Key takeaways
- Senior expertise is an action library stored in a person, retrievable only by them, and lost when they leave. Externalising it is both a risk-management and an automation prerequisite.
- Name each action with a verb a stranger would understand. Vague names are where automation fails silently.
- Aim for eight to twenty-five actions per matter type. Fewer usually means you have described phases; more usually means keystrokes.
- Mark each action routine or judgment. Routine and high-frequency is your automation queue, in that order.
- Extract by observing real closed matters, not by asking a lawyer to describe their process from memory - the two answers differ substantially.
Why 'we know how we do it' is not documentation
Within any practice area, actions repeat. A commercial lease review in one city and a commercial lease review in another are not the same matter, but they are overwhelmingly the same sequence with different facts poured in. Senior lawyers experience this as expertise. Structurally it is an action library stored in a human being, indexed by memory, retrievable only by that person, and deleted from the practice on the day they leave.
This is tolerable in a traditional practice because juniors absorb the library slowly by proximity, and the firm accepts the cost of that transfer as normal. It becomes intolerable the moment you want to automate anything, because an agent cannot absorb by proximity. It needs the sequence stated.
There is a second, quieter cost. An undocumented sequence cannot be improved deliberately. It drifts, varies by who is running it, and its failure modes are invisible because nobody can compare a given execution against a stated standard. Firms consistently discover, on documenting a sequence for the first time, that three lawyers were running materially different processes and each believed theirs was the firm's process.
The extraction method that actually works
The instinctive approach - sit the senior lawyer down and ask them to describe their process - produces a clean, confident, and substantially incomplete answer. People describing familiar expert processes systematically omit the steps they have automated in their own heads, and reconstruct a tidier sequence than the one they actually run. You will get about sixty per cent of the real library this way.
The method that works is to reconstruct from artefacts, then verify with the expert. Read what actually happened on real matters, build the sequence from evidence, and then walk the expert through your draft. They will correct it accurately, because recognising an error in a stated sequence is a much easier cognitive task than generating the sequence from scratch.
- 01
Pick one matter type and five closed matters
Highest volume first. Five is enough to see the pattern and few enough to read properly. Include at least one that went badly, because failure modes are part of the library.
- 02
Build a timeline of every action taken on each
From documents, emails, file notes, time entries, and calendar entries. Do not interpret yet - just record what was done and in what order, matter by matter.
- 03
Find the common spine
Lay the five timelines side by side. The actions appearing in all five, in roughly the same order, are the spine of your action library. The ones appearing in some are conditional branches, and you should note what condition triggers them.
- 04
Name each action with a verb
'Draft the reply', 'verify the counterparty', 'confirm the filing deadline'. Not 'preliminary work' or 'handle the response'. If a stranger cannot tell what they would physically do from the name, rename it.
- 05
Walk the expert through the draft
Ninety minutes, with the draft in front of them. Ask what is missing, what is misordered, and what condition triggers each branch. This session is where the remaining forty per cent surfaces, and it only works because there is a concrete draft to react to.
- 06
Pressure-test against three more closed matters
Run the finished library against matters you did not use to build it. Anything that does not fit is either a missing branch or evidence that you have two matter types wearing one name.
Getting the granularity right
The most common failure in workflow documentation is picking the wrong altitude. Too coarse and the document is a phase list that tells nobody anything they did not already know. Too fine and it becomes an unmaintainable procedure manual that nobody reads and that is wrong within a month.
| Altitude | Example | Verdict |
|---|---|---|
| Too coarse | 'Prepare the case' | A phase, not an action. Cannot be assigned, timed, or automated. |
| Right | 'Draft the letter of response from the standard template' | Assignable to a named person or agent, timeable, and either done or not done. |
| Right | 'Verify counterparty details against the register' | Discrete, checkable, and a genuine candidate for automation. |
| Too fine | 'Open the template document' | A keystroke. Documenting at this level produces a manual nobody maintains. |
Marking routine versus judgment
Once the library exists, go through it and mark each action. The test is not difficulty and not importance - it is whether the same inputs reliably produce the same shape of output. Assembling a first draft from a template against known facts is routine even though it takes skill. Deciding whether to concede a point in negotiation is judgment even though it takes a minute.
Sort the routine actions by frequency. That ordered list is your automation queue, and it is now a concrete engineering brief rather than an aspiration. It is also, incidentally, your delegation plan for human juniors, which is why this work pays for itself even at practices with no interest in AI at all.
The line that matters
Routine does not mean unimportant, and automating a routine action does not remove the lawyer from it. It moves the lawyer from originating the output to reviewing it - the same professional responsibility, applied to a better starting point.
Keeping the library alive
A documented workflow that is not maintained is worse than none, because people trust it and it is wrong. The maintenance burden is manageable if you build two habits from the start.
First, the library lives with the matter. A checklist attached to each matter of that type, worked through in practice, is a living document that surfaces its own errors - when someone skips a step or adds one, that is visible. A playbook in a shared drive that nobody opens during actual work will rot silently.
Second, schedule one review a quarter, thirty minutes, with whoever ran the most matters of that type. Ask two questions: what did you do that is not on the list, and what is on the list that you skipped. Both answers are improvements. This is what makes the practice get measurably better over time as an institution rather than only inside individual people's accumulating experience.
Frequently asked
How do I document a legal process my senior lawyers do from memory?
Do not start by asking them to describe it - experts systematically omit steps they have internalised, and you will capture roughly sixty per cent. Instead, reconstruct the sequence from five real closed matters using documents, emails, and file notes, then walk the expert through your draft. Correcting a concrete draft is far more accurate than generating the sequence from memory.
How many steps should a legal workflow have?
Eight to twenty-five actions for most matter types. Fewer than eight usually means you have documented phases rather than actions, which cannot be assigned or automated. More than forty usually means you have descended to keystrokes, producing a manual that will not be maintained. The right altitude is an action that a named person or agent could be given and that is either done or not done.
What is an action library in a law firm?
The written, ordered set of named, repeatable moves that carry a matter of a given type from opened to resolved. It converts tacit expertise held in individual lawyers into a firm asset that can be taught, delegated, improved deliberately, and - once routine actions are identified - automated.
Should each lawyer have their own workflow?
No, and discovering that they do is usually the most valuable finding of the documentation exercise. Variation across lawyers running the same matter type is not personalisation, it is undetected drift - and it means quality depends on which lawyer picked up the file. Converge on one sequence, with explicitly documented branches for genuine conditional differences.
How often should legal workflow documentation be updated?
Quarterly, in a thirty-minute review with whoever has run the most matters of that type, asking two questions: what did you do that is not on the list, and what is on the list that you skipped. Between reviews the checklist should live attached to each matter, so deviations surface during real work rather than being discovered a year later.