Time capture and billing narratives: AI for the billable hour

Most fee earners rebuild at least part of their week from memory. A phone call runs long between two meetings, or a contract mark-up gets read on the train. By Friday the time sheet is pieced together from the calendar and recollection, and some of the work never makes it onto a bill.

This post in our Practical AI for Law Firms series covers using AI to draft time entries and billing narratives from activity the firm already records. It also looks at the practice-management side, where the same approach helps with work in progress (WIP) and debtor follow-up. The design rule throughout is simple: the system drafts and the fee earner decides. Nothing reaches the practice management system until a person has read the entry and accepted it.

Where the drafts come from

The safest sources are records the firm already holds under its existing policies:

• Calendar events from Microsoft 365 or Google Workspace, with attendees and duration
• Sent and received email, matched to a matter by client domain, a matter number in the subject line, or the folder it was filed to
• Document management activity: documents opened, edited and saved, with version history
• Call logs from the phone system, where the firm has them

Each source produces an activity record with a time, a duration where one exists, and a candidate matter. The language model writes the narrative. It does not set the time. A 45-minute calendar event is recorded as 45 minutes, and an editing session runs from when the document was opened to when it was saved. The fee earner can change either.

Whole-of-desktop capture tools are also appearing. Threshyr runs offline and uses on-device AI to classify app and URL activity against projects and clients. Breadcrumb records screens and meetings on a Mac and keeps the history encrypted locally. For a law firm, though, recording a lawyer’s screen means recording privileged client material. In NSW, the Workplace Surveillance Act 2005 also requires notice to staff before computer surveillance starts. We would begin with system records and leave screen capture for a separate decision with its own privacy assessment.

Every line traces back to a record

The most useful pattern we have seen for this comes from an unrelated field. The open-source ifc-lite project changed its AI report drafts so that they are built from evidence snapshots. Each claim is marked supported, unverifiable or contradicted, and a contradicted claim blocks saving until the reviewer edits or removes it. The same PR records whether each block of text was generated by AI, edited by a person or written by a person.

Applied to time recording, a drafted entry looks something like this:

{
  "matter": "2026-0412",
  "date": "2026-10-13",
  "units": 8,
  "narrative": "Conference with client regarding draft lease; review landlord's proposed amendments to clause 14",
  "evidence": [
    { "type": "calendar", "id": "evt_8f2c", "minutes": 30 },
    { "type": "document", "id": "DMS-118204 v3", "minutes": 18 }
  ],
  "unsupported_phrases": [],
  "provenance": "ai_draft",
  "status": "awaiting_review"
}

A validation step checks each phrase of the narrative against the evidence list. If the narrative says “review amendments to clause 14”, there must be a document record behind it. Phrases with no record behind them are flagged before the fee earner sees the entry.

The ifc-lite authors state the limit of this kind of check plainly: “supported” means the cited values match, not that the sentence is true. The same applies here. The check confirms a record exists. It cannot tell whether the narrative describes the work fairly. That judgement belongs to the fee earner.

Review is part of the design

AI output varies between runs and can go beyond what it was given. One developer gave an AI coding agent the same bug 11 times, and 6 of the 11 resulting commits contained files the agent had never declared or mentioned. In billing, the same failure looks like a narrative that reads well but describes something that never happened, such as “considered advice” when the lawyer only opened the file.

So the review screen is the product. Entries are grouped by matter with the evidence shown beside each line. The fee earner accepts, edits or rejects each one, and edits are recorded against the AI draft. The integration to the practice management system only accepts entries that carry a recorded acceptance from the responsible fee earner. Automatic submission is left out of the build entirely, so nobody can switch it on later.

Narratives must describe work done

Under the Legal Profession Uniform Law, which applies in NSW, Victoria and Western Australia, legal costs must be fair and reasonable, and a client can ask for an itemised bill. The narrative is what the client reads, and in a dispute it is what a costs assessor reads. A model that writes persuasive narratives creates risk for the firm.

The rules we would build in:

• The prompt tells the model to describe the activity, not its quality. Phrases like “careful consideration” and “extensive review” are blocked.
• The model cannot change durations or units. Only the fee earner can.
• Activity that does not match a matter goes to an unallocated queue. The system never guesses a matter and bills it.
• Client billing guidelines, such as insurer panel rules against block billing, sit in configuration and are checked before an entry is shown for review.

Write-downs and write-offs remain partner decisions. The system can show what was recorded against a matter, but it does not recommend what to charge.

WIP and debtor follow-up

The same evidence-first approach works on the practice-management side. For WIP, a weekly summary for each partner lists unbilled time by matter and age, set against the estimate given in the costs disclosure. Firms must tell clients when an estimate changes significantly, so a matter approaching its estimate gets flagged early, and the partner decides whether to send updated disclosure.

For debtors, the model drafts reminder emails based on invoice age, the client’s payment history and earlier correspondence on the matter. The credit controller reviews and sends them. Escalation, whether that means a firmer letter, stopping work or anything involving a lien, stays with a partner, and the drafts never include threats or consequences the firm has not decided on.

Costs and limitations

Most of the cost is integration work. Practice management systems differ widely in their APIs, and some older on-premises products expose very little. Matter matching determines whether the drafts are good enough to use, so plan for a tuning period with real data.

Model costs per entry are small next to the integration effort. For data residency, models hosted in Australian regions, such as Azure’s Australia East or AWS Sydney, keep matter data onshore.

The system will miss work that leaves no record, such as calls on a personal mobile, corridor conversations and thinking time. Fee earners still add those by hand. Adoption depends on the review taking less time than writing entries from scratch. If it takes longer, people stop using it, and you should treat that as a failed pilot.

A sensible first project

Start with one practice group and only two sources, calendar and email. Drafts go into a review list for a month, and you compare them with what the group recorded manually. That comparison tells you how much unrecorded time exists and how well matter matching works. It also shows whether the review is quick enough before you connect the document management system.

PicNet builds production AI systems for Australian organisations. If drafted time entries with evidence behind every line sound useful for your firm, talk to us about what a first project could look like.