Attorney Timekeeper
Voice-captured billable time for law firms. An attorney speaks an entry from a phone; AI parses it into matter, duration, description, and amount, and writes it to a ledger the firm owns outright.
01Overview
A time-capture system for solo and small law firms, built around the reality that attorneys do not lose billable time because they cannot do arithmetic — they lose it because entry happens hours later, from memory. The system moves capture to the moment the work ends, using voice, and does the structuring afterwards.
02Problem
Billable time reconstructed at the end of the day is systematically undercounted, and the shortfall is invisible because there is nothing to compare it against. Existing legal time software is a serious purchase, a migration, and a new place for client data to live — which is precisely the objection that stops small firms adopting it. The barrier is not capability; it is friction at entry and reluctance to hand client data to another vendor.
03David's role
- Designed the system around the firm's existing productivity suite instead of introducing a new platform.
- Built the capture path, the parsing layer, the ledger structure, the calculation logic, and the audit log.
- Specified the data-ownership model so the firm's data never leaves infrastructure it already controls.
- Wrote the deployment, administration, security, and first-use documentation for non-technical staff.
- Ran the product demonstration directly with practising attorneys and shaped v1 scope from that feedback.
04Architecture
Deliberately thin. A form captures voice or typed input; a script layer parses and calculates; a spreadsheet owned by the firm is the ledger and the interface. No servers, no database, no vendor account, no subscription beyond the workspace licence the firm already pays for.
05How it works
- 01Capture
- 02Submit
- 03Parse
- 04Calculate
- 05Review
- 06Account
- 01
Capture
The attorney opens a form on a phone or computer and either dictates the entry or types it — the interaction is a few seconds long, at the moment the work ends.
- 02
Submit
The entry is written immediately to the firm's own spreadsheet. Capture never depends on the AI being available.
- 03
Parse
The model reads the natural-language entry and fills in matter, duration, description, and billable amount.
- 04
Calculate
The billable amount is computed from duration and the applicable rate.
- 05
Review
The attorney reviews, corrects, or annotates directly in the sheet — a familiar interface requiring no training.
- 06
Account
An audit log records entries and changes for billing review.
06AI & technology
- Google Gemini 2.5 Flash-Lite for natural-language parsing of spoken and typed entries
- Native device dictation for voice input, so no separate transcription service is involved
- Structured extraction into typed ledger fields with deterministic amount calculation
- AI parsing is optional and can be disabled with a single configuration change
07Integrations
- Google Forms for capture; Google Sheets as ledger, interface, and system of record
- Apps Script for parsing, calculation, and audit logging
- Google AI Studio API key held by the firm
- Operates entirely within an existing Google Workspace licence
08Human control & governance
- The firm owns the spreadsheet outright and can open, export, or delete its data at any time.
- No client data is stored on external servers; nothing is held by a vendor.
- AI output is a draft the attorney reviews — the model fills fields, the attorney remains accountable for the bill.
- Every entry and change is recorded in an audit log for billing review.
- AI parsing can be turned off entirely, leaving a fully functional manual system.
09Key design decisions
The decision, the alternative it was chosen over, and why.
- 01
Decision
Build inside Google Workspace instead of as a standalone SaaS product.
Instead of
A hosted application with its own database, accounts, and subscription.
Why
The blocking objection for small firms is not capability, it is handing client data to another vendor and learning another tool. Building where the data already lives removes the security conversation, the migration, and the recurring cost in one decision — and it means the interface is a spreadsheet every attorney already knows.
- 02
Decision
Separate capture from parsing, so submission never depends on the model.
Instead of
Parsing the entry inline and writing the structured result.
Why
The system's entire value is that entry happens in the moment. If an API outage or a rate limit can block submission, the habit breaks and the tool is abandoned. The raw entry is written first and enriched after; a parsing failure degrades to a manual entry rather than losing the time.
- 03
Decision
Make AI parsing optional and disableable by configuration.
Instead of
Making AI parsing integral, since it is the product's headline feature.
Why
Some firms will not accept AI touching client matter descriptions, and that is a legitimate position rather than an objection to overcome. A single switch turns it into a fast manual system instead of losing the client — and it forced the underlying ledger to be sound on its own.
- 04
Decision
Choose the smallest, fastest model rather than the most capable one.
Instead of
A frontier model for higher parsing accuracy on ambiguous entries.
Why
The task is short, highly structured extraction with a human reviewing every result. A larger model would add latency and cost for accuracy that the review step already provides — and negligible running cost is a real feature when the buyer is a solo practitioner deciding whether this is worth it.
10Tradeoffs
- Building on a spreadsheet gives instant familiarity and firm ownership, and accepts the scaling ceiling that comes with it — this is a solo-and-small-firm system by construction.
- Multi-attorney support is scaffolded but not complete; v1 is honestly a single-attorney product and is described that way.
- The Microsoft edition is planned; only the Google Workspace edition is built and documented.
- Deferring parsing means an entry can sit briefly unstructured — the correct trade for never blocking capture.
- Firm-owned data means firm-owned backup responsibility, which the administration documentation has to cover explicitly.
11Outcome
Version 1.0 of the Google Workspace edition is complete and documented for non-technical deployment, with setup guides, an administration guide, a security guide, a quick-start checklist, and a versioning policy. It was demonstrated to practising attorneys and shaped by that feedback. Multi-attorney support is scaffolded for v2; the Microsoft edition is in development. Client deployment outcomes are not claimed here, because they are not verified.
12Evidence
- Shipped package
- Version 1.0.0 with forms, sheet structure, Apps Script automation, documentation set, and a versioning policy.
- Documented model choice
- Gemini 2.5 Flash-Lite, selected for latency and cost, with running cost characterised for a typical small firm and an explicit disable path.
- Scope honesty
- The feature table marks multi-attorney support as scaffolded for v2 rather than presenting it as included.
- Client validation
- Demonstrated to practising attorneys, with v1 scope shaped by that feedback.