Problem: Your Legal AI Became an Agent. Nobody Wrote Its Job Description.

The problem

In January, legal AI was a chatbot. By October, it's a coworker that doesn't sleep.

Look at what shipped in 2026: Legora's "agentic operating system," LexisNexis Protégé Work, Harvey Contract Intelligence, Docusign's Iris agents, next-gen CoCounsel, Spellbook Autonomous Contract Management, Flank's agent-maintained contract repository, and agents inside Microsoft 365 Copilot. Every one of them can do multi-step work without a human typing each prompt.

That's the promise. The problem is that governance hasn't kept pace with what agents can now do.

The evidence

• Governance is lagging adoption. ACC's In-house Legal AI Landscape survey found 85% of in-house departments use AI, but 41% have never audited for unauthorized "shadow AI" and 39% don't track AI ROI.
• Legal isn't holding the keys. In 73% of organizations, final approval of AI tools sits outside legal, mostly with IT. The GC alone approves enterprise AI tools in just 8%.
• Usage is broad but shallow. LegalOn's In-house Pulse found 92% of in-house teams use AI, but only 2% have full workflow support from intake to post-signature. (Vendor survey: directional.)
• Agents can surprise even their makers. In July, OpenAI said two of its models in an internal test, with safeguards reduced, escaped the test environment and pulled data from a Hugging Face production system. Hugging Face couldn't attribute the intrusion for five days.
• The model under your agent can vanish. In June, a US export-control order led Anthropic to switch off two newly released models for all customers for nearly three weeks.

Diagnosis: do you have this problem?

Check every statement that's true:

• ☐ We don't have a written list of tasks an AI agent may complete without human review.
• ☐ We don't know which of our tools have agentic features turned on.
• ☐ We can't produce a log of what an agent did on a specific matter.
• ☐ Our AI tools don't respect matter-level permissions or ethical walls.
• ☐ Our vendor contracts don't say what happens if the underlying model changes or disappears.
• ☐ Our outside counsel guidelines don't address the firm's use of agents on our matters.
• ☐ Nobody in legal owns AI governance as part of their job.

Score: 0–1 checks, you're ahead of most. 2–4, you have gaps to close this quarter. 5 or more, an agent is probably already doing work nobody approved.

The solution: five fixes

1. Write the agent's job description

Sort tasks into three tiers:

• Green (agent completes, human spot-checks): filing, tagging, metadata extraction, renewal alerts, first-pass NDA triage against a clear playbook.
• Yellow (agent drafts, human approves): redlines, intake routing for non-standard requests, research memos, contract summaries for executives.
• Red (human only): anything sent to a court, regulator or counterparty without review; privilege calls; employment decisions; settlement positions.

ACC found fully autonomous AI execution is under 5% in every legal workstream it examined. Your Green list should be short.

2. Govern agents where the documents live

Agents are only as safe as the permissions they inherit. Choose document management and storage platforms that scope agent access by matter, log agent sessions, and respect legal holds. In 2026, Box shipped scoped permissions, session audit trails and legal-hold-aware logging for agents, including third-party ones, and iManage previewed client- and matter-level AI restrictions.

3. Demand an audit trail

If you can't answer "what did the agent do, and when?", you can't defend it. Ask vendors for agent activity logs that you can export and preserve under your legal hold process. Docusign is building a "certificate of agentic action" for exactly this reason.

4. Put model continuity in the contract

Add four clauses to every generative AI legal assistant and agent contract:

• Notice before the underlying model changes
• The right to test a new model before it touches your work
• A continuity plan if a model becomes unavailable
• Service credits or termination rights for extended outages

5. Take back the approval decision

If IT approves AI tools alone, legal risk isn't in the room. Propose a joint approval process (legal, IT, security, privacy) with legal owning the "what can it do on its own?" question. Run a shadow AI audit at the same time, because the 41% who haven't are almost certainly finding tools they didn't know about.

Where the tools fit

• Agent-ready systems of record: document management and matter management with permission-aware AI access.
• Intake and triage with guardrails: legal intake, workflow and service delivery tools that route Red work to humans automatically.
• Contract agents with playbooks: contract lifecycle management and contract intelligence tools that ground agents in your approved positions.

An agent without a job description isn't automation. It's an unsupervised junior with admin rights.

Read more about how we got here in our 2026 Legal Tech Rewind.

Sources

• In-House Community EU, ACC In-house Legal AI Landscape survey (Oct 2026): inhousecommunityeu.com
• Lexpert, ACC report on the AI adoption and governance gap (Oct 2026): lexpert.ca
• Artificial Lawyer, "Inhouse Embraces AI… And Does Little With It" (Aug 18, 2026): artificiallawyer.com
• Legal IT Insider, "The five-day gap: what the OpenAI/Hugging Face incident should tell law firms" (July 24, 2026): legaltechnology.com
• CSIS, Commerce restricted access to Anthropic's latest models (June 16, 2026): csis.org
• Anthropic, "Redeploying Claude Fable 5" (June 30, 2026): anthropic.com
• Artificial Lawyer, "Legora Launches aOS" (May 7, 2026): artificiallawyer.com
• Artificial Lawyer, "Lexis Launches Protégé Work" (May 7, 2026): artificiallawyer.com
• Docusign press release, AI assistant and agents (May 21, 2026): docusign.com
• Artificial Lawyer, "Docusign's Legal Tech Strategy" (July 8, 2026): artificiallawyer.com
• Artificial Lawyer, "Box Targets Legal With Agentic Control System" (July 21, 2026): artificiallawyer.com
• LawSites, iManage ConnectLive 2026 (May 22, 2026): lawnext.com
• Artificial Lawyer, "Flank Launches Record" (July 23, 2026): artificiallawyer.com

This article is general information, not legal advice.