and time. It describes how these averages stand to one another in a stable system; it does not, by itself, make flow predictable. The formal heart of the flow tradition is a statement about a population of items in a bounded space, and nothing else.

So the convergence is broad. The definition of flow, the flow metrics, the queueing model beneath them, and the economic theory beneath that all describe items moving through states. They acknowledge a shared origin in the Toyota Production System and the work of Ohno and Deming, an origin ProKanban names in its own history.¹¹ The lean lineage carries the same item-primacy in its very notion of value. An activity counts as value-adding only if it transforms the work item into something closer to what the customer wants, so the value is located in the item’s changed state and the action is judged by what it does to that item. Even waste, in the lean sense, is defined from the item’s side, as effort and delay the item gains nothing from. The tradition most preoccupied with scrutinising activities still judges those activities against the item. Across the lineage, when precision is required, the activities recede and the items and states remain.

A definition

The reading the word invites, and the reading the tradition converges on, can be stated directly.

A workflow is a state space of work items, together with the legitimate transitions between those states, governed by explicit policies.

Each clause earns its place against the sources. The work items are the units that move, the individual units of value ProKanban places first among the elements of a defined workflow. The states are the conditions an item occupies between a defined start and a defined finish, which is what ProKanban’s own list of required elements describes when it asks for the states that work items flow through. The transitions are the moves an item makes from one state to the next, and the policies are the explicit conditions under which those moves are permitted, which ProKanban requires as policies about how work items can flow through each state. The definition is not an invention laid over the practice. It is the practice’s own description of itself, rendered in the vocabulary of state spaces, which is the vocabulary the practice was already using without naming.

The definition also recovers the metaphors the tradition teaches with. Kanban University’s introductory image is a motorway: vehicles, the work items, move through sections of road, and a vehicle in a section is in a place, a state, not performing an activity. When a vehicle is blocked, it stays where it is and is cleared in place. The metaphor is a state space of items in motion, whatever the glossary that accompanies it says. The definition simply makes explicit what the metaphor already shows.

And the definition draws the boundary with activity precisely. A workflow, so defined, says nothing about which actions move an item between states, in what order, or by whom; it holds each item’s position and leaves the sequence of operations open. That openness is what fits high-variability work, and what the activity-centric definition gives away.

What turns on the reading

This can sound like a quarrel about words, and it is worth being plain that it is not. The reading you hold changes how the board behaves, and an earlier essay traced the dysfunctions that follow from the activity reading. They are worth naming together, because they share a single root.

Read the columns as activities, and a developer who must touch an item during verification has nowhere to put it but back in development, and the team falls to arguing whose fault the backward movement was. Read them as states, and the item never moved: it sat in verification while the work of verification, which can include writing more code, went on. The backflow dissolves, because there was never any flow backward to explain. Read the columns as activities, and each becomes the place a kind of person works, until people speak of “my step”, identity fuses to a column, hand-offs turn adversarial, and every station optimises itself at the expense of the whole. Read them as states, and a column is a condition the item is in, not a territory a role owns; the question stops being whose column this is and becomes what the item needs to advance.

Each of these is the same mistake in different clothes: treating the board as a model of who does what, when it is a model of where the work is. And this is why the labels are not as innocent as they look. A board with activity-named columns can work, and many do, because what is working is a state space whatever the columns happen to be called. But a label is not inert. “Development” on a column quietly recommends the activity reading, and the activity reading is what brings the trouble. The label neither makes the board a state space nor stops it being one; it leans on the people who read it, toward one understanding or the other. That is reason enough to choose it with care.

What the definition assumes

A definition advanced this confidently should be turned on itself, because the same examination this essay applied to the activity reading applies here too. The state-space definition rests on assumptions, and naming them is the honest way to mark where it holds and where it does not.

The first assumption is that work items have states discrete and stable enough to be worth naming. This is an existence claim, and it is not free. Some work resists clean state demarcation: a single sprawling item whose condition is truly continuous, or whose meaningful states are too numerous and too short-lived to be useful, will strain the model. The definition assumes the work decomposes into states that persist long enough to be visible and few enough to be legible, and where that assumption fails the model gives less than it promises. ProKanban itself concedes the related point that there is no right number of states, which is an admission that the definition is a frame and not an algorithm: it tells you what kind of thing a workflow is, not how to carve one.

The second assumption concerns exclusivity, and here the definition is narrower than it may sound. To say a workflow is a state space is not to say that a state space is the only legitimate model of work. Activities are real. The operations people perform are real, and modelling them, specifying how a given operation should be carried out, is a worthwhile thing to do for its own purposes. The claim is not that the activity view is wrong as a model of activity. It is that the activity view is the wrong model of a workflow, because a workflow is the structure through which work flows, and flow is the movement of items, not the performance of operations. The two models can coexist. The error is only in asking one artefact, the board, to be both.

The third assumption is the load-bearing one, and it is frame-dependent in a way worth stating outright. The definition’s advantage over the activity reading rests on the claim that what the work needs is visibility without commitment to sequence. That claim is true for high-variability knowledge work, which is the domain this essay addresses, and it is the reason the state model beats the activity chain there. But it is not a universal truth about all work everywhere. A context whose value depends on a fixed, mandated sequence of operations, certain regulated or safety-critical procedures, for instance, may be a context where the activity chain is the right model and the state space is the impoverished one. In such a context the work is low in path variability by design, and the factory’s original assumption holds again. The state-space definition is the right definition for work whose path varies. Where the path is fixed and must stay fixed, the case made here does not reach.

None of these caveats undoes the definition. They locate it. It is the right account of a workflow for the high-variability work that flow methods were built to serve, it coheres with the field’s own concept of flow and the mathematics beneath its metrics, and it fails gracefully and namably at its edges. That is about as much as a definition of a contested term can fairly claim.

A closing note

The word told us what it was. Workflow is the flow of work, and flow, on the tradition’s own account, is the movement of work through a system. A workflow is the structure through which work moves: a space of states, with the transitions between them and the policies that govern those moves. The activities are what people do to carry an item from one state to the next, and they matter, but they are not the workflow. They are the labour the workflow is silent about, and that silence is what lets the work vary.

There is an old worry that a board is too linear for work that varies, that laying states left to right imposes an order the work refuses. The worry is sound, but it is aimed at the wrong object. It holds against an activity space, where the columns are a mandated sequence of operations and a fixed order misfits variable work. It does not hold against a state space. A state is a bucket, and the bucket is the state: a condition the work can be in, holding whatever actions belong to it without ordering them. The actions inside loop, branch, recur, and run in parallel as the work demands, and the states impose no sequence on them at all. They structure the work from above without lining it up. The variability the worry defends is real, and the state-space reading leaves it exactly where it lives, inside the buckets, untouched.

The factory could leave all of this implicit, because its work was uniform and its flow was a line you could see. Knowledge work cannot, because its work varies and its flow is invisible, and so it has to draw the states it can no longer point at. The board is that drawing. It is a state space made visible, which is to say it is a workflow, which is to say it is the flow of work, which is what the word said in the first place.

¹ The Kanban Guide (ProKanban.org, May 2025), which defines flow as the movement of potential value through a system. Offered under CC BY-SA.

² Daniel S. Vacanti, Actionable Agile Metrics for Predictability, p.13, defining flow as the movement and delivery of customer value through a process.

³ The Kanban Guide (May 2025), the four mandatory flow metrics, each defined strictly in terms of work items started and finished.

⁴ Kanban University, The Official Guide to the Kanban Method (v2, December 2022), glossary entry for “Workflow.”

⁵ ISO 12651–2:2014, the workflow-management vocabulary standard, which tracks the definition maintained by the Workflow Management Coalition: a workflow as the automation of a business process during which tasks are passed from participant to participant for action according to procedural rules.

⁶ James P. Womack, Daniel T. Jones and Daniel Roos, The Machine That Changed the World, ch.2, on Ford’s moving assembly line and the alignment of operations to the movement of the work item.

⁷ The Machine That Changed the World, p.55, on Ohno grouping workers into teams and assigning each team a stretch of the line.

⁸ Donald G. Reinertsen, The Principles of Product Development Flow, on queues as the central object and cost of delay as a property of the work item. The term “workflow” appears in neither this book’s index nor that of the earlier Developing Products in Half the Time.

⁹ Vacanti, Actionable Agile Metrics for Predictability, p.24, introducing the queueing system as arrivals, items in the system, and departures, preparatory to fixing the commitment and delivery points.

¹⁰ Vacanti, Actionable Agile Metrics for Predictability, p.43, stating Little’s Law with L as the average number of items in the system, W as the average time an item spends in the system, and λ as the average number of items arriving per unit time.

¹¹ The Kanban Guide (May 2025), History of Kanban, tracing the present state of Kanban to the Toyota Production System and the work of Taiichi Ohno and W. Edwards Deming.