The why has nowhere to live

I have spent more than twenty-five years watching people capture and draw their processes, and the conversation has barely changed. Someone maps how the work is supposed to go. Six months later nobody can say who agreed to it, why one branch exists, or whether the picture still matches the work. The tools got way better. The conversation, however, stayed the same. Twenty-five years is a long time to sit in the same meeting.

I co-invented Universal Process Notation in the early 2000s at Nimbus. Since then it has been used in pharmaceuticals, financial services and food production, in some of the largest companies in those sectors, to document the kind of work an auditor asks about. Long enough to see which rules survive contact with a real organisation, and which ones quietly get ignored.

Right now a new generation of tools is clearly arriving, and they are genuinely good. Also, they are all making claims about process knowledge. Scribe has been at a $1.3bn valuation since last November. Guidde raised $50m in February. Skan $63m in August. June with $20m of pre-seed money and Marc Benioff’s fund behind it. This is not a fad, and most of it is better than what came before.

But there is a question none of them has answered yet, and it decides whether any of this captured knowledge is still worth anything next year.

What the new tools genuinely solved

Process mining solved scale. Celonis says it plainly: their output shows “how your processes actually run – not how you think they do”. Anyone who has run a discovery workshop knows exactly what that sentence points at. The room describes the process as it is perceived to work. The event log describes it as it really does. Most improvement money is buried in that gap, and until mining arrived we found it by accident.

Desktop capture solved invisibility. Skan, Soroco and Mimica watch the work that never touches a system of record: the spreadsheet, the second screen, the copy and paste between two applications that nobody will admit to. Soroco describes “capturing every variation, exception, workaround, and judgment call in workflows”, and a great deal of real work is exactly that.

The newest group solved access. ClearWork’s agent runs asynchronous interviews, and makes a true point worth repeating: “Async interviews reach the people who’d never make the workshop.” June’s agents schedule and interview stakeholders, then turn those conversations into structured requirements. The night shift, the regional team, the one person who understands the exception; they were never in the room, and now they do not have to be.

Each of these is a real advance on a real constraint. Take them seriously.

Three levels of claim

It helps to be clear about what is being offered, because three different things are currently sold under one heading.

Level one: a model draws a diagram. Give a language model your procedure document and it returns a flowchart. This works, it costs almost nothing, and it will keep getting better. What it produces is a plausible picture. Nobody asked whether it is right, nothing checks it, and nothing happens when the procedure changes. Useful as a first draft, utterly worthless as a record.

Level two: new platforms claim more, and deliver more. They watch the real work, or interview the people doing it, and some now claim to capture intent. Skan says its model “captures both what operators do, and … why”. ClearWork goes further and gates what comes back: “Nothing enters the record until a person confirms it”, with every line traceable: “Click any line. See who said it.” That is serious design, and it is ahead of most of the market.

Two things separate those claims. Skan derives the why by inference, from what the screen showed; their own co-founder and chief executive, Avinash Misra, says the hard problem is not screen observation but “the intent extraction”. ClearWork asks a human being. Inferring someone’s reason and asking for it are different acts, and only one of them survives a challenge.

Level three: where the real question sits. You ran the interviews. You asked the right person. She told you the thing that matters: claims over fifteen thousand go to the finance director rather than the department head, because of what happened in 2019, and the documented procedure has been wrong for four years.

So, where does that sentence go?

The answer needs somewhere to live

In most of these tools it becomes a swimlane flowchart, exported to Visio or draw.io. Swimlanes record sequence and handoffs, and they do it well. What they do not have is anywhere to put the reason. The arrow leaving the box has no field for why you take it. So the answer ends up in prose beside the picture, in a note, or in an interview recording nobody will ever open again.

You asked the expensive question, got a valuable answer, and then put it somewhere it cannot be checked.

This is why notation is not a matter of taste. UPN has four rules that look small and are not.

Every flowline carries a condition. The why is not an optional field somebody might fill in. A flowline without one is a defect, visible on sight. This rule exists because in twenty-five years I have never seen an optional rationale field filled in properly. Make it optional and it stays empty. Optional fields are where good intentions go to retire.

Activities are verb-noun. “Approve credit limit”, not “Credit stage”. A verb-noun name is something a named person can be held to. Stage names borrowed from a system tell you where the record sits, not what anyone does.

There are no gateways. No diamonds, no XOR labels. Decisions live in the flowlines leaving an activity, so the reasons sit on the branches, at the decision itself, rather than in a header field describing the whole process. A diamond lets you write “Approved?” and leave the criteria in somebody’s head; a flowline cannot, because the condition is the label. This is the rule people argue with most, and the one that puts intent where it belongs.

Parent and child diagrams must agree at their boundaries. The conditions on a parent activity must match the entry and exit of the diagram beneath it. A machine can check that without a human reading anything, which means a hierarchy cannot silently drift apart: the failure that eventually kills every large process repository.

None of these were designed at a whiteboard. Each one is there because something broke without it, in an organisation that had to answer to an auditor.

What structure buys you

You can audit the knowledge. “Show me every branch with no stated reason” is a query against a UPN model. Against a picture it is a reading exercise, and a missing why is invisible, because there was never a slot where its absence would show.

Software can act on it. The branches leaving a decision are its options. A branch with no stated condition gives an agent nothing to test against, so at the first edge case it fills the gap itself, confidently and without telling you. If you want an AI agent to follow your process rather than improvise one, it really needs a specification. And a drawing is not a specification, however neat its arrows.

It outlives the tool. A model in an open format can be rendered, queried and moved. The model outlives the platform. A diagram locked into the application that drew it does not.

So where has UPN been all these years?

Fair question, and the blunt version is fairer still: if the notation is that good, why does every analyst know BPMN while almost nobody has heard of UPN?

Not because BPMN is weak. BPMN is excellent at the job it was built for, which is exact and executable specification. It carries execution semantics and an interchange format, so engines run it. That creates a tooling market, a vendor ecosystem, certifications, university courses. UPN was built for a different job: getting people to agree on what actually happens and, more importantly, why. Vendors sell engines. Nobody has ever sold consensus.

Ownership mattered too. BPMN belongs to a standards body, the OMG, and is an ISO standard besides, so anyone could adopt it, teach it, put it on a CV, name it in a tender. UPN came out of a commercial nest (Nimbus Partners) and basically travelled with the product.

Then the part I find most telling. Both notations take work, but the work sits in different places. A BPMN model can be built by one analyst at a desk who knows the systems. A UPN model needs the people doing the work to say why a branch exists, and that cannot be done alone. Anything that needs agreement scales worse than anything that needs only expertise.

Underneath all of it, the industry optimises what it can instrument. Systems emit logs. People emit, well, not so much. Process mining exists because event logs exist, not because sequence is the most valuable thing to know. For twenty-five years the tooling grew where the data already was.

So UPN was trying to answer a question nobody could afford to ask at scale. Two things have changed that. AI makes the asking cheap for the first time. And here is the thing: agents create the first real demand for a specification that carries business context and has a human standing behind it.

The market is already moving this way

Elements.cloud ties agent deployment to a signed-off, versioned diagram, telling its users to “version the design documentation with a formal sign-off” when the agent is deployed. Pega promises that requests are fulfilled “in the same approved way, every time”. Camunda requires a deployed model, though not an approved one. UiPath, Celonis and ServiceNow all say their agents are grounded in process context, without anyone approving the process.

Meanwhile the platforms that do have proper sign-off kept it apart from their new AI. Signavio puts diagrams in front of “a subject matter or BPMN expert”, on the grounds that a well-defined approval process is “crucial for ensuring that a document is factually correct, does not contain typos and complies with the corporate design standards before you publish it”. Nintex makes the process owner and expert the default approvers and prompts for a typed reason when one declines. IBM Blueworks Live carries a configurable re-approval interval, typically 90, 180 or 365 days. All of them added AI diagram generation during 2025 and 2026, as a separate feature. The generator produces. The approval workflow approves something else. Nobody has wired the two together.

That is the gap, and it is worth knowing which end of it your supplier sits at.

Six questions worth asking

Short questions. The pauses before the answers are informative.

• Where does a reason live in your model, which field, on which object?
• Can you list every decision in my process that has no stated reason?
• Who approved this map, and is their name attached to it?
• What happens to that approval when the map changes?
• Can I take the model out in a form another tool can read, rather than just display?
• If I point an agent at this, what exactly does it read?

What Synfliq is built on

Synfliq produces UPN from whatever you already have: procedures, diagrams, exports, transcripts. A reason on every flowline, a role on every activity, parent and child diagrams that agree at their boundaries, provenance on every output, and packages you can take elsewhere.

That order is deliberate. The structure comes first, because expert review of a model that cannot hold a reason is review of the wrong thing. With the structure in place, the review and sign-off layer we are building next has something solid to attach to: a named expert confirming what was generated, and that confirmation surviving the next change.

Twenty-five years of this work taught me one thing above all. Organisations do not lose process knowledge because nobody wrote it down. They lose it because what they wrote down had no room for the reason, and no name against it.