Lesson objective
Identify the exceptions, decision points and handoffs in a mapped workflow, and record for each one the trigger, the off-ramp, the role that decides and whether the handoff leaves any trace.
Plain-language explanation
Why this matters now
The happy path is the part of a workflow everyone can describe. It is also, in most legal matters, the minority of the work. The exceptions are the work: the file that arrives incomplete, the party who turns out to be an affiliate, the client who changes instructions after the draft is out, the matter that has to route past an ethical wall. A map that records the happy path in detail and the exceptions as "sometimes it is different" has recorded the easy half and left out the half that decides how the workflow behaves under pressure.
This matters more with an assistive tool in the workflow than without one. Tools are trained and configured for the shape of work they are shown. The routine path is the shape that gets shown. When an exception arrives, the tool does not usually announce that it has encountered something unusual; it produces its ordinary output, with its ordinary fluency, about a situation its instructions did not cover.
Evidence: the randomized trial reported by Schwarcz, Manning, Prescott, Barry, Cleveland and Rich, AI-Powered Lawyering (posted 2 March 2025; Journal of Law and Empirical Analysis vol. 3, issue 1 (2026)), found a retrieval-augmented legal tool and a reasoning model each producing productivity gains of 50% to 130% in five of six tasks, with improved quality and — for the retrieval tool — no added hallucinations. Note the shape of that finding as well as its size: five of six. The tasks were performed by law students using 2024 tools under study conditions, which limits what can be carried across to a firm. What survives the caveats is a useful operational instinct: gains are real and they are uneven, and knowing which of your tasks is the sixth one requires knowing where your workflow stops being routine.
A question to sit with
Your map records what happens when the file is complete, the parties are clear and the deadline is comfortable. How often is that actually true — and who currently decides what happens when it is not?
What goes wrong
Exceptions recorded as a mood. "Sometimes it is different", "it depends", "we deal with those case by case". Each of these is a real practice being described by someone who has never been asked to name its trigger. The remedy is mechanical: ask for the last three times it happened.
The exception rate is unknown and assumed small. Teams routinely estimate an exception at five per cent and find, on counting, that it is a third. The estimate is not dishonest; exceptions are handled and forgotten, while the routine path is what people remember doing all day.
A decision is recorded without its decider. The map says "conflicts cleared or flagged" and does not say who decides, or what happens to the matter while the question is open. Under deadline, an unowned decision resolves in favour of proceeding — module 1 made the same point about escalation, and it is the same mechanism.
Authority is inferred from a job title. Who may decide and who actually decides are different questions, and the map must record the second while noting the first. Where they differ, that is a finding for the supervising attorney, not a cell to tidy up.
Handoffs that leave no trace. A file that moves by a spoken word, a corridor conversation or an unrecorded email has moved invisibly. Nobody can say when it moved, who had it, or how long it waited. A handoff with no record is also a supervision gap: Law / rule: Model Rules 5.1 and 5.3 place responsibility on supervising lawyers for the work of other lawyers and non-lawyer assistance, and Rule 1.4 governs communication with the client. These are model provisions, adopted differently in each jurisdiction; check the ones that bind you. Supervision of a step you cannot see happening is supervision in name.
Rework loops treated as normal life. Work that goes backwards is invisible in most measurement, because the step gets counted once and performed twice. A loop that runs on one matter in five is a substantial hidden cost and usually a sign that an earlier step has an undefined input.
The practical answer
Take the three things in turn, and give each a structure rather than a sentence.
Exceptions: name the trigger, the off-ramp and the frequency.
| Field | What to record | Why |
|---|---|---|
| Trigger | The observable condition: "the claim file arrives without a policy schedule" | A trigger you cannot observe cannot be detected by anyone, or automated later |
| Frequency | Counted, not estimated, over a defined recent period | Estimates of exception rates are consistently too low |
| Off-ramp | What is actually done, step by step | Off-ramps are where undocumented method concentrates |
| Who handles it | The role, and whether it differs from the routine role | Exceptions often fall to whoever is senior and available |
| Return path | Where the matter rejoins the main workflow, or does not | A file that never rejoins is an orphaned matter |
| Tool behaviour | What the assistive step does when this input arrives | Usually the answer is "produces ordinary output", which is the finding |
Count exceptions from the file, not from memory. Take the last thirty completed matters in this workflow and classify each as routine or exceptional against your named triggers. Thirty is small, and you should say so when reporting it, but a counted thirty beats a remembered impression every time. AI Coalition Network recommendation: report the exception rate as a fraction with its denominator visible — "nine of the last thirty" — rather than as a percentage, so that nobody mistakes a small sample for a measurement.
Decision points: four cells, every time.
| Cell | Example |
|---|---|
| The question actually asked | "Does this matter need the practice head’s clearance before work begins?" |
| The role that answers it | Conflicts analyst, escalating to practice head |
| What the matter does while open | Held; no substantive work begins |
| Where the answer is recorded | The conflicts report field in the matter system |
Classify each decision by what it does. This classification is what module 3 tiers and module 4 writes down.
| Decision type | Description | Can an assistive tool be near it? |
|---|---|---|
| Gate | Permits or blocks the next step | It may assemble the information; it may not decide |
| Route | Sends the matter down one of several paths | It may propose a route with reasons, for a person to confirm |
| Judgement | Applies professional skill to facts | The judgement stays with the qualified person; drafting support is a separate question |
| Ministerial | Applies a fixed, written rule to observable facts | Suitable for assistance, with the output checkable against the rule |
The dividing line worth remembering: a tool may prepare what a decision needs. A tool may not be the place the decision is recorded as having been made. Best practice: where an assistive step proposes a route or a classification, the map records both the proposal and the human confirmation as separate rows, because collapsing them into one row is how a proposal quietly becomes a decision.
Handoffs: record the method, the wait and whether it leaves a trace. Part 3 of malc-l3-workflow-map.md is built for this, and three of its columns matter most.
| Column | What it exposes |
|---|---|
| Handoff method | Verbal and email handoffs are invisible to every later measurement |
| Typical wait and longest observed wait | The gap between them is the workflow's real unpredictability |
| Is the handoff recorded anywhere? | A "no" here means a step nobody can supervise or time |
Trace one real file end to end. Interviews give you steps; following an artefact gives you handoffs. Take one completed matter, find every point at which the work changed hands, and record the timestamp if one exists and "none" if it does not. A workflow with four handoffs and one timestamp has one visible junction and three blind ones.
Rework loops: from which step, back to which step, how often, who notices. A loop is not a failure to be hidden; it is a measurement. Record its trigger honestly, because the trigger almost always names an earlier step whose inputs or acceptance criteria are undefined — and that earlier step, not the step doing the rework, is where the fix belongs.
Remember this
- Exceptions are the real work; record each with a trigger, a counted frequency, an off-ramp, a role and a return path.
- Count exception rates from the last thirty matters and report the fraction with its denominator, never a remembered percentage.
- Every decision point needs four cells: the question, the role that answers, what the matter does meanwhile, and where the answer is recorded.
- A tool may prepare a decision; it may never be where the decision is recorded as made. Keep proposal and confirmation as separate rows.
- A handoff with no record is a blind junction: nobody can time it, supervise it or notice when it stalls.
Legal example
Aldreth Sowande is a fictional 35-lawyer commercial practice. Its invented knowledge management lead, Tomas Iversen, maps the firm's standard-form supplier agreement review, expecting a tidy six-step workflow with one decision point. The happy path is indeed six steps. The exceptions take him another week. Asking each of the four reviewers to describe the last three times a review did not go normally, he collects eleven instances and resolves them into four triggers. First: the counterparty sends a marked-up version of a form the firm has never seen, which happened four times in the eleven. Second: the client asks for the review to cover a jurisdiction outside the playbook. Third: the agreement incorporates a separate document by reference that nobody has. Fourth: the client wants the review turned round inside a day. He then counts the last thirty completed reviews against those four triggers and finds eleven exceptional, which he reports as "eleven of the last thirty" rather than 37%, because thirty is a small sample and he wants it visible. The estimate he had been given at the start of the week was "maybe one in ten". The decision points surprise him more. The map as written showed one: whether the review passes to a partner. Walking it, he finds three. Somebody decides whether a deviation from the playbook is material. Somebody decides whether an unfamiliar form is reviewed against the playbook at all or reviewed from scratch. Somebody decides whether the same-day request is accepted. Two of the three had no recorded owner, and both defaulted, under deadline, to whichever reviewer had the file. Handoffs are worse. Four handoffs exist and one leaves a trace — the partner sign-off, which is in the document management system. The other three happen by email or by a message in a chat channel, so nobody can say how long a review sits between steps, which means the fictional firm's belief that its turnaround time is a day and a half is an average of the parts it can see. Tomas's recommendation is not to automate anything yet. It is to give the three unowned decisions an owner and to make one invisible handoff visible, on the ground that a tool introduced before either would be measured against a baseline the firm does not actually have.
Responsible-use warning
Never let an exception path inherit the risk treatment of the routine path it branches from. Exceptions exist precisely because something unusual or something wrong is true, so the consequence of a bad output on an exception path is typically higher, while the tool is less dependable and the reviewer less alert — an unusual input rarely produces an unusual-looking output. A workflow mapped and tiered only on its happy path will therefore apply its lightest controls exactly where its heaviest ones are needed, and nothing in the routine path's performance record will reveal the problem.
Human-review requirement
A licensed attorney owns every output that reaches a client or a court, and exception handling is where that ownership is most often diffused by accident. The supervising attorney owns three judgements here: which decision points are professional judgements that may not be delegated to an assistive step at all, whether the recorded owner of each gate is the person who actually decides in practice, and whether any exception path may proceed with tool assistance before its own controls have been set. Where a decision point currently has no named owner, the attorney names one before the workflow changes, not after.
Practical exercise
Extend your as-is map into its exceptions, decisions and handoffs. Use only fictional or non-confidential material and Parts 2 and 3 of malc-l3-workflow-map.md.
- Ask each person who performs the workflow to describe the last three times it did not go normally. Record the instances verbatim before interpreting them.
- Resolve the instances into named triggers, each written as an observable condition rather than a category.
- Take the last thirty completed matters in this workflow and classify each as routine or exceptional against your triggers. Report the result as a fraction with its denominator.
- For each trigger, record the off-ramp step by step, the role that handles it, and where the matter rejoins the workflow — or that it does not.
- List every decision point. For each, write the question actually asked, the role that answers, what the matter does while the question is open, and where the answer is recorded. Write "not recorded" where that is the truth.
- Classify each decision as gate, route, judgement or ministerial, and write one line on why.
- Follow one completed matter from trigger to completion and record every handoff: method, whether a timestamp exists, typical wait and the longest wait anyone can recall.
- Identify one rework loop. Record where it starts, where it sends work back to, its trigger, who notices, and what the earlier step would need in order for the loop to stop existing.
- Write three sentences naming the exception path you would be least comfortable running with an assistive tool today, and why.
Knowledge check
-
Why does asking about "the last three times" work better than asking what the exceptions are?
Answer guidance: Because a general question produces categories — vague, tidy and usually flattering. Instances produce the things the map needs: an observable trigger, a countable frequency, the off-ramp actually taken and the role that took it. You record the instances first and decide what they mean afterwards.
-
What four cells must every decision point on the map carry?
Answer guidance: The question actually asked, the role that answers it, what happens to the matter while the question is open, and where the answer is recorded. Without the third, an unowned decision resolves under deadline in favour of proceeding; without the fourth, nobody can later show that the decision was made at all.
-
An assistive step proposes which of three routes a matter should take, and a person confirms it. How should the map record that?
Answer guidance: As two separate rows: the proposal and the human confirmation. Collapsing them into one row is how a proposal quietly becomes a decision. A tool may prepare what a decision needs; it may not be the place the decision is recorded as having been made.
-
Why is a handoff that happens by email or conversation a problem for an operation rather than merely untidy?
Answer guidance: Because it leaves no trace. Nobody can say when the work changed hands, who held it, or how long it waited, so the step cannot be timed, cannot be supervised and cannot be seen to have stalled. A turnaround figure built from only the visible junctions is an average of the parts the firm happens to be able to see.
Sources and further reading
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-14.
- Daniel Schwarcz, Sam Manning, J.J. Prescott, Patrick Barry, David R. Cleveland and Beverly Rich (Journal of Law and Empirical Analysis) — AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice (2025-03-02). Accessed 2026-09-14.
- American Bar Association — Model Rules of Professional Conduct. Accessed 2026-09-14.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-14.
- National Institute of Standards and Technology (AI Resource Center) — NIST AI RMF Playbook. Accessed 2026-09-14.