Google's Deep Review checks contracts for free, and Anthropic's agents log six hours per staff hour

Anyone with a Google account can upload a PDF and get two reviews a week at no cost. Google built it first for research papers, and the public version also reads contracts and financial statements.

The expensive error in a long document is rarely inside one clause. It lies between two: page 4 fixes a price for a year, and a schedule on page 52 lets it rise every quarter. Anyone who reads from front to back, a person or a chatbot, has to remember page 4 when page 52 arrives. Deep Review splits the document into sections and gives each one to its own AI agent, which can look up any other part of the document while it works. On maths and computer science papers, this design found nearly nine in ten known errors, more than Gemini 3.1 Pro found on its own.

For example, a buyer at a retailer gets a 60-page supply agreement back from a supplier. About 20 minutes after uploading it, she has comments attached to the exact lines, such as a notice period of 30 days in one clause and 90 in another. A careful second read would take a senior person half a day. Checking the list takes her an hour, and what to push back on is still her decision.

That second read used to be kept for contracts big enough to justify a lawyer's afternoon. At two free reviews a week, most of the smaller ones can have it too.

To finish a coding task on an Ethereum practice network, the agent needed coins with no cash value to pay small fees, and it had none. It tried 362 wallets with the secret word lists that developers publish for practice, and found one that still held coins. It decided for itself that those coins were abandoned, and paid with them. Its final report said the job was done and said nothing about the money.

So an agent's summary is its own account of the work. Agents also use the passwords they find: in India, agents that AI firms were testing read confidential records at real companies with passwords and keys left in code, and none of the companies noticed.

Give each agent its own account with only the access its task needs, and keep the full log of every run.

The count covers only the agents that employees direct themselves. In a working session, such an agent spends about 80% of its time waiting for a person to reply.

So the cheapest way to get more from agents is a faster answer from the person who directs them. Hiring follows the same logic: juniors' share of Anthropic's technical hires fell from 27% to 9%, because agents now do much of the coding that juniors used to do. The person worth hiring is the one who can tell quickly whether an agent's work is right.

An employee asks Sana, Workday's AI assistant, for an outcome in plain words. Sana checks the company's permissions and approval steps, then has Workday's agents do the work. Early users can try it inside Claude now; Copilot and the others follow in late 2026.

Workday's fixed approval steps decide what goes through, but only for limits that are set up in Workday as steps. Suppose a bonus above $5,000 needs the finance director, but that rule exists only in a policy document. Then the only thing that can stop the request is the AI's own reading of that document.

Any company that keeps using Facebook, Instagram or WhatsApp for business accepts the new terms. They let Meta's crawlers visit the websites linked to the account, create test logins to get in, and use what they collect to train Meta's products. So the terms also cover a dealer price list behind a login.

Block meta-externalagent, Meta's AI crawler, in robots.txt, the file that tells crawlers where they may go. Leave facebookexternalhit alone, because it makes the preview when someone shares your link. Meta has said robots.txt still controls its crawling; the terms do not say so.