Can you build a process model from meeting transcripts?
Yes, partially. Meeting transcripts are rich in who does what, where work changes hands, which exceptions occur and why a step exists, and poor at the order of steps, how long each takes, which systems are used and the routine steps nobody bothers to discuss, so a model built from transcripts is a well-evidenced draft that one document and one observed run of the job can finish.
Written for operations and transformation leads who have hours of recorded meetings and no current process documentation. Reviewed by Mike Bentley, founder, .
What a meeting transcript reliably tells you about a process
A meeting transcript is a record of people explaining work to each other, and people explain the parts of work that need explaining. That makes a transcript a strong source for six things a process model needs and a formal document rarely holds.
The first is actors. Speakers name themselves, their colleagues and their teams in the course of describing what happens, so a transcript usually tells you who performs a step more reliably than an organisation chart does. The second is hand-offs, because the moment work leaves one person for another is exactly the moment people complain about, ask about or negotiate. The third is decisions and their criteria: when someone says "if the amount is over the delegation we send it to the regional manager", you have a decision, its rule and its outcome in one sentence.
The remaining three are pain points, exceptions and vocabulary. People spend meeting time on what goes wrong and on the cases that do not fit, and they use the words the organisation really uses for its forms, queues and statuses rather than the words a template would impose. A model that carries that vocabulary is one the people doing the work will recognise as theirs.
What a meeting transcript leaves out
A meeting transcript is a poor source for the skeleton of a process: the sequence of steps, how long each one takes, which applications and screens are used, and how much volume passes through. Nobody in a meeting says "then I open the second tab and copy the reference number into the field at the top", because everyone in the room already knows, or nobody cares.
The order of steps suffers most. Conversation jumps between the start of a process and its end, returns to an exception, and then settles a point about approvals that belongs in the middle. The words "then" and "after that" in a transcript describe the order of the conversation at least as often as the order of the work.
System names are the second casualty. Speakers say "the system", "the portal" or "the spreadsheet", and an automatic transcript turns product names and acronyms into whatever sounds closest. Durations and volumes are the third: a transcript tells you a step feels slow or happens a lot, which is a reading of frustration rather than of time or count.
The quiet routine steps are the hardest gap to see, because they leave no trace at all. The step that always works is never discussed, and a model built only from discussion will have a hole exactly where the process is most reliable.
How to extract a process model from transcripts
The method that holds up for building a process model from transcripts is to read each transcript for claims rather than for a summary. A claim is one statement about how work happens, kept in the speaker's exact words, with the transcript, the speaker and the position in the recording recorded beside it. A summary loses the words, and the words are the evidence you will need when two people disagree about what was said.
As each claim is captured, sort it into one of three piles: how the work runs today, what someone intends or wishes it did, and what a vendor or consultant proposed. Meetings about improvement mix all three in the same breath, and a model that quietly promotes an intention or a sales pitch into a current-state step is wrong in a way nobody will spot until it is built on.
- Capture claims, one statement each, with the exact words and their source position kept as evidence.
- Label every claim as current state, intention or proposal, and build the current-state model from the first pile only.
- Assemble the actors, hand-offs, decisions and exceptions the claims support into steps and decision points, and attach the supporting claims to each.
- Where the claims say nothing about an order, a duration, a system or a step, mark the gap in the model explicitly and leave it open rather than filling it with what seems likely.
- Where two transcripts contradict each other, record both claims against the step instead of choosing one.
How to verify a process model built from talk
A model assembled from transcripts has been verified when the people quoted in it have seen it and the gaps it declared have been closed by a source that is not more talk. Three checks do most of the work.
First, show the draft to the people whose words it rests on, with their quotes attached to the steps. Reading your own sentence beside a box on a diagram is a strong prompt to say "that is not what I meant" or "that is only true for the overseas cases", and both corrections improve the model. Ask each reviewer one further question: which step is missing? People rarely notice a hole in a model they are reading, but they notice one when asked to look for it.
Second, add one document. A policy, a procedure, a form or a checklist pins down the things talk is weakest on: the official order, the mandatory fields, the approval thresholds, the names of the systems of record. Where the document and the transcripts disagree, you have found either an out-of-date document or an undocumented workaround, and either is worth knowing.
Third, watch the job done once. A single consented recording of one person completing the process end to end supplies the sequence, the screens, the applications and the quiet routine steps in one pass, at a cost interviews cannot match. The transcripts tell you why the process is shaped the way it is; the recording tells you what the shape is.
The failure modes when modelling from transcripts
Most ways a transcript-built process model goes wrong share a cause: the modeller treats the transcript as a description of the process rather than as evidence about it. Four failures do most of the damage.
All four are cheap to avoid while claims are being captured and expensive to undo once a diagram exists, because a drawn process acquires an authority its evidence never had.
- Treating a confident speaker as the process. The person who talks most about a process is often its manager, its critic or its newest member, and none of them does the work daily. A claim from the person who performs a step outweighs a fluent summary from the person who oversees it.
- Merging two variants into one. When the Sydney team and the Perth team describe the same process differently, the easy move is to draw an average that neither performs. Keep both as variants, attached to their quotes, until someone with authority decides which is the standard.
- Inventing order. A diagram demands a sequence and a transcript rarely supplies one, so the modeller fills it in from common sense. Common sense is an assumption, and an assumption drawn as an arrow looks like a fact. Mark unknown order as unknown.
- Losing the quote. Once a step is paraphrased and its source dropped, nobody can check it, and the model becomes one more opinion about the process. The citation is what makes a transcript-built model worth more than the meeting it came from.
Where transcripts fit beside event-log process mining
Process mining reconstructs a process from the event logs a transactional system writes as it works, and answers questions about order, duration and volume well, because those are what a log records. Process mining needs the work to leave events in instrumented systems that share a case identifier, and most processes an operations lead worries about do not: they live in people, email, documents and a spreadsheet between two systems. The Learn article on process mining without event logs sets out that boundary.
Transcripts, documents and observed work are the evidence for that second kind of process. Transcripts carry the reasons, the actors and the exceptions; a document carries the rules and the official order; one recording of the job carries the sequence and the systems. Together they cover most of what a log would show, plus the reasons behind the steps, which a log never does.
Doing this by hand is slow, which is where tooling earns a place. Omni by Valstra is one option: it reads meeting transcripts and documents as text and, when it builds a process from them, reports the gaps the sources leave rather than inventing detail, as its published constraints page sets out. Omni does not read audio or video, so a recording is transcribed first, and a person accepts every proposed step before it joins the model. A spreadsheet of claims and a careful analyst produce the same model more slowly; the method does not change.
Common questions
- Do we need to record every meeting?
- No. Record the working sessions where people walk through how the process runs, with the consent of everyone present. A few of those on one process are worth more than months of status meetings, which mention the work without describing it.
- Are AI transcripts accurate enough?
- Accurate enough for who does what, where work changes hands and why, which is what a transcript is good for anyway. Automatic transcription most often garbles product names, acronyms and numbers, so check those against a document or the speaker rather than trusting the text.
- What about privacy and consent for recording meetings?
- Tell participants a meeting is being recorded and why, get their consent, and follow your organisation's privacy and records policies, including where the transcripts are stored and who can read them. Whether and how you may record a particular conversation is a question for your privacy or legal team, and this article is not legal advice, so ask them before you start.
- How many transcripts do we need before the model is useful?
- Enough that the steps stop changing between transcripts. While each new transcript adds a step, a variant or a decision, keep going; when a new transcript only confirms what the model already holds, you have reached the limit of what talk can tell you and it is time to add a document and a recording of the job done once.
See it against your own processes
A demo runs on your material, not a canned dataset. Bring a process you find hard to explain.