Sullivan & Cromwell, one of the oldest names in corporate law, built an AI tool that reads a draft agreement and flags the problems a junior associate would spend a night hunting for. It is a narrow, unglamorous job, and that is exactly why it works. The interesting part is not that a model can read a contract, but what the firm had to wrap around it to trust the answer.
A LAW FIRMSullivan & Cromwell built an agent that red-flags draft agreements
Sullivan & Cromwell, a 150-year-old Wall Street law firm, worked with OpenAI to build a tool it calls Agreement Analyzer. The job is specific: take a draft contract that a client or counterparty has sent over, and surface the clauses that create risk. Missing indemnities, odd termination triggers, definitions that quietly contradict each other. The kind of reading a junior associate does slowly and a partner cannot afford to do at all.
The mechanism matters more than the model. A general chatbot asked to review a contract will summarise it and sound confident. What the firm built instead points the model at a defined checklist of issues the firm already cares about, so the output lands as a list of flags tied to specific sections rather than a vague opinion. That framing is what makes a lawyer willing to act on it, because the tool is answering the firm's question, not inventing its own.
The result is not fewer lawyers. It is lawyers who start from a marked-up draft instead of a blank one, which moves the expensive human judgment to where it actually belongs.
Carvana stopped silencing alerts and started killing their causes
Carvana, the used-car retailer known for its glass vending-machine towers, runs software that throws off production alerts all day long, the automated warnings that tell engineers something might be breaking. The usual response to alert overload is to mute the noisy ones, which is how a real failure ends up buried under a hundred ignored pings. Carvana used Claude Tag, Anthropic's tool for labelling and grouping streams of events, to do the opposite.
Rather than treating each alert as its own fire, the system clusters related alerts and traces them back to a shared root cause. Fix the one underlying issue, and the dozens of downstream warnings it was generating simply stop. That shift, from firefighting symptoms to closing causes, cut the firm's production alerts by 56 percent.
The number is the evidence; the method is the point. Fewer alerts here does not mean problems were hidden. It means the thing generating them got fixed, which is the version of 'quieter' an on-call engineer actually wants.
Meta's new open model runs on a Mac, no cloud required
Meta and Apple released Muse Spark Open, a model with openly published weights that runs locally on Apple silicon. 'Open-weight' means the model's actual parameters are downloadable, so it runs on your machine instead of calling a server. Running on Apple silicon means a modern Mac can do the work on its own chip, with nothing sent to anyone's cloud.
For a business, the consequence is about where data goes. A model running on the laptop can read sensitive documents, client records or draft contracts without any of it leaving the building, which removes the single biggest objection most regulated or privacy-conscious operators raise about AI. It also removes the per-token bill, because there is no API call to pay for.
The trade is capability. A model small enough to run on a laptop will not match the frontier systems on the hardest reasoning. But for drafting, summarising and classifying, a local model that costs nothing per use and never phones home is a genuinely different option than it was a week ago.
Expedia's AI chief: the model is swappable, your data is not
Xavier Amatriain, ten months into running AI at the travel company Expedia, laid out a buying principle worth more than most product launches: stay deliberately flexible about which models you deploy, and treat your own data as the thing that actually compounds. Models improve and get cheaper every few months, so wiring your business permanently to one is a bet against the whole market.
The practical version is architecture. If your systems call models through a layer that can route to whichever one is best or cheapest this quarter, swapping is a config change rather than a rebuild. What you cannot swap, and what competitors cannot copy, is the proprietary data that makes any model's answers specific to your business.
For a smaller operator the lesson scales down cleanly. Do not fall in love with a vendor. Fall in love with keeping your own data clean, labelled and portable, because that is the asset every model you ever use will depend on.
Models are fungible, data isn't
Reply and tell me the one task you keep meaning to hand off, and I'll tell you how we'd build an agent for it at Openhour.
