What this automation does
Most unbilled time is not lost because people work for free. It is lost because the 11-minute call at 4:40 p.m. never makes it onto the timesheet. This workflow rebuilds the day from the systems that already recorded it.
- Collect the day’s activity. Calendar events, Teams or Zoom meetings with their actual duration, sent emails and replies, phone calls from the firm’s phone system, and documents opened or edited in SharePoint or your document management system.
- Group related activity. Twelve emails and a document edit on the same lease amendment become one work session, not thirteen entries.
- Match to a client and matter. Matter numbers, folder paths, email domains and contact lists from the billing system do most of the matching. The model reads content only to break ties.
- Draft the narrative and duration. The model writes a task-level narrative in your firm’s style (“Review and revise draft lease amendment; correspondence with opposing counsel re: assignment clause”) and suggests a duration rounded to the tenth of an hour.
- Queue for review. Each timekeeper gets a list of drafts each evening or the next morning. They accept, edit, merge, split or discard.
- Post approved entries. Approved entries are written to the billing or practice management system through its API, with the source activity kept for audit.
When it is worth building
This is worth building when partners believe meaningful time is going unrecorded and associates or staff reconstruct their week on Friday afternoon. Firms that bill by the hour feel this directly. A firm on flat fees or retainers still benefits from better effort data for pricing, but the return is less obvious.
It is not worth building if your timekeepers already run a timer they trust, or if a commercial time capture tool already reads the systems you use. It is also a poor fit if your activity lives mostly in tools with no API access, such as a desktop-only document system that cannot export audit logs.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Activity sources | Microsoft Graph for Outlook, calendar and Teams, SharePoint audit data, phone system call logs | Most firm activity already lives in Microsoft 365 with good API coverage |
| Practice management | Clio, or your billing system’s API | Source of clients, matters, rates and task codes, and destination for approved entries |
| Activity store | PostgreSQL in your cloud tenant | Holds a rolling window of activity metadata, matched matters and drafts |
| Orchestration | n8n or a small scheduled service | Nightly collection runs, retries and per-timekeeper batching |
| Drafting model | Claude or OpenAI’s GPT models through the API | Groups activity and writes narratives that follow firm and client billing rules |
| Review interface | A simple web view or Teams adaptive card | Timekeepers approve drafts where they already work |
Clio publishes a documented API that registered apps can use with OAuth. Other legal and accounting billing systems range from open APIs to partner-only access or import files. That difference can add weeks, so confirm it during scoping.
How to build it
- Agree the rules with billing partners first. Minimum increment, rounding, what counts as billable, banned narrative phrases, and which clients require UTBMS task and activity codes or LEDES output. Write these down; they become prompts and code checks.
- Connect activity sources with least privilege. Use application permissions scoped to the timekeepers in the pilot. Collect metadata (time, participants, subject, file path) by default and pull message bodies only when matching needs them.
- Build the matter matcher. Start with deterministic rules: matter numbers in subjects and paths, client domains, contacts on the matter. Measure how much this alone covers before adding the model.
- Write the drafting prompt with examples. Give the model ten or twenty real, approved narratives from your firm as style examples, plus the client’s billing guidelines when they exist. Ask for structured output with matter ID, task code, activity code, duration and narrative.
- Validate in code. Reject block-billed narratives with several tasks under one duration, durations longer than the underlying activity, entries on closed matters, and narratives containing banned phrases.
- Ship the review queue. Make accept and edit one click each. If review takes longer than writing time from scratch, nobody will use it.
- Pilot with five to ten timekeepers for a month. Compare recorded hours and write-downs against the same people’s prior months before rolling out.
What breaks in production
- Personal and administrative email. Lunch plans and firm-wide announcements must be filtered out before drafting, or the queue fills with noise.
- Shared matters across many clients. Insurance defense panels and corporate groups often share contacts across matters. Domain matching fails, so the model needs matter descriptions to choose.
- Over-generous durations. A meeting scheduled for an hour that ended after 20 minutes should not draft as 1.0. Use actual join and leave times where the meeting platform provides them.
- Narratives that reveal privileged strategy. Bills are sometimes produced to third parties. Keep narratives descriptive and neutral, and test the prompt against that rule.
- Trust erosion. If the first week of drafts is poor, timekeepers stop opening the queue. Run in silent mode first and only show drafts once matching accuracy is acceptable.
Timeline and cost
A pilot covering Microsoft 365 activity and one billing system typically takes four to eight weeks, including the silent-mode period. Lumeor builds this in the range of $12,000 to $35,000, depending on the number of activity sources, how open your billing system’s API is, and how many client-specific billing rule sets you need to enforce. If meeting notes are part of the picture, our guide to automating meeting summaries and action items covers that piece, and the cost of adding AI to existing software explains what drives the range.
How to measure success
- Recorded hours per timekeeper per week, compared with the same period last year
- Share of drafts accepted with no edits or light edits
- Write-downs and client billing disputes on matters using drafted entries
- Days from work performed to time entered
- Timekeeper minutes spent on time entry per week, measured by a short survey before and after