An AI agent for legal work takes a goal, plans the steps, runs them against your documents and sources, and hands back a draft with your name still on the review. In-house teams are running them today on first-pass contract review, scheduled regulatory scans, and quarterly vendor terms checks.
The agents that earn their place share a history. Somebody wrote the workflow down first, step by step, the way you would brief a lateral in their first week: the intake questions, the review order, the always/never rules. Teams that skip that step get a fast tool producing work nobody wants to sign.
That sequencing came up on CZ and Friends well before "agent" landed on every legal AI homepage. Nicole Altman, Senior Counsel at Instacart, put the enthusiasm in order:
"Agents are like the shiny new thing. Everybody's talking about agents. And that's great and good, but there's so much amazing things that AI can do just with the chatbots."
Nicole and Kelly Noguchi, Instacart's Senior Legal Technology and Operations Manager, built an AI-first legal culture at public-company scale, and they got there by standardizing the unglamorous parts before automating any of them.
Cecilia Ziniti built GC AI, the enterprise-grade legal AI platform for in-house teams, after three tours as a general counsel. It runs agents today: Chat 2.0 coordinates multiple agents on a single request, Research deploys simultaneous agents across primary law and the web, and Playbooks run agentic multi-step contract review. As of September 2026, 2,000+ in-house legal teams run it daily, including the legal teams at Snyk, Columbia Sportswear, and Eventbrite.
What Is an AI Agent, and How Is It Different From a Chatbot?
An AI agent for legal work is software that takes a goal, breaks it into steps, executes those steps with tools like document analysis and web research, and returns a finished draft for a lawyer's review. A chatbot answers the question you asked. An agent owns the task you delegated: it plans, runs the intermediate steps, checks its own work against the instructions you gave it, and delivers the output in the format you specified.
The distinction matters for buying decisions because "agent" now appears on legal AI landing pages across the market. The test is behavioral. If the system can run a multi-step workflow you defined, on documents you supplied, and hand back a deliverable you would grade, it is functioning as an agent. If it needs you to steer each step, it is a chat interface with better marketing.
The 5 AI Agents In-House Legal Teams Run Today
Five agents dominate in-house adoption, and they share one trait: high volume, repeatable structure, and a lawyer at the end of the line.
First-pass contract review. An agent runs your negotiation standards against each incoming NDA or MSA and returns findings with quoted language.
Regulatory monitoring. Scheduled scans of named regulators and jurisdictions, summarized against your risk profile, the workflow behind AI for compliance monitoring. GC AI's Regulatory Monitoring Skill Creator packages this as a named build: set your jurisdictions, risk areas, and a materiality threshold, and the agent verifies every flagged item against primary sources and returns a prioritized action-items table on the schedule you set.
Vendor terms re-checks. Quarterly re-review of key vendors' terms of service for changes that matter to your data posture.
Research memos. Multi-agent research across primary law with citations, compressed from an afternoon to minutes.
Document extraction. Pulling renewal dates, liability caps, and change-of-control triggers across a contract stack.
Here is the shape of the vendor re-check as a scheduled workflow: each quarter, pull the current terms of service for five named vendors, compare each against the version on file in Files, flag changes to data use, subprocessors, or liability, quote the changed language exactly, and deliver a one-page summary marked review-needed or no-action. A lawyer reads the summary in four minutes. The agent did the reading that never fit anyone's week.
In GC AI, the scheduled versions of these run through Automations. Here is one running a recurring legal task end to end:
The Gap Between a Demo and a Deliverable
Enterprise AI has a graveyard, and it is well documented. MIT's NANDA initiative reported that roughly 95% of corporate generative AI pilots stall before measurable impact, and that tools from specialized vendors succeed roughly three times as often as internal builds. The study's core finding lands hard for legal: generic AI excels for individuals and stalls in enterprises, because it never learns the organization's workflows.
Kelly Noguchi lived that finding before it had a citation. The gap between one lawyer's personal productivity and a scalable, enterprise-ready process is where the challenge lives, she told Cecilia, and teams cannot "glaze over" the work of turning excitement into execution. Nicole put the same point in the voice of an early adopter:
"It's magical what it can do. These tools now open up these possibilities that were not possible before. But it takes time. It takes a lot of investment of time and grit and patience to get stuff out of them."
The teams that cross the gap follow a sequence. First they standardize the workflow while a human still runs it: the intake questions, the review steps, the output format, the always/never rules. Then they hand the standardized version to the machine. In GC AI that standard is a Skill, built in plain language, and the build process is documented step by step in how to build Skills in GC AI.
Share it from the Skill Library and the team runs one process; schedule it with Automations and the process runs itself.
An agent you can trust is a workflow you already standardized, running on a schedule, with your name still on the review.
Is It Legal to Use AI Agents on Client Work?
Yes. Lawyers may use AI agents on client work, and the professional duties travel with the lawyer at each step. Three duties do the heavy lifting:
Competence: The State Bar of California's generative AI guidance, updated May 2026 specifically to address agentic AI, expects lawyers to develop a reasonable understanding of an AI system's capabilities, limitations, and risks before deploying it.
Review: The same guidance draws a bright line for agentic filings: a document reaches a court only after lawyer review and approval. The practical rule generalizes beyond litigation, and GC AI's breakdown of ABA Formal Opinion 512 covers the duty as it applies to everyday advice work.
Confidentiality: An agent touches more documents than a single chat, so the platform's data posture carries more weight. GC AI publishes its full subprocessor list, maintains zero-data-retention agreements with its LLM providers wherever feasible, and holds SOC 2 Type II and SOC 3 certification, and the Heppner ruling explainer covers why counsel-directed use matters for privilege.
Frame the question the way a regulator would: the agent is a fast non-lawyer assistant, and ABA Model Rule 5.3 already tells you how to supervise one of those. The full duty map, with the cases that set the verification standard, is in AI legal ethics in 2026.
What Is the Best AI Agent for In-House Legal Work?
The right answer depends on whose work the agent has to survive. Most "best legal AI agent" roundups rank platforms for law firms, where the job is billable matter work. In-house teams grade on volume, turnaround, and whether the output holds up to a business partner who wants an answer today.
Four criteria separate the platforms that hold up:
It runs your standard, not a generic one: The agent should execute the playbook your team negotiated, with your fallback positions, not a vendor's idea of market.
Every claim traces to a source you can open: Exact Quote gives character-level citations back to the underlying document, which is what makes a first pass reviewable in minutes.
It runs without a human starting it: Scheduled execution is the difference between a workflow you use and a workflow that happens.
The data posture survives your own vendor review: SOC 2 Type II and SOC 3, GDPR compliance, zero data retention agreements with the model providers, AES-256 encryption.
GC AI was built against those four for in-house teams specifically. For how the wider market maps against them, see the best legal AI tools for in-house counsel.
Does the Model Behind the Agent Matter?
The system around the model decides more than the model itself. A frontier model with no legal system prompt, no document grounding, and no verification layer produces confident prose and unverifiable citations. GC AI's In-House Legal Bench, published May 2026, measured the difference across 100 in-house tasks scored against 1,200+ attorney-developed criteria:
GC AI: 86.8%
ChatGPT (GPT-5.5): 79.8%
Claude (Opus 4.7): 68.4%
Gemini (3.1 Pro): 57.5%
The category cut matters more than the total when you are deciding what to hand over. GC AI led all ten task categories, and the two widest margins fall on the work in-house teams delegate to agents first, regulatory tracking and legal research:
Legal Task Category | GC AI | ChatGPT | Claude | Gemini | GC AI Margin |
Regulatory tracking | 88.6% | 73.5% | 68.2% | 45.0% | +15.1 |
Legal research | 88.3% | 75.6% | 66.2% | 61.7% | +12.7 |
Contract analysis | 82.7% | 72.8% | 66.3% | 42.9% | +9.9 |
Checklists | 89.9% | 81.9% | 73.4% | 59.3% | +8.0 |
Comparison and benchmarking | 91.4% | 84.7% | 81.4% | 72.9% | +6.7 |
Extracting information and data | 82.0% | 76.9% | 57.0% | 56.3% | +5.1 |
Risk assessment | 89.0% | 84.2% | 71.1% | 59.2% | +4.8 |
Drafting | 87.6% | 83.4% | 74.9% | 66.4% | +4.2 |
Summarizing documents | 81.6% | 77.5% | 63.7% | 57.5% | +4.1 |
Legal strategy | 86.3% | 84.5% | 63.0% | 58.0% | +1.8 |
The platform's advantage comes from the layer above the model: a legal-specific system prompt, Exact Quote's character-level citations, and workflows tuned to in-house tasks, running on models from OpenAI, Anthropic, Cohere, Reducto, and Google. The buying question shifts accordingly. Ask which platform turns the model into supervised legal work product, and see the full comparison in GC AI vs ChatGPT.
Give an Agent Its First Job
Start with a workflow you already trust yourself to run: the weekly NDA, the quarterly vendor check, the recurring regulatory scan. Standardize it, test it, then schedule it.
Frequently Asked Questions
Can AI Agents Review Contracts Without a Lawyer?
AI agents produce the first pass, and the lawyer owns the judgment. An agent can run a playbook against an NDA, quote the relevant language, and flag deviations in minutes, which converts the lawyer's role from reading everything to reviewing findings. The State Bar of California's 2026 guidance makes that division explicit, requiring lawyer review before any output acts on anyone's behalf.
Does Using an AI Agent Waive Attorney-Client Privilege?
Where the documents go decides it. In a cautionary aside in Munir v Secretary of State for the Home Department ([2026] UKUT 81), the UK Upper Tribunal observed that uploading confidential material into an open-source tool such as ChatGPT places it in the public domain and risks waiving legal professional privilege, while distinguishing closed, enterprise-grade tools that don't expose data publicly. The US analysis turns on counsel-directed use and the terms governing the platform, covered in the Heppner ruling explainer. The working rule for agents: route privileged material only through platforms with contractual confidentiality and zero data retention wherever feasible, backed by SOC 2 Type II certification, and keep a lawyer directing the work.
Are AI Agents Safe for Confidential Legal Work?
Yes, on infrastructure built for legal work. GC AI is SOC 2 Type II and SOC 3 certified, GDPR compliant, with zero data retention agreements with OpenAI and Anthropic, and AES-256 encryption. Buyers evaluating any agentic platform should ask how the vendor isolates client data, whether model providers retain prompts, and how counsel-directed use supports privilege.
What Is the Difference Between an AI Agent and a Skill in GC AI?
A Skill is the standardized workflow: the instructions, intake questions, steps, and always/never rules that encode how your team runs a task. An agent is the execution layer that carries a multi-step task through to a deliverable. Save a workflow as a Skill, schedule it with Automations, and you are running a supervised agent on your own standards.
What Should an In-House Team Automate First?
Automate the task you already run the same way each time: recurring NDA review, quarterly vendor terms checks, or a scheduled regulatory scan. High volume plus repeatable structure plus lawyer review at the end is the profile where agents return time fastest, and it is the pattern in-house teams report across GC AI's 2,000+ customer base.
