The first useful AI project for a busy legal team may be a better way to receive a question. In an October 21 account from OpenAI, AT&T describes a planned system that would accept requests in ordinary language, sort them, find relevant past work, and give a lawyer a starting point. The lawyer would still make the decision. Business leaders outside legal can borrow the intake pattern using technology that fits their teams. Give one team a clear front door for a recurring kind of request, then make sure the right person can see what came in, what was found, and what remains unresolved.
In brief: Start with intake and keep decisions with the responsible person. Put requests in one place, attach the records needed for a first review, and route uncertain cases to someone empowered to decide. Measure the reliability and timing of the whole resolution, including human review.
What AT&T is actually proposing
AT&T’s legal department has hundreds of lawyers and legal professionals around the world, according to OpenAI’s account. A request from the business can require someone to find the facts, locate an earlier decision, and work out which specialist should see it. Chief Compliance Officer Bill Ryan describes a proposed intake system as a “front door” where people can state their needs in ordinary language. It would classify and prioritize requests, connect the responsible legal professional to relevant precedent and past work, and suggest first steps. AT&T general counsel David McAtee is explicit about the boundary: “The decision making will always remain with the lawyers.”
That system remains a proposed next step whose results are still to be established. AT&T’s October 8 announcement says the companies will design, build, and evaluate legal workflows. LegalEdge, an existing internal team launched in January 2026, handled nearly 100 matters in its first six months. That count predates evidence of results from the proposed intake. It does not show faster legal responses or lower costs.
Why begin at the entrance? People have to work out where to send a request, which documents matter, and what the company decided last time. Scattered details delay the person best placed to answer.
Keep a person responsible for the answer
Imagine a sales manager asking legal to review an unusual contract clause. The request arrives by email with the latest draft but not the earlier customer agreement. A second manager sends background in a chat. A lawyer who handled a related dispute last year knows why one proposed change was rejected, but that decision lives in an old matter file.
A useful intake process would collect the request, documents, deadline, and business contact in one place. It might surface the older matter for a lawyer to inspect, or flag conflicting versions. The sales manager needs to know who owns the request and when to expect an answer. Legal judgment stays with the lawyer throughout.
This is where simplicity matters. A single entry point only helps if people can use it without learning the legal department’s filing system. If the form asks for information the requester does not have, staff will go back to email. If the system silently guesses which contract version is current, it may make the work faster to receive and harder to trust. A clear front door is useful only if it leaves a visible path to the person who can decide.
OpenAI describes what the planned system would do. Neither company’s account provides a before-and-after measure for this new intake. Treat the design as a pattern worth testing; its return remains unproven.
Try the handoff with one kind of request
Test the idea with a request your team sees repeatedly: contract exceptions, policy questions, or vendor approvals. Ask the person who currently receives those requests what is usually missing. Then create a short intake path that captures only what that person needs to assign the matter: requester, decision needed, deadline, relevant files, and a way to flag uncertainty. Define who checks the record before anyone relies on a suggested precedent.
Run a small sample through the old and new paths. Count the full time from the first request through acknowledgement to a usable legal response. Note how often a reviewer had to ask for missing material, how many requests went to the wrong person, and whether the lawyer could inspect the original source behind a suggested past decision. Where confidential files are involved, decide which records the system may read before loading them. An answer built from the wrong customer agreement needs correction, however quickly it arrives.
Give the team a way to correct the intake record. If a lawyer finds that two matters only looked alike because they shared a customer name, that correction needs to stay with the request where the next reviewer can find it. The person routing the next similar question should be able to see why the earlier suggestion was rejected.
The same exercise works outside legal. HR has policy exceptions. Finance has disputed invoices. Operations has incident reports that arrive without the information needed to act. The shape is familiar: a request, scattered context, a person with the authority to decide, and time lost between them. Start by improving that passage of work. Add automation only where the team can see it helping.
AT&T is trying to make its accumulated legal knowledge easier for its lawyers to use while keeping judgment with them. For the broader question of how people and AI share work, see Human-AI Collaboration in Practice. Make the route to the decision shorter. Keep its owner unmistakable.
Research and draft: Mai. Editorial direction: John Lipe.