Software's new AI tax, an autonomous agent breaches Hugging Face, and AI hiring tools show bias

Enterprise software vendors are adding AI features to their products and charging 20 to 37 percent more when contracts renew, whether or not customers ever use those features. The spend-management firm Tropic calls it the AI tax.

The underlying cost of AI keeps falling. Baytech Consulting estimates that the infrastructure costs vendors themselves pay dropped 40 to 70 percent over the past two years. The vendors are keeping the savings and charging more on top.

The increase rarely arrives on a line marked AI. It comes through a forced move to a new product tier, a switch to pricing by credits consumed, or a bundle that folds AI into the base plan with no option to decline. Gartner expects business software spending to grow 15.2 percent in 2026, and says companies are setting aside about 9 percent of their entire technology budget just to absorb price increases on software they already run.

Microsoft's July 2026 changes show the shape of it. Business Standard rose from $12.50 to $14 per user a month and the E3 plan from $36 to $39, with AI tools such as Copilot Chat folded in, separate from the $30 per user the full Copilot still costs. Adobe's Pro tier went to $70 from $60, and Atlassian raised some cloud plans by up to 10 percent.

Before Levi Strauss switched on a single AI agent, it spent years on the unglamorous part: retiring more than 90 legacy systems and consolidating nine customized ERP platforms into one standardized global system on Microsoft's Azure cloud. The program began in 2019 and is due to finish across 110 countries by mid-2027.

Only with that base in place did the automation follow. Levi's says agents now built into the system process sales orders, check vendor compliance, and capture invoices, the routine back-office work that ties up staff time. That order matters, because an AI assistant is only as reliable as the data underneath it, and Levi's spent seven years making that data consistent before letting software act on it.

For the thousands of companies still running SAP's older ECC system, the deadline is fixed: SAP ends mainstream support at the end of 2027. At the close of 2024, by Gartner's count, only about 39 percent of ECC customers had licensed the replacement, and a full migration commonly takes 18 to 36 months.

Over a single weekend, an autonomous AI agent broke into the production systems of Hugging Face, the platform that stores more than 45,000 AI models for over 50,000 organizations, and ran more than 17,000 actions on its own. It entered through a malicious dataset, harvested credentials, and moved from one internal system to the next, reaching internal datasets and the access keys its services rely on before the company caught and evicted it.

When Hugging Face's security team fed the attack logs to the commercial AI models it normally uses, the models refused. Their safety filters could not tell an incident responder from an attacker, and blocked the real attack data as harmful. The team switched to an open model it runs on its own machines, Z.ai's GLM 5.2, and rebuilt the full timeline in hours instead of days.

The attacker faced no such limits. A hostile agent obeys no usage policy and works at machine speed, and Hugging Face said the campaign matched the 'agentic attacker' scenario the industry had been forecasting. Any business that keeps data in the cloud now shares that exposure; Hugging Face has told users to rotate their access tokens and check recent account activity.

The next time someone applies for a job, an AI may read the résumé before any person does. A study published in July suggests it may be less even-handed than a human.

Researchers at Princeton and the University of Chicago put ChatGPT, Claude, and Gemini through a hiring game first designed to test people, asking each model to fill 20 jobs with candidates from four invented ethnic groups. The models sorted candidates into narrow roles by group far more sharply than the human players had, about 65 percent higher on the study's segregation scale, where 2 means every group is confined to a single kind of job. Humans scored 0.84; OpenAI's o3 reached 1.83, near the maximum. One coauthor, Ryan Liu of Princeton, said the models 'really are eager to create generalizations from limited data.'

The legal responsibility belongs to the employer, not the software vendor. New York City already requires yearly bias audits of automated hiring tools, and the EU's AI Act sets obligations for high-risk hiring systems from August 2, 2026, a deadline a proposed delay has not yet moved. In the Mobley v. Workday case, a US court has let age-discrimination claims proceed on the theory that hiring software can act as the employer's agent, over a system Workday says handled about 1.1 billion applications.

Halfway through the 2026 season, Major League Baseball switched off a feature on the dugout iPads. The change, set out in a June 11 memo, targets the moment a tool stops informing a decision and starts making it.

By the league's estimate, up to a third of teams had loaded a custom tab that ran outside programs, including AI that recommended pitch selection and substitutions in real time. The tablets keep their video and league data; the recommendation programs are switched off. One club, according to a former player, had paid several hundred thousand dollars for its program.