Most company AI money is spent outside IT, and ServiceNow sells a help desk that starts in a day

Corporate AI spending has also doubled in less than a year, to 3.3% of revenue. For a company with $200 million in sales, that is about $6.6 million a year. About $5.3 million of it is in budgets run by marketing, finance, operations and sales.

Most of it arrives in pieces small enough to approve without a business case. The CRM renewal comes with an AI add-on, and the marketing team pays for an image tool by card. Nobody adds the pieces up, and nobody checks them against a result.

The money is in the right hands, though. Companies that get value from AI put 70% of the effort into people, organization and processes, and only 30% into technology, data and algorithms. Only the head of accounts payable can change how an invoice gets approved. So the return on an AI tool in that team depends on a decision made in that team. The spending moved out of IT, and the responsibility for its return moved with it.

For example, a finance director could ask every budget owner for one line per AI tool: what it costs a year, which piece of work it changed, and which number moved, such as days to close the month or invoices handled per clerk. Expect about a third of the lines to have nothing in the last column. Those tools are the first conversation at renewal, and the lines with real answers are the case for spending more.

Griffin, from the AI video start-up Tavus, follows faces, pauses and tone of voice while it talks. After a one-minute live call, 48% of the people who talked to it believed it was a real person. Earlier systems, including Tavus's own, reached 2% at most. For now, only selected testers can use a lighter version.

Once tools like this are widely available, a face on a screen will prove very little. A final interview for a remote job will need a check that does not depend on the face, such as an identity document verified separately.

Flow answers staff questions in Slack, Microsoft Teams or email using the company's own information, and it completes routine requests inside the same conversation. When a person solves a new kind of request in a chat, one click lets Flow handle it the next time. Buyers need no ServiceNow platform and no implementation project, and they pay by use.

Until now, buying ServiceNow meant a long project. Its own light version gives every buyer of large service software a fair question at the next renewal: what would the same job cost if it started in a day?

Sidekick, Shopify's AI assistant, could already edit themes and code. Canvas lays out every page of the store in one workspace and shows each change on the store's real code as Sidekick makes it. A custom store takes about 20 minutes instead of weeks.

A new product page layout or a size guide used to be a job for a freelancer. Now it is a request in a chat. Canvas is open only to some stores for now, and blocks from apps, such as product reviews, do not yet appear in themes made in it.

Argon is the first new generation of Google's flagship model since Gemini 3 last November. It can find, confirm and fix serious flaws in software on its own, and it found a critical flaw in hospital software used worldwide that earlier models had missed. Paid API customers and Google AI Ultra subscribers get it next.

So the first work this model does will reach most companies as security updates for software they already run. Expect more urgent patches from vendors in the coming months. The risk is in the days between a patch's release and its installation.

Developers at more than 300 organizations asked coding agents to fix a screen and attach before-and-after pictures for a colleague to review. GitHub accepts those pictures only through its website. The agents work from the command line, where that upload does not exist. So many agents created a public folder of code, often in the developer's personal account, put the pictures there and linked to them. The pictures showed customer billing records and unreleased features. Security scanners check text and let images through, so nobody noticed.

The agents did what they were asked. Each one met a blocked path and found another, because nobody had given it an approved place for its pictures. Before an agent starts a job, someone should decide where its output may go. Otherwise the agent decides.

Most plants see energy as one bill a month. AI software reads the meter on each machine and shows where the waste is: compressed air leaking overnight, motors idling between orders. One part of the bill, the capacity charge, depends on what a plant draws in the grid's busiest hours, and it is climbing where data centers multiply. By July, one brick maker was paying $12,000 a month for it, up from $1,600. The same software can move the heaviest jobs out of those hours.