Lesson objective
Separate hands-on time from elapsed time across a mapped workflow, locate the queues and rework that consume most of the elapsed time, and identify which steps a faster tool would and would not change.
Plain-language explanation
Why this matters now
Most legal AI proposals are time arguments. The task takes six hours and will take two. The abstraction takes a week and will take a day. Very few of these arguments distinguish between the two kinds of time in a workflow, and the distinction decides whether the promised saving can exist at all.
Hands-on time is time somebody is working on the matter. Elapsed time is time between the trigger and completion. In most legal workflows the second is several times the first, and the difference is not work. It is queueing: the file sitting in an inbox, waiting for a signature, waiting for a document from the client, waiting for the only person who can decide to come back from court.
An assistive tool acts on hands-on time. It does not clear queues. A step that consumes ninety minutes of hands-on time inside a workflow with an elapsed time of nine working days can be made instant and the client will notice almost nothing, because the ninety minutes was never what they were waiting for.
Industry survey: Clio's 2025 Legal Trends Report (16 October 2025) reports adoption plateauing at 79%, with 45% of wide adopters having adjusted prices and 36% reporting a positive revenue effect. It is a vendor survey of its own market and should be read as one. What it points at is that the conversation has moved from capability to economics — and an economic claim about a workflow is only as good as the time data underneath it.
Professional guidance: ABA Model Rule 1.5 governs fees, and Rules 1.1, 1.4, 5.1 and 5.3 govern competence, client communication and supervision. These are model rules; each jurisdiction adopts its own numbering and wording, and you must read the version that applies where you practise. The operational consequence for this lesson is narrow and important: any claim about time saved in a workflow eventually becomes a claim about what may be billed, so the measurement has to be honest before anyone repeats it outside the firm.
A question to sit with
Your proposal says the workflow takes eight days and will take three. Which of those days is somebody working, and which is the file sitting on a desk waiting for a person who is in court?
What goes wrong
Elapsed time reported as effort. The map says "two days" for a step. Asked what that means, the performer says the file usually comes back the day after tomorrow. That is a queue with a work item somewhere inside it, and treating it as ninety minutes of effort or as two days of effort both mislead, in opposite directions.
Optimising the visible step. The step with a name and an owner gets improved, because it is the step you can see. The wait either side of it has no owner and no name, and so it is invisible to the improvement effort and survives it intact.
Hands-on time estimated from memory, upward or downward. People underestimate tasks they do often and overestimate tasks they dislike. Neither error is deliberate and both are large. A handful of timed observations beats any amount of confident recollection.
Rework counted once. A step performed twice because work came back appears in most time records as one step. A loop running on one matter in five can be the largest single consumer of hands-on time in a workflow and never appear in a summary of it.
The review burden is left out entirely. Module 1 made this point about pilots and it recurs here as a measurement rule: a workflow that halves producer minutes and doubles reviewer minutes has moved work to a more expensive person. If reviewer time is not on the map, the map cannot show that.
Benchmark speed treated as workflow speed. How fast a tool returns an output is not how fast the step goes. The step includes assembling the input, checking the output against a source, deciding what to do with a flagged item, and recording what was done. Evidence: the randomized trial in AI-Powered Lawyering (Schwarcz and others, posted 2 March 2025; Journal of Law and Empirical Analysis vol. 3, issue 1 (2026)) measured productivity gains of 50% to 130% in five of six tasks, with quality improvements and no added hallucinations from the retrieval-augmented tool — measured on law students using 2024 tools in study conditions. Those are task-level gains under controlled conditions, not workflow-level gains in a firm, and the gap between the two is exactly the queueing, checking and recording this lesson is about.
The practical answer
Record two durations in every step row, and never merge them. Part 2 of malc-l3-workflow-map.md asks for hands-on and elapsed separately for exactly this reason.
| Duration | Definition | How to capture it |
|---|---|---|
| Hands-on | Time a person is working on this matter at this step | Timed observation of three to five real runs, or a short timed self-record |
| Elapsed | Trigger of the step to the handoff onwards | Timestamps where they exist; the file's own history where they do not |
Then build the queue table. Part 3 of the template is where most of the finding lives.
| Between steps | Handoff method | Typical wait | Longest observed | Cause of the wait | Recorded anywhere? |
|---|
The "longest observed" column earns its place. A workflow whose typical wait is half a day and whose longest is nine days is not a workflow with a half-day wait; it is one with a dependency on a person or an event that sometimes is not available, and the client experience is set by the tail, not the median.
Do the arithmetic that exposes the ceiling. For any step you propose to speed up, compute three numbers and write them in a sentence.
| Number | Definition |
|---|---|
| Step hands-on time | What the step costs today |
| Total workflow elapsed time | Trigger to completion |
| Best possible improvement | Step hands-on time as a share of total elapsed time |
If a step consumes ninety minutes inside an elapsed time of nine working days, then eliminating the step entirely improves end-to-end time by roughly two per cent. AI Coalition Network recommendation: write that sentence into any automation proposal before the benefits section, in the form "even if this step became instantaneous, end-to-end time would fall by about X". It is the fastest way to find out whether you are solving the problem the client actually experiences.
Distinguish the three things a time change can be for. They are not the same and they need different evidence.
| Goal | What it improves | The measure that proves it |
|---|---|---|
| Capacity | More matters through the same team | Hands-on minutes per matter |
| Turnaround | The client waits less | End-to-end elapsed time, including the tail |
| Cost mix | Cheaper people doing more of the work | Producer minutes and reviewer minutes, separately |
A proposal that claims all three usually has evidence for none. Pick one, and measure the one you picked.
Measure reviewer minutes as a first-class number. Best practice: record producer minutes per deliverable and reviewer minutes per deliverable as two separate figures in every row where an assistive step exists, and keep them separate in every report. Best practice: the NIST AI RMF (AI RMF 1.0, 26 January 2023) names Measure as one of its four functions, and the NIST AI RMF Playbook offers a menu of concrete actions — documented roles, inventories, decommissioning plans — from which a firm picks what fits its size. Both are voluntary, and neither requires anything of a law firm; they are useful because they supply shared vocabulary. Nothing in this paragraph is required by NIST.
Count rework properly. For each loop found in the previous lesson, record how often it runs and the hands-on minutes it consumes when it does, then express it as minutes per matter across all matters rather than per affected matter. A loop costing three hours on one matter in five costs thirty-six minutes per matter, and that is the number that belongs beside the step times.
Why does this keep surprising experienced teams?
Because hands-on time is what people experience and elapsed time is what clients experience, and the two are almost never discussed in the same meeting. Practitioners describe the work they do; clients describe the waiting they do. A map with both columns filled in is often the first document in which the two accounts of the same workflow sit side by side, and the gap between them is usually large enough to change what the firm decides to fix.
Remember this
- Hands-on time and elapsed time are different quantities; a map that merges them cannot support any claim about either.
- Assistive tools act on hands-on time and do not clear queues, so a workflow dominated by waiting will not get faster end to end.
- Compute the best possible improvement before the benefits section: "even if this step were instantaneous, end-to-end time would fall by about X".
- Capacity, turnaround and cost mix are three different goals with three different measures; a proposal claiming all three usually proves none.
- Reviewer minutes are a first-class number. A workflow that halves producer time and doubles review time has redistributed cost, not saved it.
Legal example
Pemberton Adeyemi LLP is a fictional 80-lawyer real estate practice, and its invented operations manager, Grace Ohlsson, is asked to evaluate an AI-assisted title report summarisation step. The business case in front of her is straightforward: the summarisation currently takes a paralegal about two hours per transaction, the firm runs roughly 120 transactions a month, and the vendor proposal says the step would take fifteen minutes. That is 210 hours a month, and the practice head has already described it in a partners' meeting as cutting the firm's turnaround time in half. Grace maps the durations. Hands-on time across the whole workflow, measured by timing four real transactions end to end, comes to about six and a half hours spread across nine steps. Elapsed time from instruction to issued report averages eleven working days. The summarisation step is indeed the largest single block of hands-on time. It is also two hours inside eleven days. She writes the sentence the lesson asks for: even if summarisation became instantaneous, end-to-end time would fall by roughly two per cent. Then she builds the queue table, and the picture changes. Between the searches being ordered and the results arriving, the file waits an average of four days, with a longest observed wait of nine. Between the draft report and partner sign-off, it waits an average of a day and a half, with a longest observed wait of six, because one partner covers the whole team and is regularly in completion meetings. Two handoffs leave no trace at all, so the fictional firm's belief that its turnaround is "about a week" is an average of the segments it can see. Grace also measures the review side, which the business case had not mentioned. The paralegal's summary is currently checked by an associate in about twenty minutes because the paralegal marks which entries came from which document. On a sample run of the proposed workflow, the associate takes fifty minutes, because nothing in the generated summary shows what has been verified. Grace's report recommends proceeding, but with the justification changed. The step is worth automating for capacity — it frees roughly 200 paralegal hours a month — and it should not be described to clients or partners as a turnaround improvement, because the turnaround is set by a search provider and a single signing partner. She adds one condition: the summary must mark which entries have been checked against which document, or the saving moves from a paralegal to an associate and shrinks by two thirds.
Responsible-use warning
Never quote a step-level time saving as an end-to-end improvement. In most matter workflows the elapsed time is dominated by queues — waiting for a document, a search result, a client instruction or the one person who can sign — and an assistive tool acts only on the time somebody is working. A proposal that promises a faster client experience on the strength of a faster step will be judged, months later, against a turnaround figure that has barely moved, and the resulting loss of credibility is usually spent on the next proposal too. Say which of capacity, turnaround or cost mix you are improving, and measure that one.
Human-review requirement
A licensed attorney owns every output that reaches a client or a court, and time measurement touches that ownership in two places. The supervising attorney owns whether the review gate remains performable: if reviewer minutes per deliverable rise while volume rises, the gate will thin in practice rather than slow the work down, and the attorney decides whether to narrow scope or add reviewers before that happens. The attorney also owns any statement about time that leaves the firm. Where a workflow's hours change, what the client is told and how the bill reflects time actually expended are the attorney's judgements under the fee and communication rules that apply where they practise, not the operations team's.
Practical exercise
Measure the two kinds of time across your mapped workflow. Use only fictional or non-confidential material, and Parts 2 and 3 of malc-l3-workflow-map.md.
- For each step, record hands-on time from timed observation of at least three real runs. Note beside each figure how it was captured, and mark any figure that came from memory.
- For each step, record elapsed time from the step's trigger to the point the work is handed on. Use timestamps where they exist and write "estimated" where they do not.
- Sum both columns. Write the total hands-on time and the total elapsed time for the workflow as two separate sentences.
- Build the queue table for every handoff: method, typical wait, longest observed wait, cause, and whether it is recorded anywhere.
- Identify the two longest waits. For each, name what the workflow is waiting for and who, if anyone, owns that wait.
- Choose the step you would most like to automate. Divide its hands-on time by total elapsed time and write the resulting sentence: "even if this step were instantaneous, end-to-end time would fall by about X".
- Add the rework you found in the previous lesson: for each loop, record its frequency and hands-on cost, and express it as minutes per matter across all matters.
- Record producer minutes and reviewer minutes separately for one deliverable, timed rather than recalled.
- State in three sentences which of capacity, turnaround or cost mix your proposal is actually improving, what measure would prove it, and what you would have to stop claiming.
Knowledge check
-
Why can a large step-level time saving produce almost no change in what the client experiences?
Answer guidance: Because the client experiences elapsed time, which in most matter workflows is dominated by queueing rather than work. An assistive tool acts on hands-on time. Two hours removed from a workflow with an eleven-day elapsed time changes the end-to-end figure by about two per cent, because the two hours were never what the client was waiting for.
-
What are the three distinct goals a time improvement can serve, and how is each proved?
Answer guidance: Capacity, proved by hands-on minutes per matter; turnaround, proved by end-to-end elapsed time including the long tail; and cost mix, proved by producer minutes and reviewer minutes recorded separately. A proposal that claims all three usually has evidence for none, so pick one and measure that one.
-
Why must reviewer minutes be recorded separately from producer minutes?
Answer guidance: Because a workflow can halve the producing time while doubling the checking time, which moves work to a more expensive and scarcer person. It also has a supervision consequence: if the review burden exceeds what named reviewers can perform at volume, the gate thins in practice rather than the work slowing down, and only a separate reviewer figure makes that visible.
-
A step's wait is typically half a day and was once nine days. Which number describes the workflow?
Answer guidance: Both, and the long one matters more than teams expect. A nine-day tail means the step depends on a person or an event that is sometimes unavailable, and the client's experience of the firm is set by the bad runs rather than the median. Recording the longest observed wait beside the typical one is what makes that dependency visible.
Sources and further reading
- Clio (Themis Solutions) — 2025 Legal Trends Report (2025-10-16). 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.
- 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.
- National Institute of Standards and Technology (AI Resource Center) — NIST AI RMF Playbook. 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.