On November 6, GC Collective and Streamline AI hosted Building Your First Agentic Junior Associate, a practical, hands-on virtual workshop for in-house legal leaders and their teams.
James Cosgrove, Vice President and General Counsel at CookUnity, showed how he has built AI agents that now handle most of his team's business-as-usual legal work. Evan Michalski, Principal Solutions Architect at Streamline AI, helped turn James's approach into steps any legal team can take, whether you're building your first agent or connecting your tenth.
Meet the speakers
James Cosgrove
Vice President and General Counsel
CookUnity
Brought the in-house view of a General Counsel deploying AI agents in his own team's legal work.
Evan Michalski
Principal Solutions Architect
Streamline AI
Covered the technical and implementation side: how agents are designed and connected to legal workflows.
A three-lawyer team, a high-growth company
James joined CookUnity, a prepared meal service, as its first legal hire in March 2024. The company grew fast, and its riskiest areas are specialized: consumer protection, data privacy and food law. With a limited budget, he chose not to hire junior versions of himself. He hired subject matter experts in food law and healthcare law instead, and built agents to handle the rest.
Since he started building in earnest in March this year:
Legal costs, covering outside counsel and technology including AI usage, fell 28% quarter over quarter.
Roughly 80% of business-as-usual work (commercial contracts, outside counsel invoices, tracking legislation and litigation) is now handled by agents, with a lawyer reviewing every output.
Average turnaround on legal requests dropped from about 7.5 days to 2.8.
Watch the recording and get the resources
Complete the form below for instant access to the full workshop recording and two practical resources:
The Agentic Junior Associate Prompt Kit: five copy-ready prompts that take you from your team's know-how to a working, tested agent.
The Commercial & DPA Training Playbook: the fictional sample playbook James used in the live demo, so you can build and test your first skill.
Thank you to James, Evan and the Streamline AI team for a generous, practical session.
What you need, and what you don't
An enterprise AI account, so your data is not used to train the models. Coordinate with IT on what the tools can access.
A small, clear first objective.
As much good data and precedent as you can give it. The better the inputs, the more the output sounds like you.
Curiosity. Talk to it, typed or spoken, and ask it to explain anything you're unsure about.
What you don't need is a technical background. James is a former political science major and M&A lawyer.
Four building blocks
Playbook: the rules the AI applies to a task, such as how you typically mark up a SaaS agreement.
Skill: a defined task you hand off to the AI, which applies your playbook to it.
Workflow: a sequence of activities that moves work into and out of a skill.
Agent: a bundle of skills and workflows, so you only step in at the end to review.
How-To in 5 steps
Step 1: Capture your judgment in a playbook
James connected his email, Slack and document management system, then asked the AI to review his past drafts, from first to final, and turn the back-and-forth with opposing counsel into rules. It took about half an hour. His advice: expect the first version to be about 80% right, and review it closely. His first draft overweighted assignment clauses, a leftover from years of M&A work. If you don't have a deep archive of precedent, ask the AI to interview you about how you approach a review.
Step 2: Build a skill that applies it
In the live demo, James described every step he takes from intake to deliverable, then told the AI to apply the playbook rather than explaining how to review the document. Where his instructions had gaps, it asked him questions. In about nine minutes it built the skill, applied it to a sample SaaS agreement, and produced a summary memo and a redline. Two design choices worth borrowing:
Triage by materiality. Have it rate each issue green, yellow or red, and redline only the red ones, so the business gets something usable and you stay out of the weeds.
Explain every edit. A margin comment on each change, pointing to the relevant playbook section, keeps the playbook honest and forces a real human review. That's how James caught the agent applying his vendor playbook in reverse on a deal where CookUnity was the vendor.
His key takeaway: tell your AI what decisions it should make, what evidence it should use, and what a finished result looks like. Evan added that it learns fastest from examples.
Steps 3 to 5: Workflows, agents and testing
Workflows get work into and out of a skill. They need triggers, finishing instructions and approval boundaries; James never lets a deliverable go out without human review. One of his workflows sweeps his inbox weekly for outside counsel invoices, categorizes them and sends them to accounts payable. He hasn't forwarded one by hand since.
His agent bundles it all together. Requests arrive through Streamline from a dedicated Slack channel and email address. A scheduled sweep routes each one to the right playbook (NDA, MSA, DPA, real property or marketing promotions), converts PDFs to Word, checks legal citations against a legal research service, posts the final redline back to Streamline and files drafts in Google Drive. He spends an hour each morning reviewing the results. In the session's closing demo, a new request went from intake to a finished redline and summary in six and a half minutes.
Testing is the step not to skip. Keep throwing new scenarios at your skills, and ask the AI to review your process and suggest improvements.
How to get started
Pick one annoying, repeatable task and brainstorm with the AI about how to solve it.
Define what the finished deliverable looks like.
Build in a human review gate.
Connect it to a system of record, such as your document management system or CLM.
Not sure where to begin? Evan suggested asking your AI for a morning brief across your email, Slack and open requests. The tasks that keep showing up are often your first agent.
Two tips for advanced builders
Optimize for token efficiency. Once a skill works well, ask the AI to streamline it, so it isn't rebuilding your whole process every day. It made James's budget go much further.
Stay portable. If another AI provider offers better pricing, ask your current tool to export your skills and workflows, then import them into the new one.
