Senior time went into finding the passage, not arguing it.

The issue

The firm's starting point, on Microsoft 365 Copilot and Copilot Studio, for its Innovation Lead and Head of Knowledge Management:

• 59% of deposition prep spent reading transcripts before knowing what mattered.
• $1.8M of research and review time written off as non-billable last year.
• 9 days from a research request going in to an answer coming back.

Associates read deposition transcripts end to end before knowing which parts mattered, and answered research questions with no way of finding out whether the firm had already answered them.

The obvious fix is one searchable index across every matter. That was the one fix this firm could not make. Two vendor pilots had already been stopped for it: both indexed everything and filtered the results afterwards, which means the model has already read the confidential file.

Issue analysis

We coded timesheet narratives from 214 depositions and 380 research memos by activity, rather than relying on self-reported time. A typical 40-hour deposition-prep block broke down like this:

Only 12% of a prep block reached analysis, and 33% of memo hours repeated earlier firm work.

Solution suggested

Not one index for the firm. An index per lawyer. The agent inherits the matter permissions the person already holds, and the retrieval scope is built from them before any query runs. A file the lawyer cannot open is one the model never reads, never embeds and cannot quote.

The lawyer asks, signed in as themselves; their rights define what is searchable; and the client wall is enforced before retrieval, not after. Every statement cites page and line, research answers come from the firm's own precedent, and prompt, sources and permissions are logged per answer.

Solution implementation

Eighteen weeks to a firm-wide rollout:

1. Weeks 1–3 · Permissions mirrored: matter access rebuilt as per-user retrieval scopes, reviewed with the General Counsel.
2. Weeks 4–8 · Transcript pipeline: parsed to page and line, benchmarked blind against 40 human summaries.
3. Weeks 9–14 · Research agent: firm precedent indexed per scope, with answers drawn only from it.
4. Weeks 15–18 · Pilot, then firm-wide: two practice groups, the audit log fed into KM reporting, then all four offices.

Outcome and impact

Two quarters in, a deposition-prep block went from 40.0 hours to 16.8, with reading down from 23.6 hours to 5.4 and analysis up from 4.8 hours to 9.6. The week halves. Judgement doubles.

• ↓58% deposition prep: 40.0 → 16.8 hours.
• ↓61% research hours: 18 → 7 hours per memo.
• 2× analysis time: 4.8 → 9.6 hours per prep.
• Cited and scoped, secure by design: no answer crosses a client wall.

The reading is done before the lawyer opens the file. If a confidentiality review has already stopped an AI pilot, the conversation to have is about the architecture before the features.

Scenario reference: Microsoft's Copilot scenario library for Legal catalogues the patterns this builds on: case and precedent analysis, and legal guidance from existing matter context. Deposition summarisation extends that pattern rather than appearing as a listed scenario.

An illustrative engagement. The scenario and figures are representative, not the audited results of a single named client.