Law Firm Time Tracking With AI: Catch Unrecorded Time Before Billing

Start with the timers and activity capture already in your practice management system, because they keep data inside the platform. For gaps they miss, export only matter numbers, durations and activity types, never names or subject lines, and ask an AI tool under business terms to compare that with logged time.

That order matters. Most unrecorded time sits in short tasks: a ten-minute call, three emails, a document opened and closed. A suggestion engine can spot those from metadata alone, so client facts never need to leave the building. Capture is the only job here; turning the findings into good narratives is a separate job, covered in fixing time narratives before invoices go out.

Follow me on Instagram@sagnikteaches

Where unrecorded time actually hides

Before any tool, know what you are hunting. In small firms the leaks cluster in five places, and each leaves a different trace.

Connect on LinkedInSagnik Bhattacharya

These sizes are illustrative ranges for planning, not measured averages. Your own four-week sample, described below, replaces them.

Subscribe on YouTube@sagnikteaches

Step 1: use what your practice system already captures

Before any outside tool, find out what you already own. Smokeball's AutoTime page, for example, says that whether you are in Word, Outlook or Smokeball, it logs each activity to the right matter automatically and turns that activity into billable entries you control before invoicing. Clio's time-tracking page describes timers on desktop and mobile, plus AI-drafted time entries from emails sent in Outlook, edits to documents saved in OneDrive and outbound calls made in its mobile app, with nothing reaching a bill until someone approves it. Check on your own account which of these your plan includes before you rely on them.

Spend one hour on these checks before spending anything else:

1. Open the vendor's time-capture help article and write down which applications and event types it watches. Anything outside that list is a leak the platform cannot see. If the page names Outlook email, OneDrive documents and mobile-app calls, then a lawyer who drafts in a desktop folder or rings clients from a desk phone is working outside it.
2. Look at the setting that controls whether captured activity becomes a draft entry or is only logged. A draft nobody reviews is invisible time.
3. Check whether lawyers actually switch timers on. A feature enabled for the firm but unused by half the team recovers nothing.
4. Pull a report of captured but unbilled activity for the last four weeks. If it is long, your gap is review discipline, not detection.

Step four may cover most of the problem. If your platform already suggests entries and nobody accepts them, adding a second tool just makes a second pile.

Step 2: decide what may leave your systems

If the platform leaves gaps, you can use a general AI assistant to compare activity records with logged entries. The risk is client data. Calendar titles such as "Call with [client surname] re boundary dispute" are confidential information, and typing them into a consumer chatbot is exactly what the confidentiality tutorial warns against. Use this three-tier rule.

Time capture rarely needs more than green. The question you are asking is "which 20-minute blocks have no entry", and a matter number plus a duration answers it.

Settings to confirm on the AI tool itself

• Business or enterprise terms in force, not a personal account.
• Training on your inputs switched off, and the vendor's current data terms say so.
• Chat history retention set to the shortest option, or off.
• No connected apps (calendar, email, drive) switched on for this task. You are supplying a file, not granting access to the mailbox.
• Only named people on the account, with their sign-ins protected by two-step verification.

Check these against the vendor's own pages today, because terms and setting names change. If a setting is not clearly stated, treat the answer as "no" until you have it in writing.

Step 3: build the redacted export

Export two things for the same date range, for each lawyer separately: the activity feed (calendar, call log, email sent, documents edited) and the time entries recorded.

Then reduce the activity feed to green data. A spreadsheet does this in minutes: delete the subject and title columns, replace the client name with the matter number using a lookup, and keep date, start, duration and type. Run a final search for any surname you know is a current client; it should return nothing.

Date,Lawyer,Matter,Type,Start,Duration_min
2026-10-12,AB,M-1042,Call,09:10,12
2026-10-12,AB,M-1042,Email x3,09:30,9
2026-10-12,AB,M-0988,Document,10:05,25
2026-10-12,AB,(none),Meeting,11:00,60

The "(none)" row is useful: it marks activity that could not be linked to a matter, which is often where time vanishes.

Step 4: ask for gaps, not entries

With the two tables ready, give the AI one narrow job: list blocks of activity that have no logged time against the same matter and day. Do not ask it to write entries.

You are helping a law firm find unrecorded time. Below are two CSV
tables for one lawyer: ACTIVITY (calendar, calls, emails, documents)
and ENTRIES (time already recorded). All client identifiers have been
replaced by matter numbers.

For each day, list activity rows that have no ENTRIES row for the same
matter and date, or where the ENTRIES duration is more than 50% lower
than the activity duration. Output a table with: Date, Matter,
Activity type, Activity minutes, Entered minutes, Gap minutes.
Do not guess what the work was. Do not invent matters. If a row has
no matter, list it separately under "Unmatched".

ACTIVITY:
[paste]

ENTRIES:
[paste]

Here is an illustrative output for one lawyer on one day, with what you would fix in it.

Date        Matter  Type      Activity  Entered  Gap
2026-10-12  M-1042  Call      12        0        12
2026-10-12  M-1042  Email x3  9         6        3
2026-10-12  M-0988  Document  25        15       10
Unmatched: Meeting 11:00, 60 min

What to correct: the three-email row shows a gap of three minutes, which is below your threshold of practical value, so you would drop it. The meeting without a matter is the high-value item, because 60 minutes is real money; the lawyer needs to say which matter it belongs to. A tool cannot decide that.

A four-lawyer litigation boutique's four-week audit

Picture a four-lawyer litigation firm (illustrative) with a blended rate of $300 an hour. The office manager exports four weeks of metadata for two lawyers who say they are "pretty good" at timekeeping.

The AI comparison lists 62 gap rows. Review shows 14 are noise (tiny email bursts, a calendar block that was a lunch). Of the remaining 48, the lawyers confirm 36 as genuine work, totalling 1,080 minutes, or 18 hours. At $300 an hour that is $5,400 across two lawyers in four weeks, about $2,700 each. Eleven rows are unmatched meetings; the lawyers assign nine to matters and decide two were internal.

These are made-up numbers, but the arithmetic is the method: gap minutes found, minus noise, times the rate, then check it against what you could realistically recover. Two cautions. Some of that time will be written down or written off at billing because the matter is fixed-fee or the client disputes it. And the first month always finds the most, because it contains a backlog of habits; later months show the real run rate.

Sorting the gap list into three piles

Hand each lawyer their own list, never a firm-wide league table, and ask them to sort every row into one of three piles within 48 hours while memory is fresh.

1. Bill it. Create the entry now, using the real narrative from the lawyer's own recollection.
2. Not billable. Internal, administrative or covered by a fixed fee. Mark why, so patterns appear.
3. Not mine or not sure. Assign it to someone else or flag it to the supervising partner.

The second pile is useful data too. If one matter keeps showing large non-billable blocks, the fixed fee may be wrong.

More ways the same approach finds money

Travel and waiting time

Filter the calendar export for rows with a location and no entry the same day. A hearing with no time entry for the waiting is the kind of row this finds. Ask the AI to list calendar blocks longer than 90 minutes with no matching entry.

Matters with activity but no time at all

Ask for any matter with five or more activity rows in a fortnight and zero entries. That usually means a file worked on by someone who forgot to start a timer, or a matter opened without billing instructions.

Late-entered time

If your system stores the entry's creation date, compare it with the work date. Entries created more than seven days after the event are the ones most likely to be understated; the gap between event and entry is a proxy for how much detail was lost.

Support-staff time

Paralegals and assistants often work against a lawyer's matter without their own entries. Run the same comparison for them if your billing basis allows it, again by matter number only.

A before-and-after day

Before (reconstructed at 6 pm):
  Matter M-1042 - various - 0.5
  Matter M-0988 - review - 0.2
After (from captured activity, reviewed at 4 pm):
  Matter M-1042 - call 12 min; emails 9 min; hearing prep 25 min - 0.8
  Matter M-0988 - document review 25 min - 0.4
  Matter M-0771 - meeting 60 min - 1.0

The "after" has three matters instead of two, finer durations and a lawyer who can actually remember what the meeting was about, because the review happened the same afternoon.

Three capture routes compared

Before you commit effort, compare the options on the points that matter to a small firm. The effort figures are rough planning estimates for a firm of around five fee earners, not vendor claims.

Start at the top row and move down only when the evidence says to. A dedicated product also means another vendor holding information about your clients' matters, which has to be explained to clients and your insurer.

A four-week pilot you can run without disrupting anyone

Run it on two volunteers. Volunteers talk about it positively afterwards; conscripts do not.

1. Week 1, baseline. Export the last four weeks of activity and entries for both lawyers. Build the redacted file. Do not tell anyone the results yet.
2. Week 2, first comparison. Run the gap prompt, remove noise, and sit with each lawyer for twenty minutes to sort their own list into the three piles.
3. Week 3, change one habit. Pick a single behaviour, such as accepting suggested entries daily at 4 pm, and ask for it. Changing three things at once tells you nothing.
4. Week 4, repeat and compare. Run the same comparison on the new week's data. Look at gap minutes per lawyer per week.

If the second week's gap falls by half or more, habit was the main cause and a reminder routine will hold it. If it barely moves, the capture itself is missing something and you should check the platform's coverage again.

Phone calls, the leak most exports miss

Calls are the hardest time to capture because they happen outside the document system. Three practical fixes cost little. First, check whether your phone system or softphone logs calls by number and whether it can link numbers to matters. Second, add a one-tap "log call" action to the lawyers' mobiles, such as a shortcut that opens a pre-filled time entry. Third, set a rule that any call over five minutes gets an entry before the next call begins.

For the export, a call log contributes only the lawyer, number's matter link, start time and duration. If the number is not linked to a matter, it falls into the "Unmatched" list, which is correct: a human has to say what it was.

Fixed-fee and capped matters: capture still matters

A lawyer on a fixed-fee conveyancing or probate matter may think timekeeping is pointless. It is not. Without recorded time you cannot tell whether the fee was right. Run the same comparison for fixed-fee matters and total the minutes per matter after the file closes. If a typical fixed-fee matter takes 11 hours against a fee priced on 8, you have found a pricing problem that no amount of better billing narratives will fix.

Illustrative figures: a firm prices a standard residential lease extension at $2,400, assuming 8 hours at $300. Capturing activity across ten closed files shows an average of 10.5 hours. The fee is underpriced by 2.5 hours a file, or $750, so either the fee goes up to around $3,150, or the process gets tightened. That is a decision you can only make with captured time.

A note to send the team before you start

Transparency prevents most of the resentment. Adapt this note (illustrative wording):

Subject: A four-week look at unrecorded time

We are testing a way to spot time that was worked but not logged.
Two volunteers are taking part. The test uses matter numbers and
durations only, no client names, and each person sees only their own
results. It will not be used to rank or assess anyone. After four
weeks we will decide together whether to keep it. Questions to [name].

The monthly routine for a two-lawyer firm

A two-lawyer boutique has no office manager, so the owner does the export. The routine shrinks to one hour on the first Friday of the month. The owner exports activity and entries for the previous four weeks, strips subject lines with a spreadsheet filter, runs the prompt and reviews the output with each lawyer for ten minutes. The AI subscription costs a typical business-plan price per seat; confirm the current price on the vendor's page before budgeting. Against the hour of time, a recovered ten hours a month at the firm's rate makes the cost easy to justify, but only if the entries are real and billable.

If the firm already pays for a legal platform with AI features, compare the cost of the add-on with this manual route. Per-seat pricing is the heart of legal AI pricing for small firms, and for the choice between a general assistant and a legal-specific tool see ChatGPT or a legal AI tool.

Five ways the gap report misleads you

These are the realistic failures, and how each one shows up.

• Names leak into the export. A matter description column is left in. Check the file with a find for three or four current client surnames before every upload.
• The AI invents a matter. If a row says "Matter M-1099" and no such file exists, your prompt allowed guessing. Add "use only matter numbers present in the data" and verify every matter against your system.
• Double counting. A call logged in the softphone and in the calendar appears twice. Ask for overlaps in start times and drop the duplicate.
• Capturing time a client will not pay for. Finding it is not the same as billing it. Check engagement terms, caps and budgets before the entry becomes an invoice line.
• Surveillance creep. If staff think the output ranks them, they will game it. Keep it personal and coaching-only, and say so in writing.

Checking the pilot recovered billable hours

After one month, compare three numbers for the pilot lawyers against the month before: hours recorded, hours billed after write-offs, and the share of entries created the same day. A genuine improvement shows in the second and third, not only the first. If recorded hours rose and billed hours did not, you captured time that cannot be charged.

Repeat the exercise quarterly for two lawyers rather than building a permanent pipeline. If the gaps shrink, the problem was habit and the platform's own capture will carry the load. If they stay large, look at tools with deeper capture and judge them with the questions in what to automate first in a small law firm.

If you would rather map this against your own systems, that is what my AI implementation consultation covers: which feature to switch on, what to redact and what to leave manual. Keep the stricter risk tiers from five legal workflows ranked by risk beside the export rules above.

Questions about AI time capture in a law firm

Should AI write the time entries for us?

Not as part of this job. Capturing that something happened is one task; writing a client-ready narrative is another, with its own checks. Get the capture right first, then pass the findings to whoever writes narratives.

Is it safe to export a calendar to a chatbot?

Only after you strip it. Calendar titles often contain client names and matter facts. Replace them with matter numbers and durations before anything leaves your practice system, and use a tool whose business terms bar training on your data.

How much time do firms usually miss?

There is no reliable figure to quote, because it depends on the firm. Measure your own: compare four weeks of calendar and email activity with logged entries for two lawyers, and use that gap, not an industry claim.

Will fee earners resent it?

They will if it looks like surveillance. Frame it as a prompt that suggests entries for the lawyer to accept or reject, show them only their own data, and never use it to rank people.

Further reads

• AI Time Capture for Solicitors: Recover Hours You Never Billed — A broader look at recovering hours lawyers never billed.
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Intake calls are another place time goes unlogged.
• How to Summarise Bundles and Transcripts With AI Safely — Safe handling of large documents once time is captured.
• Can Lawyers Use AI Note Takers in Client Meetings? — Meeting records raise the same confidentiality questions.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Drafting is the next job to speed up after capture.
• WIP Report Template for Service Firms: Build One From Timesheets — A WIP report for a service firm built from a timesheet export: ageing buckets, write-offs, roll-forward and lock-up days, with sample figures.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Smokeball AutoTime feature page; Clio legal time and expense tracking page; vendor data terms for the AI tool you choose.