Companies cut back heavy AI use as token bills double, and ADP keeps agents out of payroll

A few months ago, executives at some technology companies treated heavy token use as a mark of a strong employee. The same companies are now cutting that use back, and the arithmetic explains why.

Jue Wang, a Bain consultant who advises large firms on their AI spending, told the Associated Press that their token costs have been "doubling, almost every other month." A company paying $200 of tokens a month for each of 20,000 developers, a scale she says her firm often deals with, has a monthly token bill of $4 million, and that is the figure doing the doubling.

Per-token prices keep falling, so the money is going somewhere else. Bain's analysts wrote that agents consume tokens on multistep reasoning and error correction, and that consumption is climbing faster than prices drop: teams point agents at more parts of the business, the tasks get harder, and most companies want the newest and most expensive models even where an older one would do. In some parts of a business, Bain found, the cost of running agents has passed what the same work costs when people offshore do it.

Vincent Gusdorf, head of AI analytics at Moody's Ratings and author of a new report urging more discipline, named the failure in one line: "It's very easy to create something you don't need with AI."

What the spending bought is thinly documented. In Bain's Automation and AI Pathfinder Survey 2026, with 951 respondents, 40% of the companies that tracked their AI spending recorded cost savings under 10%. Among the companies whose AI investments fell short, 90% plan to raise their AI budgets next year.

Perplexity's desktop agent reached Windows on July 28. Personal Computer opens files on the machine it runs on, edits spreadsheets, and works inside Microsoft Office. It ran on Mac first, and Windows is where most business paperwork is done.

The price is $200 a month, on the Max and Enterprise Max plans. Behind one window, Perplexity hands each task to one of more than 20 outside models from OpenAI, Anthropic and Google, so the person using it never chooses a model. The company says the version sold to businesses connects to more than 400 other systems, Salesforce and Snowflake among them, and it gives an owner a log of what the agent did and a switch that stops it.

The reasoning itself does not run on the PC. Tech Times reported that Perplexity runs the models on its own servers, so a spreadsheet the agent opens on a company laptop is sent out to Perplexity and comes back changed.

"We cannot be mostly right." That is how Roberto Masiero, ADP's first chief AI officer, described the standard for AI at a company that processes pay for one in six American workers, in an interview with InformationWeek. ADP will not let AI agents run payroll on their own.

Masiero said the pressure in payroll compliance is to keep the AI on a path where it does not invent things and does not stray outside the limits set for it. So ADP sorts the work by what a mistake costs. Where an error means redoing a draft, an agent goes ahead. Where it means an underpayment, a corrected filing and a deadline set by a tax authority, a person signs.

ADP's Assist agents, introduced on January 28, find variances in a payroll run, guide tax registration, answer policy questions and produce reports, on a platform the company says covers 1.1 million clients in 140 countries and 42 million wage earners. The agent proposes the correction; a payroll practitioner is the one who accepts it.

Atlassian's Teamwork Lab showed the same email to 961 US knowledge workers between March 30 and April 7 this year, a summary of a business proposal from the same imagined colleague, and changed one thing: whether the reader was told AI had helped write it. When they were told, the odds a coworker labeled the writer lazy went from 3% to 24%, and they were 24 percentage points less likely to recommend that person for a high-visibility project.

Most workers are not hiding anything. Atlassian found 94% of knowledge workers used AI in the past month and nearly 80% say they are open about it with managers and peers. Anthropic's education lead says teaching people specific tricks for using AI has stopped working. The penalty nearly disappeared, Atlassian found, in teams whose leaders used AI visibly themselves.