Automatic time capture & billing-narrative generation for professional services

Under the billable-hour model, time that is not recorded is revenue lost, and narratives that violate client guidelines are cut or rejected. AI watches the digital trail of work (emails, documents, calls, calendar, apps, site visits), assigns it to the right client and matter, and drafts the entry with a compliant narrative. The professional reviews and approves in minutes instead of reconstructing the week from memory.

Problem & value

• Today: lawyers, tax advisers, auditors, consultants and agencies log time with timers or from memory at day or week end. Small items (a short call, a quick email) are lost; narratives are terse or inconsistent; billing staff and partners rewrite them before invoicing. Corporate clients apply outside-counsel guidelines (OCGs) and reject or reduce invoices with block billing, vague descriptions or wrong codes [1].
• AI value: more complete capture, fewer late entries, consistent narratives, fewer invoice write-downs and rejections, less non-billable admin for fee earners and billing teams.
• Evidence level: vendors (e.g. Smokeball [3]) market recovery of additional billable time, with figures often quoted in the 10-30% range that we could not verify; treat these as vendor figures, not independent benchmarks. Run your own pilot and compare approved hours per person before and after.
• Reality check: the AI drafts, the professional certifies. Time records are a statement to the client; an unreviewed AI narrative is a professional-conduct risk, not just a data-quality one.

Variants

• Law firm (core): capture across Outlook, DMS (iManage, NetDocuments), Teams/Zoom, phone and drafting tools; map to client-matter numbers; UTBMS codes and LEDES 1998B/XML 2.x export for corporate clients [1].
• Tax advisers, auditors, consultancies: engagement/phase based capture, budget-versus-actual alerts, DATEV or ERP/PSA integration; narratives less code-driven.
• Freelancers, agencies and Handwerk (trades): timesheet generation from calendar, repositories, ticket tools and GPS/geofenced site time on mobile; output is a customer timesheet or Stundenzettel feeding invoicing.
• Client-side (legal ops): the mirror image: AI reviews incoming legal invoices against OCGs (rate cards, caps, block billing, staffing) and flags exceptions; see DOC-001 for the invoice-intake pattern.
• Inside AI tools: time entries generated automatically from tasks done in a legal AI assistant (Laurel and Legora announced such an integration) [2].

How it works

activity sources (email, calendar, DMS events, calls/meetings, apps, GPS)
  -> normalise to activity events (who, when, duration, object, snippet or metadata)
  -> matter classification (rules + embeddings + LLM over history; confidence score)
  -> grouping into entries (merge fragments, rounding rules, minimum increments)
  -> narrative drafting (LLM, firm style + client guideline + UTBMS code selection)
  -> compliance check (OCG rules: block billing, vague verbs, caps, rate, prohibited tasks)
  -> professional review (accept / edit / reject; low-confidence items shown first)
  -> export to billing/PSA (Clio, Aderant, Elite, 3E, Intapp, DATEV) -> invoice / LEDES file

• Classification: an assignment problem with strong priors (open matters, recent contacts, document paths, subject lines). A small classifier or embedding retrieval plus an LLM tie-breaker works; unknown activity goes to an "unassigned" inbox rather than a guess.
• Narrative drafting: the LLM turns activity metadata into a one-line description ("Draft and revise share purchase agreement, section 8 warranties; review client comments"). Prefer metadata and short excerpts over full document content (data minimisation).
• Compliance layer: deterministic rules per client (maximum entry length, forbidden words, required task codes) executed after the LLM, because rules must be reliable, not probabilistic.
• Human-in-the-loop: daily or weekly review prompt; nothing reaches an invoice without approval. Corrections feed few-shot examples and rules.

Data preparation

The system needs activity signals (calendar, e-mail, documents, phone and application usage, often from Microsoft 365 or Google Workspace), the list of clients and matters from the practice or billing system, past timesheet entries and billing narratives, and the billing guidelines of key clients (often PDFs or e-mails). Gaps are typical: matter numbers are inconsistent, closed matters stay open, old narratives vary in style, and client guidelines exist only as scattered documents.

Preparation steps: - Clean the client and matter list (close dead matters, unique numbers, responsible persons), see PREP-005. - Export 6-12 months of approved time entries as labelled examples for matter assignment, see PREP-020. - Collect good narratives and client guidelines as templates and rules, see PREP-016 and PREP-014. - Check access rights and decide which signals are captured at all (no private content, no content of privileged mails), see PREP-024 and PREP-025. - Build a small test set to check assignment accuracy, see PREP-018.

Tools: the billing or practice system API, Microsoft Graph or Google APIs, a spreadsheet for rules. Rough share of effort: about 25-35 percent, mostly matter-list cleanup and the decision on which signals to capture. Use GUIDE-003 as a checklist.

Implementation paths

Which path for whom

• SMB (A): Mature time-capture products with EU hosting exist at per-user prices and the process (matters, clients, timesheets) is standard.
• Enterprise (B): Billing guidelines of large clients, matter systems and billing integration are specific, and client confidentiality favours a controlled build in an EU region.
• Developer (C): Open-weight small models and open-source tooling on HW2 keep activity data local and allow a tailored prototype.

Path A — Buy (recommended)

• Law-specific AI time capture: Laurel (US, widely deployed in Canadian, UK and US firms), Intapp Time (automatic capture of meetings, emails and document editing plus timers and voice input) [4], Clio Timekeeping and Clio Duo suggestions in Clio Manage [5], Smokeball AutoTime (nightly entries from in-platform activity) [3]. Third-party add-ons: Billables AI, MIRA (Clio ecosystem).
• PSA/ERP-integrated: Microsoft Dynamics 365 Project Operations, Workday, SAP, Aderant and Elite 3E (UK/US legal ERP) with their AI add-ons.
• Freelancers/agencies/trades: Timely (automatic memory tracker), Toggl Track, Clockify; for trades, Craftboxx, Streit/Plancraft-type field apps with GPS time (check EU hosting and works-council fit). German time-recording tools often offer EU hosting; verify per vendor.
• Choose A when: the firm already runs a supported practice-management or billing system. Check EU data residency, subprocessor list, whether content or only metadata leaves the device, and LEDES export.

Path B — Build on cloud APIs

• Microsoft Graph (mail, calendar, Teams) and Google Workspace APIs as sources; EU-region LLM endpoints (Azure OpenAI in EU Data Zone, Mistral on EU infrastructure, Claude or Gemini via EU-region cloud offerings) for narratives; structured output with JSON schema for entry objects.
• Orchestration with n8n or a small service (AGENT-002); billing system via REST API (e.g. Clio Manage API). Add an eval set of historic approved entries to measure narrative acceptance rate.
• Choose B when: a boutique or mid-sized firm needs custom matter logic or a niche billing tool, and has developers.

Path C — Self-host (open source)

• Open-weight LLMs in the 8-32B class (Qwen3, Mistral Small, Llama) served with vLLM or Ollama; embeddings with bge-m3 or multilingual-e5; Whisper-class ASR for call notes (SPEECH-001). Capture agent as a local desktop/Outlook add-in so raw content never leaves the firm.
• Hardware: HW2 is sufficient for a firm with hundreds of users because workloads are short and batched nightly.
• Choose C when: professional secrecy (Berufsgeheimnis, §203 StGB), mandate confidentiality or works-council demands rule out SaaS.

Path D — Custom training

Rarely needed. Fine-tune a small model on 1-3 years of approved narratives per firm for style fidelity and cost, or train a matter classifier on labelled history. Needs a few thousand approved entries and consent/anonymisation of client data.

State of the art (2026)

• Passive capture has moved from niche to default feature of legal practice software; leading products combine event capture with LLM narratives and a compliance check against client guidelines [2][3][4].
• Newer: time entries produced directly from agentic legal-AI work (task completed -> entry), and capture of AI-assisted work, which raises the open question of how to bill work that took minutes with AI [2].
• Limits: attribution errors on multi-matter days; phone, in-person and thinking time stay invisible; narratives can sound generic or invent detail not in the metadata; code selection (UTBMS) is error-prone for unusual tasks; some practitioners report only part of the suggestions are usable unedited (anecdotal, unverified). Partner adoption depends on trust, so show the source behind each entry.

Shared resources & combinations

• Same stack: email/calendar connectors and classification with AGENT-001 (inbox triage) and AGENT-012 (scheduling); ASR and summaries from SPEECH-001; CRM/matter master data with AGENT-011.
• Downstream: invoice generation, collections and write-off analysis (AGENT-023); controls monitoring of billing policy (AGENT-018); client-side invoice review reuses DOC-001 extraction.
• One local LLM server (HW2) can serve narratives, triage and summaries together.

Effort, cost & KPIs

• Path A: weeks (pilot with 10-20 timekeepers, then rollout); per-user subscription; main cost is change management.
• Path B: 2-4 months incl. billing integration and eval set. Path C: 3-5 months plus operations.
• KPIs: recorded hours per timekeeper, share of AI drafts accepted unedited, time-to-entry (lag between work and entry), write-down and invoice-rejection rate, realisation rate, review minutes per week, share of unassigned activity.

People & change

Involve management (billing policy), the department lead (partners, billing manager), end users (timekeepers) and IT early. The data protection officer must be involved before the pilot, because activity data is personal data and a DPIA is likely (systematic evaluation of people, Art. 35 GDPR); sign an AVV with every processor (GUIDE-005).

Works council: yes, if one exists. Passive activity capture is objectively suited to monitor employees' behaviour and performance, so §87(1) no. 6 BetrVG co-determination applies; §90(1) no. 3 requires early information, and §80(3) allows an AI expert. Agree a works agreement that excludes performance evaluation of individuals and limits retention (GUIDE-004).

Support AI literacy (Art. 4 AI Act) with a short session on what is captured and how suggestions are corrected (GUIDE-006). Start with 10-20 volunteer timekeepers as pilot group and AI champions (GUIDE-008, GUIDE-009). Typical concern: "I am being watched". Address it with transparency, data visible only to the person, no use for performance review, and a pause switch. Humans always approve entries before billing.

Risks, regulation & compliance

• EU AI Act: typically minimal risk (Art. 4 AI literacy applies). Caution: if the same activity data is used to evaluate, rank or discipline employees or to allocate tasks based on behaviour, Annex III point 4 (employment, worker management and monitoring) can make it high-risk [6]. Art. 5(1)(f) prohibits inferring emotions of workers; do not add sentiment or "engagement" analysis. Keep the tool scoped to billing support.
• GDPR: activity logs are employee personal data and client communication content. Needs a legal basis (Art. 6; in Germany §26 BDSG and works agreement), transparency, data minimisation (metadata first), DPIA likely for continuous capture, retention limits and processor contracts with EU transfer safeguards [7]. In Germany a works council has co-determination on technical monitoring (§87 Abs. 1 Nr. 6 BetrVG). GPS time capture of field staff must be limited to working hours and job sites.
• Professional secrecy and client confidentiality: lawyers' duties (e.g. BRAO §43a, §203 StGB) restrict sending mandate content to third parties; prefer metadata-only capture, EU processing and subcontractor commitments. Some clients prohibit AI use on their matters; honour matter-level opt-outs.
• Billing integrity: inflated or fabricated time is an ethics and fraud risk; keep audit trail of draft vs approved entries and never auto-submit.
• Working-time law: timesheets for billing are not the legally required working-time record; EU case law requires employers to record working time (CJEU C-55/18) so keep the purposes separate and documented [8].

Getting started

1. Pick 10-20 volunteers; measure baseline: hours recorded per week, entry lag, write-downs, rejected invoices.
2. Agree scope with the data protection officer and works council: sources, metadata-only, retention, no performance ranking.
3. Pilot one product from Path A with matter list and two client guidelines encoded as rules.
4. Review daily for two weeks; log accept/edit/reject reasons and tune rules and style examples.
5. Decide rollout using realisation, acceptance rate and partner feedback; add client-side invoice review later.

Sources

1. Onit — Using AI to solve UTBMS issues for law firms and legal departments — https://onit.com/blog/ai-solve-utbms-ledes-outside-counsel-corporate-legal-departments
2. City AM — Laurel and Legora partnership turning legal AI tasks into time entries — https://www.cityam.com/?p=2465400
3. Smokeball — Getting started with AutoTime — https://support.smokeball.com/hc/en-gb/articles/5860708227863
4. Intapp — Intapp Time product page — https://www.intapp.com/time/
5. Clio Help Center — Get started with Clio Timekeeping — https://help.clio.com/hc/articles/50852535786139-Clio-Timekeeping
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
8. CJEU, Case C-55/18 (CCOO), working-time recording — https://curia.europa.eu/juris/liste.jsf?num=C-55/18
9. American Bar Association — AI time tracking for lawyers — https://www.americanbar.org/groups/law_practice/resources/law-technology-today/2025/ai-time-tracking-for-lawyers/
10. MLT Aikins — Launches Laurel as its AI timekeeping platform — https://www.mltaikins.com/news/mlt-aikins-launches-laurel-as-its-ai-timekeeping-platform/