Agentic AI uses 5 to 30 times more tokens per task, and 2026 AI spending is up 63.4%

An agentic AI system uses between 5 and 30 times more tokens to complete one task than a chatbot uses to answer one question, by Gartner's count. Neither the model nor the published per-token rate has to change for that to happen. Finishing a task means reading files, planning steps, calling other software, checking the result and often doing part of it over.

The tokens themselves keep getting cheaper to produce. Gartner forecast in March that by 2030 the cost to providers of running a model with a trillion parameters will be more than 90% below what it was in 2025. In the same forecast Gartner says those savings will not be passed on to business customers in full, and that token use is growing faster than token prices are falling, so what companies pay to run these systems goes up.

The per-token rate is the part a supplier publishes. The token count belongs to the buyer: how many steps a workflow takes, how much of a contract or a case file the model re-reads at each step, how often it tries again. A department can add no new users and sign no new contract, and still pay more this quarter than last.

Gartner's July forecast puts worldwide company spending on AI models and platforms at $64 billion this year, up 63.4% from $39 billion in 2025. Generative models alone account for $23.4 billion of that, more than half of what companies spent on all AI platforms and models put together last year.

The annual review a manager writes about a member of their team is now often drafted by a machine before a person edits it. More than half of US workers use AI for that kind of writing, their performance reviews and their work emails included.

Gallup's latest quarterly survey found 52% of US employees using AI at work at least a few times a year, up from 27% two years ago; writing and editing is the most common use, at 51%. In a survey by the language learning company Preply, 63% of workers said they had used AI to avoid a difficult conversation at work.

A review and a warning letter are the documents read again months later, by an investigator or a lawyer, if the employee contests what happened. The manager's name is on them either way. The same holds for an email that gives a customer a date, a price or an assurance.

Most employers have set no rule for which messages a person must still write.

Since 2 August, a company anywhere in the world that shows AI-made content to people in the EU has to say so visibly, under Article 50 of the EU AI Act, with fines up to 15 million euros or 3% of worldwide turnover.

The business using the tool owes the label, not the lab that built it. A deepfake, meaning AI-made images, video or audio of a real-looking person or event, must be disclosed at the latest when a person first sees it. AI-written text published to inform the public on a matter of public interest carries the same duty, unless a person reviewed it and took editorial responsibility for what it says.

The Commission's guidelines, adopted on 20 July, put clearly impossible pictures, a dragon or unaided human flight, outside the definition.

Chegg's revenue for the first quarter of 2026 was $63.3 million, down 48% in a year, and its academic services line fell 57% to $45.7 million. The decline in its traffic from non-subscribers widened from 8% in the second quarter of 2024 to 49% by January 2025, as Google's AI Overviews answered homework questions on the results page.

Pearson sells into the same market and its profit rose 14% in the first half of 2026. Its largest business, Assessment and Qualifications, produced £803 million of revenue and £157 million of adjusted operating profit in that half. It sells the exam, the marking and the certificate an employer or a licensing board accepts.

AI can explain the syllabus, work every practice problem and write the study plan, free and at any hour. It cannot issue a score that a hospital's credentialing office or a state licensing board will treat as proof, because someone has to be accountable for that number being right.

Other industries divide the same way, wherever the customer pays less for information than for a named party standing behind the result. Pearson, in its own first-half presentation, says AI now handles 40% of its customer interactions.