Lesson objective

Produce an as-is map of a matter workflow by walking it with the people who perform it, recording each step as a verb and an object with its role, inputs, outputs and error modes.

Plain-language explanation

Why this matters now

Module 1 ended with a readiness gate, and the first line of that gate was a mapped workflow. This lesson is where the mapping happens, and it is the least glamorous work in the level. It is also the work that decides whether everything after it is useful or merely energetic.

The reason is simple and easy to underrate. A tool introduced into a process nobody has described does not make the process faster. It makes the failure harder to locate. When a defect appears three months later, the firm cannot say which step produced it, because the firm has never agreed what the steps are. Every diagnosis becomes an argument between people who each remember a different version of the same workflow, and all of them are describing something real.

The commercial pressure to skip this step is genuine. Industry survey: Clio's 2025 Legal Trends Report, published 16 October 2025, describes adoption plateauing at 79% and a shift in attention towards pricing — 45% of wide adopters had adjusted prices and 36% reported a positive revenue effect. That is a vendor survey of its own market rather than a census, and it should be read as one. What it describes, though, is a profession past the stage of asking whether to use these tools and into the stage of asking what they change about the economics of the work. You cannot answer that about a workflow you cannot describe.

Best practice: the NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, published 26 January 2023) calls this function Map, and places it before Measure and Manage for the same reason this module comes before module 13. The AI RMF is a voluntary framework. Adopting it neither creates nor discharges a professional duty, and nothing in it tells you what the rules where you practise require of you.

A question to sit with

You are about to change a workflow you have never watched anyone perform. Whose description of it are you changing — the policy's, your own memory's, or the one that actually happens on a Tuesday afternoon under deadline?

What goes wrong

Mapping the policy instead of the practice. The fastest way to produce a map is to read the procedure document and write it out as steps. The result is clean, agreeable, and a description of something that does not occur. Every real workflow has accumulated adaptations that nobody recorded because each one was sensible at the time. Those adaptations are where the risk sits, and a policy-derived map is precisely the map that cannot see them.

Mapping from the top of the hierarchy. A partner's account of a workflow is an account of the parts a partner touches. The steps that consume most of the elapsed time are usually performed by people who were not asked. This is not a criticism of partners; it is a structural feature of how work is visible from different seats.

Steps named after departments. "Intake." "Review." "Conflicts." These are not steps, they are places. A step is one verb and one object performed by one role and then handed on: extract the stated dates from the complaint. If you cannot write it that way, you have not found the step yet, you have found a container holding several.

Naming people rather than roles. A map that says "Priya runs the comparison" dies the week Priya changes teams. Best practice: every cell names a role, and individual names live in staffing notes only. The map has to survive its informants.

Recording only the happy path. People describe the version of the workflow that goes smoothly, because that is the version that has a shape. The exceptions get compressed into "sometimes it is different", which is the single most expensive sentence in workflow mapping. Module 2's second lesson takes that sentence apart.

Listing errors nobody can detect. A map that records an error mode without recording how the error is currently noticed has recorded a wish. If the honest answer is "we would not find out", write that down. It is the most valuable finding on the page.

The practical answer

Work from the malc-l3-workflow-map.md template, and fill it by walking rather than by writing. The template has six parts; this lesson covers Parts 1 and 2, the next lesson covers Parts 2 and 3, and the lesson after covers Part 3 and the baseline.

Part 1 is a header, and it is harder than it looks. Several of its fields are diagnostic in themselves.

Header field Why it is diagnostic
Trigger — what starts this workflow Teams often find three triggers where they assumed one, and one of them bypasses the record
Completion — what "done" means, observably If nobody can name an observable finish, the workflow has no boundary and cannot be measured
Volume per month, and how counted "How counted" exposes work that never enters a system
Client-facing? Court- or regulator-facing? Sets the consequence axis you will use in module 3
Highest data classification handled Decides which tools may touch the workflow at all
Rules or duties that constrain it Written in by you, with a record of where you checked

That last row is the template's deliberate blank. It does not tell you what any rule requires, because requirements differ by jurisdiction and change. You write in the constraint and record where you checked it.

Walk the workflow, in this order.

Step in the walk What you do What it produces
1. Sit with a daily performer Watch one real run end to end, saying nothing The steps that exist but are never mentioned
2. Interview separately Ask two more performers to describe it without conferring The variance between accounts
3. Follow the artefact Trace one real file from trigger to completion The handoffs and the waits
4. Ask what they were not asked "What did you do that is not in what I have written?" The workarounds
5. Play it back Read your map aloud to the performers Corrections, which arrive fast when the map is wrong

Separate accounts before shared ones. Interviewing three people together produces one account, usually the most senior person's, and the disagreements never surface. Interviewing them separately produces three accounts and a list of differences, and that list is the real output of the exercise. Module 1 called the differences drift; here they are simply data.

Write each row to the column discipline. The template's guidance is worth restating because each column has a characteristic failure.

Column Write this Not this
Step A verb and an object A department or a state
Who performs it The role The individual
Inputs Documents, approvals, access, engagement status Documents only
Outputs The artefact and where it is saved "Reviewed"
Decision point The question asked, and the role that answers The decision without the decider
Exceptions The named off-ramp and what happens on it "Sometimes it is different"
Duration Hands-on and elapsed, separately Elapsed reported as effort
Error modes What goes wrong and how it is currently detected An error with no detection route
AI column Appropriate / Conditional / Not appropriate, with the reason A judgement based on how easy the task looks

Resist marking the AI column while you walk. The template puts it last on the row for a reason. The moment a mapper starts deciding what a tool could do, the map begins bending towards what would be convenient to automate, and steps that are inconvenient get compressed. AI Coalition Network recommendation: complete the whole step grid with the AI column blank, then go back and fill it in a separate sitting, with the performers present.

Three patterns belong in "Not appropriate" almost every time. The template names them and they are worth memorising, because each looks easy.

  1. Fluent but unverifiable in the time available. A reviewer who cannot check the substance checks whether it reads well instead, and fluency is exactly what these tools produce most reliably.
  2. Asked for an absence. What is missing from a document, a party never named, a deadline nobody calendared. There is nothing on the page for the tool to point at, and nothing for the reviewer to check against.
  3. Output becomes an input to a later step without being read on its own. An error is laundered into the file and acquires the authority of something already in it.

Expect the map to find things that have nothing to do with AI. This is normal and it is a feature. Firms that map a workflow honestly typically discover an unrecorded intake route, a permissions default nobody set, or a step that exists only because a system used three years ago required it. Best practice: record those findings separately and fix them on their own merits, rather than folding them into the automation proposal where they will be judged by whether they help the business case.

Why not just start with the to-be design?

Because a to-be design built on an imagined as-is inherits every assumption nobody checked. The commonest result is a redesign that optimises a step consuming ten per cent of the elapsed time, leaving a queue that consumes sixty per cent untouched — and then a post-implementation report that cannot explain why nothing got faster. The next two lessons exist to prevent exactly that.

Remember this

  • Map the workflow as it is actually performed, by walking it, not as the procedure document describes it.
  • A step is one verb, one object, one role, then a handoff. Departments and states are not steps.
  • Name roles, never individuals; the map has to outlive the people who described it.
  • An error mode with no detection route is the most valuable line on the map — write it down as a gap.
  • Fill the AI-appropriateness column in a separate sitting, after the grid is complete, or the map bends towards what is convenient to automate.

Legal example

Loxley Fenwick LLP is a fictional 60-lawyer insurance defence firm. Its invented legal operations lead, Naomi Brathwaite, is asked to introduce AI-assisted summarisation into the firm's new-claim assessment workflow. She is given four weeks and a strong steer from the practice head that the summarisation step is the bottleneck. She spends the first week mapping instead. The procedure document describes seven steps. Naomi sits with a claims paralegal for one full run and finds eleven, and two of the four extra steps turn out to matter. The first is a call the paralegal makes to the referring adjuster whenever the claim file arrives without the policy schedule, which is roughly a third of the time. Nobody had recorded it because it is not part of anyone's job description; it just gets done. The second is a manual re-keying of dates from a scanned acknowledgement letter into the matter system, which exists because a system replaced three years ago could not read the letter format, and which is now performed out of habit into a system that can. She then interviews two more paralegals separately. All three describe the assessment step differently. One reads the full claim file before summarising, one reads the first and last pages and the schedule, one reads whatever the adjuster highlighted. All three believed they were doing the standard thing. Naomi writes the differences down without adjudicating them, because the point of the map is to record what happens. When she fills the error-modes column, the most useful cell is nearly blank: the workflow's worst failure is a summary that omits a coverage exclusion, and when she asks how that is currently detected, the honest answer is that it is not, unless the exclusion becomes relevant later in the matter. She writes "no current detection" in the cell. Her four-week report contains a recommendation the practice head did not expect. The summarisation step is not the bottleneck, and the fictional firm's first change should be to define what must be read before a summary is written, because no tool can fix an input set that three people define three ways.

Responsible-use warning

Never build a workflow map from the procedure document, the intranet page or a senior person's description alone. Each of those is a description of the intended workflow, and the risk in a workflow lives almost entirely in the difference between the intended version and the performed one — the undocumented call, the informal workaround, the step people skip when the deadline is tight. A map that matches the policy exactly has almost certainly not been walked, and it will produce a redesign that solves problems nobody has while leaving the real ones in place, invisible and now harder to find because a tool has been added on top of them.

Human-review requirement

A licensed attorney owns every output that reaches a client or a court, and mapping does not change that. What the supervising attorney owns at this stage is the accuracy of the map itself: that it describes the work as performed rather than as intended, that every step marked Conditional or Appropriate names the verification step that will follow it, and that any constraint arising from a professional duty has been written in with the source recorded rather than assumed. Where a map shows an error mode with no detection route, the attorney decides whether the workflow may continue in its current form at all — that decision is not the mapper's to make.

Practical exercise

Produce an as-is map of one workflow you personally touch. Use only fictional or non-confidential material, and use the malc-l3-workflow-map.md template.

  1. Choose a workflow with a clear trigger and an observable completion, running at least ten times a month. Complete Part 1 of the template, leaving no field blank; write "unknown" where you do not know, and mark it to find out.
  2. Watch one real run end to end without intervening. Write down every action you see, including the ones the performer does not mention.
  3. Convert what you saw into the Part 2 grid. Every step must be a verb and an object, and every "Who" cell must name a role.
  4. Interview two further performers separately, without letting them confer. Record their accounts as separate columns beside your own.
  5. List every difference between the three accounts. Mark each as genuine professional judgement or as undefined method, and give a one-line reason.
  6. Complete the error-modes column. For each error, write how it is currently detected. Where the honest answer is that it is not, write "no current detection" and do not soften it.
  7. Count your steps. If you have fewer than eight, you have almost certainly recorded containers rather than steps; split them and count again.
  8. Play the map back to one performer and record every correction they make. Note how many corrections concerned the happy path and how many concerned exceptions.
  9. Leave the AI-appropriateness column blank. You will complete it after the next two lessons.

Knowledge check

  1. Why is a map built from the procedure document worse than no map at all?

    Answer guidance: Because it is confidently wrong. The risk in a workflow sits in the difference between the intended version and the performed one, and a policy-derived map is precisely the map that cannot see that difference. It then gets relied on for tiering, SOP writing and measurement, so one unchecked assumption propagates through every later decision in the level.

  2. What is the test for whether you have found a step or a container?

    Answer guidance: A step is one verb and one object, performed by one role, and then handed on: "extract the stated dates from the complaint". If the label is a department, a place or a state — intake, review, cleared — it is a container holding several steps you have not yet separated.

  3. Why should three performers be interviewed separately rather than together?

    Answer guidance: A joint interview produces one account, usually the most senior person's, and the disagreements never surface. Separate interviews produce three accounts and a list of differences, and that list is the actual output: it shows where the method is undefined, which is what the SOP in module 4 will have to settle.

  4. What should you write in the error-modes column when an error has no way of being noticed?

    Answer guidance: Write that it has no current detection, plainly. It is the most valuable finding on the map, because it identifies a failure the workflow is structurally blind to. Softening it into "detected at review" when no review looks for it converts a known gap into a false assurance that later decisions will rest on.

Sources and further reading