← Insights home
Webinar recap · On demand

How to build a legal agent
(ft. Claude Cowork)

Missed it live, or want to go back over the setup? This is the recap of the session where Paul Lacey built the foundations on screen: the shared brain on Claude Cowork, the scheduled tasks that keep it alive, what changes when an agent owns the outcome, and the honest conversation about trust and supervision with Lorna Khemraz and Lili Breidenbach. The full recording and the slide deck are both here.

Paul Lacey VP Product, Flank · with Jake, Lorna & Lili
Format Live build + conversation
Recorded Wed 8 July 2026
Read time ~7 minutes
The short version Watch Part 1 · Build Part 2 · Outcomes Part 3 · Trust Takeaways
In one minute

The short version

Individual lawyers are getting faster with AI everywhere. The question Jake opened with is the one that actually matters: has the throughput of your legal function changed, or are individuals just quicker at their own tasks? That gap is what the whole session was about.

Paul answered it in three moves. First, the foundation: a shared, layered knowledge base (the "brain") that every Claude Cowork seat in the company plugs into, kept alive by scheduled tasks rather than left to rot like most knowledge bases. Second, the work you don't want done faster but gone entirely: purpose-built Flank agents that pick up requests where they already arrive and own the outcome. And third, with Lorna and Lili, the honest part: what it actually takes to trust an agent's work, and why the supervision burden falls over time instead of staying flat.

Read
This recap
The highlights of the live build, the supervision conversation, and the takeaways.
Start reading ↓
Watch
The recording
The full session on video, including the panel with Lorna and Lili.
Watch the session ↓
Download
The slide deck
The full presentation as a PDF. Share it with your team, or with IT and procurement.
Download the deck ↓
The deck

Download the full presentation

Every slide from the session: the shared-brain layers, the Cowork setup, the supervision surface, and the effort curve from the trust conversation.

Download PDF
Watch

The full session

Paul's live build, the Flank walkthrough, and the conversation with Lorna and Lili. If you'd rather skim, the written recap below covers the essentials.

Part 1 · Build the foundation

Four primitives, one shared brain

Everything Paul showed is the setup Flank actually runs on, live on screen, not a canned demo. It starts with the four building blocks of Claude Cowork.

01
Connectors
The tools Cowork can read and write to: your file storage, Slack, calendars, knowledge bases. Paul's build needed just two, Box (where the brain lives) and Slack (to DM him).
02
Skills
A prompt that captures how you like a task done. The trick: don't write it yourself. Paul asked Claude to build his "brand check" skill, and it wired in the brain's guidelines on its own.
03
Tasks
Actual runs of work, including scheduled ones. Paul set one up live: every Monday, a Slack DM nudging him to write a LinkedIn post, then an offer to check it against the brand.
04
Projects
Workspaces where related tasks share context and memory. Start one per recurring job, and it gets smarter about that job as you use it.
Part 1 · Build the foundation

The shared brain: three layers of context

The interesting part isn't Cowork itself. It's the "brain" behind it: one knowledge base every seat in the company plugs into, so the whole team works from the same context instead of diverging.

Foundational
The core truths: positioning, brand, who your customers are, the problem you solve. Changes yearly at most, and updates stay manual. Some things you never hand to an AI.
Operational
What's true now and for the next weeks or quarter: current projects, campaigns, personnel and roles. Refreshed monthly, partly by scheduled tasks.
Real time
The daily stream: market signals, competitor moves, what customers are saying. A weekly task combs Slack, call transcripts, and the web, and clears the noise out.

Under the hood it's nothing exotic: plain markdown files in Box. Two design choices do the heavy lifting. Every document opens with instructions for agents, because the system is written for AI to read, not humans. And a single README tells Claude how to navigate the whole thing, so it goes straight to the right folder instead of reading everything. Faster, cheaper, more reliable.

How it got built

Nobody wrote the brain by hand. The team pointed Claude at everything lying around: laptops, drives, LinkedIn, the website. Claude drafted the documents, flagged gaps and contradictions, and humans reviewed. Start with a small, coherent version one. It's easy to grow from there.

Part 1 · Build the foundation

Context that goes stale is context that gets abandoned

Every legal team knows the SharePoint graveyard: built with goodwill, dead within a year. The difference here is that the brain maintains itself, on a schedule that matches each layer.

1
Weekly · divergence check and signal synthesis
Scheduled tasks comb the real-time layer, kill anything stale or conflicting, and date everything so agents know what to trust.
2
Monthly · operational refresh
Roles, projects, and current work get reviewed. Some of it syncs from systems like Personio; some stays a human decision.
3
Manual · the foundational layer
Core truths change rarely and deliberately. When you spot drift, describe the problem to Claude. It's genuinely good at designing its own maintenance tasks.
The line worth remembering

AI tools are only as good as the context they're given. When the context goes stale, they stop being used.

And the "marketing brain" label is a misnomer. Jake's framing: this matters more in legal than almost anywhere else. Consistency of positions, house language, knowing who's an existing client, how a specific customer likes things formatted. Enterprise teams are fragmented and lossy; people leave and take context with them. The brain is the team member who holds all of it and never resigns.

Part 2 · Work that should leave the desk

A tool makes you faster. Some work you just want gone.

Paul's honest boundary: Cowork is excellent, and Flank uses it everywhere. But a tool can only make a person faster. For high-volume, routine, quasi-legal work, the kind you'd push to an ALSP or an offshore center, faster isn't the goal. Off the team's plate is the goal.

That's the domain-specific side. Flank agents deploy where requests already arrive: shared inboxes, Salesforce, Ironclad, Teams. No new tool for the business to learn, no change management. Paul showed it end to end: he emailed a licence agreement drafting request with deal documents attached, the way anyone would email legal, and the agent replied with a draft and a summary of what it did.

What legal gets
  • A single view of all demand across every channel, often for the first time.
  • Triage by your rules: ignore it, route it to an agent, or route it to a named human.
  • Agents that own whole categories: NDAs, drafting, review, Q&A.
What supervision looks like
  • Playbooks get codified into rules. Low-risk points auto-confirm and get redlined without a human.
  • Only high-risk points, or ones the agent isn't confident about, get surfaced for a lawyer to check.
  • You decide where that line sits, and it moves as trust builds.
Why the surface matters

If a lawyer has to check everything, you've saved nothing. The supervision panel exists so the lawyer checks only what's essential. Customers start by supervising a lot, and the burden drops week by week, toward only the genuinely high-risk third.

Part 3 · Trust and supervision

The conversation: Lorna and Lili on letting go

Lorna Khemraz leads legal AI alignment at Flank and spends her days inside real enterprise deployments. Her opening admission set the tone: the first reaction to "the agent does the work" is confusion, then scepticism. And that's rational.

Every legal task hides dozens of micro-decisions, and lawyers are rightly wary of outcomes they can't trace. Handed a finished review with no visibility, Lorna said she'd do what any lawyer would: check the whole thing end to end, because she's still accountable. At which point you might as well have done it yourself.

What changes that is visibility plus consistency. With the supervision panel, there are no unanswered questions: why the agent made each call, what it escalated and why. Over time she checks less and less, "because the agent is effectively acting as an extension of me. It's working to my standards."

Standard, not preference
Five lawyers review the same contract five ways, and all five can be right. The work is stripping preferences back to an objective, documented standard. That's harder than it sounds, and it's the real starting point.
🤝
Like onboarding a hire
Dialing in an agent is the same process as a new team member: right context, right guidance, supervised closely at first. The signal it's working is that supervision involvement declines.
The honest curve
Paul to customers, verbatim: it won't feel easy on day one, two, or three. There's real upfront investment. But it pays off quicker than people expect, dropping toward supervision of only the high-risk exceptions.
Lili's frame

Legal teams have made this trust journey before, every time work moved to a new hire, a service center, an ALSP, or a law firm. Competence was never the whole question; context was. You supervise everything at first, then let go task by task. The only difference now is that the new resource is infinitely scalable. And for some work you'll never go fully autonomous, which is fine: your people stop doing the work and become the arbiters of what good looks like.

If you remember five things

The takeaways

  1. Aim at the function, not the individual. Lawyers getting faster at their own tasks is not the same as your legal function shipping more, cheaper, with better throughput. Build for the second thing.
  2. Foundation before agents: a shared brain. One layered knowledge base (foundational, operational, real time) that every seat plugs into, written for agents to read, navigated through a single README.
  3. Freshness is the whole game. Scheduled tasks patrol each layer on its own rhythm. Context that goes stale gets abandoned, exactly like every shared drive you've ever seen die.
  4. Tools make people faster. Agents remove work. For routine, high-volume categories, deploy agents where requests already arrive and let them own the outcome, with supervision focused only on what's high risk or uncertain.
  5. Trust is a journey with a familiar shape. Codify a standard rather than preferences, supervise like you'd onboard a hire, and expect the burden to fall. Your lawyers become the arbiters of the standard, not the people doing the work.
The next session · Wed 16 September

This thread continues. Lorna Khemraz goes deep on exactly this territory in the next webinar: The 2027 operating model for in-house legal. Building systems your team can actually rely on, and the skills it takes to run them.

Save your seat for 16 Sept

Take it with you

Download the deck

Share the full presentation with your team, IT, or procurement. Everything Paul walked through, in one PDF.

Download PDF
Subscribe

The Intake

Weekly briefings on what's actually changing in legal AI: the market shifts, the regulatory moves, and the structural questions that matter for enterprise legal teams. We'll also send details of the next webinar.

Subscribe on Substack
Flank

Outsource legal work to supervised agents

Enterprise legal teams use Flank to handle high-volume contracting end to end: NDAs, MSA redlines, procurement, and triage. Agents that know your templates, your terms, and your escalation rules.

Learn more at flank.ai