Only 26% of firms can say what AI costs, and the ones who can see returns five times as often

Twenty-six percent of companies can say clearly what their AI costs them. That is from KPMG's latest quarterly survey, which asked more than 2,100 senior leaders across 20 countries between 28 April and 25 May.

The same survey separates that group from everyone else on results. Among the companies with full visibility of what AI costs to run, 15% report established returns; among those without it, 3% do. Five times the rate, from the same technology. In the same round, 49% said they had cut back an agent rollout when its running costs outran the value it produced.

Seeing the number is harder than it sounds. The spending is split between a cloud bill and software subscriptions held by a dozen teams, and neither one labels it AI.

That second part grows without anyone approving a purchase. The software spending platform Cledara, reading payment data across thousands of vendors, puts AI platforms at 10% of company software spend, up from 3% twelve months earlier. Cledara also puts standard renewal increases at 8% to 12% a year, and the most aggressive vendors at 15% to 25%. Vendors apply those increases to tools the finance team pays for today, with the AI feature built into the renewal.

Three months earlier, 13% of the companies in KPMG's survey described AI as part of everyday work. In this round, 22% did.

Since 19 August, chatting with Replit's AI has cost nothing extra. Free Mode runs conversation, planning and small tasks on GPT-5.6 Luna, the cheapest model in OpenAI's GPT-5.6 family, and none of it draws on the AI credits a subscriber buys for the building work. Replit's plans are $20 a month for Core and $100 for Pro, and the everyday allowance inside them is now up to 30 times what it was.

For someone who has never written code, the rounds of describing, testing and correcting are the part that used to consume the credits, and that part is now included. A small internal tracker, a form, a shift calculator: the experiment costs a subscription and an afternoon.

Where it ends is where it ended before. Building and running the finished thing still spends credits, and anything the business depends on, an app holding customer records or touching money, still needs a person accountable for who can see the data and what happens when it breaks.

Free Mode is the default for Core and Pro subscribers, and it still requires one of those two subscriptions.

Which model should a company standardize on, GPT, Claude or Gemini? The price of running one has fallen so far that the choice is close to reversible. Stanford's AI Index put the cost of a model performing at GPT-3.5's level at $20 per million tokens in November 2022 and $0.07 by October 2024, roughly 280 times cheaper in under two years.

Every correction a claims handler makes to a drafted letter, every answer marked wrong, every outcome recorded next to the recommendation that produced it: that record exists only in the tool where the work was done, and no provider sells it.

So the question at renewal is what a company would lose besides its login if it changed provider next quarter. If the answer is nothing, the model was doing all of the work and the cheapest one wins. Switching model providers is an afternoon of configuration. Getting three years of your own corrections out of a vendor's system is whatever the contract said when it was signed.

OpenAI's new Mac plugin lets ChatGPT read, search and send a person's iMessages and text messages. Reading the Messages history on a Mac takes Full Disk Access, the permission that covers every file on the machine, not only the chats: the downloads folder, the contract a supplier sent last week.

On a work laptop, that is the full record of who an employee talks to and what they were sent. Apple, whose Messages app this reaches into, is suing OpenAI in a separate dispute over trade secrets.

Full Disk Access is granted app by app, and on a managed Mac an administrator can allow or withhold it for this one, before anyone on the team decides for themselves.

Most of the leaders Deloitte surveyed expect AI agents to change jobs at their company inside 18 months. One in five say the company is ready to redesign how the work is done. Deloitte asked more than 500 US business and technology leaders.

Redesigning the work means more than adding a tool to it. An agent that answers customer emails without a person reading them changes who checks the answers, who is told when one was wrong, and how the wrong ones come back. In most companies those jobs belong to nobody by name, because the old process handed them to whoever was doing the task.

The same question is open at the other end of the chain. An independent report graded OpenAI, Anthropic, Google, Meta and xAI on their readiness to shut down a model that behaves badly, and found that most have published no rule for when they would do it or how fast. Anthropic, which built its reputation on safety, is among the ones that have not.

A 20-second video clip with sound, made from one photo or one line of typed description, now costs nothing. Black Forest Labs, the maker of the Flux image generators, released Flux 3 free to use.

That covers much of what a small firm buys visuals for: a product clip for a listing, a short explainer for new staff. The sound is generated along with the picture, so a stock music track is not part of the cost either.