AI for Legal Department: What AT&T Is Building

Why AT&T's legal overhaul matters beyond telecom

Until recently, bringing AI into a legal team meant a lawyer pasting a clause into a chatbot and hoping for the best. Legal was the department where new technology went to wait. Contracts, compliance questions and risk memos moved through email threads and the memory of senior people. Now a telecom group with roots going back to 1885 is publicly redesigning that machinery with OpenAI, and the design starts with a single question: what happens to a request before a lawyer ever reads it?

That question is the reason this story deserves a CEO's attention right now. The answer AT&T gives is more specific, and more cautious, than the headlines suggest. It also puts a quiet price tag on something most executives treat as a fixed cost. We will get to the numbers the company has disclosed, the ones it has not, and what a company one-hundredth of AT&T's size can copy from it.

AI for legal department work is no longer a lab experiment. It is a management decision about how fast the business can get a reliable answer from its own lawyers. In October 2026 AT&T announced that it is acting as a flagship in-house legal design partner for OpenAI, with the declared aim of building what it calls the AI-enabled in-house legal department of the future. For owners and COOs, the useful part is not the brand names. It is the architecture: where AI sits, where it stops, and who stays accountable.

What AT&T actually announced

The facts matter here, because much of the commentary has inflated them.

AT&T announced the collaboration on October 8, 2026. OpenAI followed on October 21 with its own account, titled "How AT&T is rethinking in-house legal work with OpenAI." Both rest on LegalEdge, an in-house legal agency that AT&T launched in January 2026. OpenAI describes it as a team of attorneys, legal professionals, technologists and project managers focused on applying AI to major legal matters.

The numbers that are public

AT&T says LegalEdge handled nearly 100 matters in its first six months. Trade coverage reports that it started with roughly 38 legal professionals, after a pilot of 24 to 30 people. AT&T's legal function overall counts hundreds of lawyers and legal professionals worldwide, which is exactly why expertise and past work end up scattered across teams.

Notice the sequence. AT&T built the team first and added the AI layer afterwards. It did not buy a tool and then look for a use for it.

What has not been published

There is no disclosed percentage of cost saved, no hours-per-contract figure, no accuracy benchmark. One outlet reports that thousands of hours of work have already been redirected in-house this year, but AT&T's own materials don't give a savings number, and I won't invent one. The planned intake system is also still in development. Independent coverage of OpenAI's post was explicit that it is a development effort, not a released product.

Success, in the words of AT&T's chief compliance officer, will not be measured by how much technology gets deployed.

That sentence is the most useful line in the whole announcement. It tells you what to ask your own vendors.

The intake idea

Bill Ryan, AT&T's chief compliance officer, describes one planned initiative as an intake system with "a front door where we have natural language intake." Requests arrive from across the business. Today someone has to understand the issue, collect context and decide what is urgent before any advice can be given. According to OpenAI's account, the aim of such a system would be to give lawyers a more informed starting point, with less time spent gathering the material they need to assess a matter.

The lawyer still decides. General Counsel David McAtee put it flatly: legal judgment remains human.

Use cases: AI for legal department work that is realistic today

AT&T has named a few application areas: answering frequently asked questions, gathering and synthesising information for attorneys, and knowledge management. Add what is already proven in the wider market and you get a practical shortlist.

Intake, triage and routing

A business unit types a question in plain language: "Can we sign this vendor on their paper?" An agent classifies it (contract, privacy, employment, regulatory), flags urgency, pulls the three most similar earlier matters and routes it to the right person. This is the AT&T scenario in miniature, and it works for a 5-person legal team as well as a 500-person one.

Self-service answers for repeat questions

Every legal department has a list of questions it answers monthly: signing authority, standard NDA terms, data-retention rules. AT&T says that for some tasks, such as FAQs, an agent may provide the response. The lawyers decide which tasks qualify. That boundary is the control, and it should be written down, not assumed.

Research and context assembly

Before advice comes preparation: reading the file, finding the earlier opinion, summarising the regulation. This is the repetitive layer the AT&T partnership wants software to handle, so lawyers keep their time for judgment. Architecture choices matter here; if you are weighing how an agent should retrieve your documents, our comparison of RAG vs CAG architectures is a useful starting point.

Drafting and review support

Contract drafting is the most visible category in the market. For a view of how a specialist legal AI vendor approaches it, see our breakdown of Harvey with GPT-6 Astra for legal drafting. Drafting is also where errors cost the most, so it needs the tightest review loop.

Savings: how to model the return, and what the comparison looks like

AT&T gave no savings figure, so any number you see attributed to this program is someone's guess. What you can do is build your own model from your own data. The structure below is my analysis, not AT&T's disclosure.

Dimension	Email and ticket intake (typical today)	AI-assisted intake (AT&T-style design)
First step on a request	A lawyer reads it and asks clarifying questions	An agent structures the request and gathers context first
Prioritisation	By whoever opens the inbox first	Classified and ranked by risk and urgency
Use of past work	Depends on who remembers a similar matter	Precedent and prior research surfaced automatically
Routine questions	Each one consumes lawyer time	Agent answers approved FAQ categories
Lawyer's time split	Heavy on gathering and sorting	Heavier on analysis and advice
Outside counsel spend	Routine work often goes out the door	Routine work stays in-house; external firms for specialist matters
Main risk	Delays and missed issues	Wrong or leaked output if guardrails are weak
Accountability	Individual lawyer	Individual lawyer, plus a documented AI policy

The outside counsel row is not hypothetical. Coverage of the announcement says AT&T plans to trim its roster of outside law firms and to favour firms that adopt AI. Legal commentators draw the same conclusion: in-house teams may buy less routine support and select external firms for specialist depth and the ability to work with internal systems.

A back-of-envelope model

Take a mid-sized company, purely as an illustration. Suppose legal receives 2,000 requests a year, and structured intake plus context assembly saves 1.5 hours per request. That is 3,000 hours. At a loaded internal cost of $90 an hour, it comes to $270,000 a year in capacity, before counting any reduction in outside counsel fees. Every one of those inputs is an assumption you should replace with your own ticket data, and that is the point: if you cannot count your requests today, you cannot prove a return tomorrow.

Risks Legal Ops must control before scaling

Legal is a place where a confident wrong answer is worse than no answer. Coverage of the AT&T partnership stresses that AI tools save time but errors carry serious consequences, and confidential information needs strict protection. Four risks deserve a named owner each.

Hallucinated or outdated answers. An agent that answers FAQs must cite the internal source it used, and the approved-answer library needs an expiry date for each entry.

Confidentiality. Privileged material, deal data and employee matters should never leave your control surface. Review vendor terms on training and retention; our piece on corporate data privacy in AI training covers what changed in this area.

Silent scope creep. An FAQ agent slowly starts answering questions nobody approved. AT&T's answer is structural: lawyers determine how agents are used, and human review is built into every workflow. Copy that, and log what the agent actually handles.

Vendor dependence. If the knowledge layer lives in a vendor's black box, switching later is painful. A short evaluation checklist helps; see how to choose an AI agent vendor.

The control that matters most is not technical. It is a written list of what the agent may answer alone, what it may only prepare, and what it must never touch.

How to start an AI for legal department pilot without a Fortune 50 budget

You don't need hundreds of lawyers. You need a narrow door and a measuring stick.

1. Count the traffic. Export 6 to 12 months of legal requests. Tag them by type and by who asked. You will almost certainly find that a handful of categories make up most of the volume.
2. Pick one category. NDAs, vendor questionnaires or signing-authority questions are common candidates. Choose something repetitive and low in stakes.
3. Build the front door first. Structured intake and routing before any automated answers. This mirrors AT&T's sequence and gives you data on your own request mix.
4. Define success beyond usage. AT&T says it will judge results on quality, access to legal knowledge, reusable capabilities and more room for work that needs experience, rather than technology deployed. Add your own: turnaround time, escalation rate, percentage of agent answers a lawyer later corrected.
5. Review monthly with the general counsel. Expand the agent's scope only by written decision.

Done this way, the pilot should give you something executives rarely get from legal: a dashboard. You see how many requests came in, how long each class took and where work piles up, so you can decide on evidence rather than on the loudest complaint. The calm of knowing that routine contracts are moving, and that nothing risky is being answered by a machine alone, is a real operational gain, not a soft one.

There is a reputational side too. A leader who can tell the board exactly where AI is used in legal, where it is forbidden and what it returned in the first quarter comes across as someone who runs technology on purpose. Investors tend to trust that more than a vague line about "exploring AI."

FAQ

Can AI replace in-house lawyers? No, and AT&T does not claim it can. The company says legal judgment remains human and that lawyers determine how AI agents are used. The realistic goal is to take over repetitive preparation and knowledge retrieval so lawyers have more time for strategy and advice.

Is AT&T's AI intake system already live? Not as a finished product. OpenAI's account describes a planned natural-language intake system, and independent coverage characterises it as a development effort. What is already running is the LegalEdge team, which AT&T says handled nearly 100 matters in its first six months.

How much does AI for a legal department save? AT&T has not published a savings figure, so any exact percentage attributed to this program is unverified. Build your own estimate from request volume, hours saved per request and your outside counsel spend, then test it in a pilot.

What should a smaller company automate first? Start with intake and routing, then repeat questions with approved answers, such as standard NDA terms or signing authority. These categories are high in volume, low in risk and easy to measure.

What are the biggest risks of legal AI agents? Wrong or outdated answers, leakage of confidential material, scope creep beyond what lawyers approved, and lock-in to a single vendor. Each needs a named owner and a written policy before the agent goes near live matters.

From announcement to action

The AT&T story is useful because it is unglamorous. No claim of replacing counsel, no miracle percentage: a team built first, a front door for requests, lawyers holding the pen, and AI placed where it removes friction. That is a pattern you can copy at any scale.

If your legal requests still travel by email and memory, the next step is a map, not a purchase: what comes in, what repeats, what an agent could safely prepare. Our team can help you build that map and size a first pilot. Book a 15-minute consultation and bring your last quarter's request list.