Throttle by friction

Friction used to set how much work got done. Asking was costly, notes were rare, and a contract could sit unread. AI made the start cheap for the person and left the company with someone still signing the result. That split is where the new work appears. For most companies it is already larger than the work removed. The tokens, and the hours spent checking them, are the tax for staying in that market. An early gain covers the tax later only when it is spent on automation that takes the signature off ordinary work. A buyback, a dividend, or a project that fails to clear work leaves the throttle where it was.

The open side is already visible in two companies' own accounts. Aetna's chief digital and technology officer said on August 20, 2026 that a cost tool, a member asking in ordinary language what a procedure will cost, was used by less than 10% of people in the old model and is now used by almost 70%, before the procedure. Novant Health's chief AI officer said on October 15, 2026 that a little over 40% of its surgery-agent conversations happen after hours and on weekends. Each figure is one executive describing his own company. Asking got cheap. The company still owes the answer.

A person pays for a bad sentence with a minute. About two thirds of all writing messages on consumer ChatGPT ask it to rework text the user supplied, including editing, translation, and summaries. Writing is about 40% of work messages on those plans, which is a different base, so the two shares do not multiply. In a randomized trial across 66 firms and 7,137 knowledge workers, people offered Microsoft 365 Copilot spent 1.4 hours a week less on email, 12%. People who used it spent about 2 hours less, 17%. Meeting time did not change.

A company prices the same miss as a customer, a filing, or the brand. A costing built on OpenAI's study of expert tasks put an expert's review at about 113 times what the model spent producing the work. That multiple is this project's derivation, not a figure OpenAI published. In an investor's survey of 100 security chiefs at companies with more than 1,000 employees, 52% were comfortable with an agent in security operations acting with a person's approval. In April 2026 logs of Google's Gemini products, attempts to automate a non-routine task end to end were under 10% of those conversations. Inside a queue that already existed, the machine does clear work: across 5,172 support agents at one software firm, issues resolved per hour rose 15%.

People moved first, and the firm questions are different questions. In May 2026, 52% of US employees had used AI at work at least a few times a year, up from 21% in May 2023. Daily use rose from 4% to 15% over that span. In May 2026, 47% said their organization had integrated AI, which is what employees believe about the firm. The Census asks the business. In the survey period ending October 18, 2026, 24.6% of US businesses said they had used AI in any function in the past two weeks, up from 17.3% in the period ending November 30, 2025. The comparison that holds the worker still is Danish. In 11 AI-exposed occupations, 41% of workers had used a chatbot at work where the employer did nothing, against 93% where the employer encouraged it, supplied a tool, and trained people. In Microsoft's 2024 survey of knowledge workers, 78% of AI users had brought their own tool. The firm, in that sample, was not the one that went first.

What comes through the open side is a record someone has to own. In a Japanese government survey of managers, around 70% said their firm uses generative AI to help prepare minutes and emails, the most common kind of work they were asked about. Among legal professionals who already use generative AI, a Thomson Reuters survey in October and November 2025 found the regular uses were legal research, 80%, document review, 74%, and document summarization, 73%. Among German firms with 20 or more employees that use AI, customer contact is the most common use, 72%, and the legal or tax department is still uncommon, 8%. Of US firms that use AI, 57% use it in only one to three of fifteen functions. Bitkom's 2026 survey, the source of the German shares, also records that 57% of firms with 20 or more employees use AI, and that no AI-using firm said it fully used the potential.

The new reading is weakly measured, and it points the same way. In September 2025, 40% of 1,150 US desk workers had received AI workslop in the past month, at about 2 hours each. On high-AI engineering teams, developers merged 98% more pull requests and review time rose 91%, with no company-level improvement in that vendor's telemetry. In a vendor survey of employees who use AI at large companies, nearly 40% of the time they said AI saved them went back into fixing the output, and 14% reported a clear net. These are not the Copilot trial.

The week stays about the same length. Workers told the St. Louis Fed that AI saved 2.1% of US work hours in August 2026, and assisted 5.3%. Danish registers rule out effects on earnings or hours above 2%, while adopters report saving about 3% of work hours. US finance chiefs said AI raised their firms' 2025 labor productivity 1.8%, while their own estimates of the effect on revenue and headcount imply 0.6%. The authors put most of that gap down to revenue that arrives later.

Token burning behaves like a tax because the rate falls and the take rises. At a fixed capability, the price per token fell 9x to 900x a year depending on the benchmark, as of March 2025, and Epoch's later estimate puts the cost of a task down about 13x a year. Bain, citing Wells Fargo, found the price per token had halved from December 2024 to December 2025 while tokens consumed grew 4.5x, and the cost of a task often flat. On Ramp, the blended OpenAI and Anthropic price was $0.64 per million tokens in the week ending October 11, 2026, down 25.4% from $0.86, while token spending rose 23.0%. Uber used up its 2026 AI coding budget in about four months and capped each employee at $1,500 a month per tool. The bill is the cost of operating where the other side has already opened the door.

A short-term gain makes sense as profit when it is reinvested in automation that raises the throttle. In McKinsey's survey of May and June 2026, 89% of respondents reported regular AI use in at least one function, 80% said their own productivity improved, and 37% said AI had contributed to their organization's earnings. About 6% attributed 5% or more of earnings to AI. Nearly three quarters of that group had redesigned the workflow, against about a quarter of the others. These are people reporting on their own organizations. BCG classes 5% of companies as future-built, 35% as scaling, and 60% as laggards, and associates the first group with 1.7x the revenue growth and 3.6x the three-year shareholder return of the laggards. Those are associations in a self-rated sample. This site calls that thin set the Joneses. Only 12% of large organizations in KPMG's pulse consistently weigh AI value against cost, and 49% had scaled back, delayed, or paused agent deployments because expected cost outweighed value.

Gartner forecast on September 9, 2026 that, by 2027, three quarters of organizations that take AI productivity as cost savings will be eclipsed by competitors that reinvest those gains in innovation, modernization, and upskilling. That is a forecast. The reinvestment that meets this asymmetry is narrower. It is automation that clears ordinary work without a person on every line. Hand the early surplus back as a buyback or a dividend, and the owners have the cash while the building keeps the old signature. No study follows AI savings into buybacks or dividends. Spend the surplus on a project that does not clear work, and the firm has paid for the tokens and still has the person. As CIO Dive reports an S&P survey, 42% of companies were abandoning most of their AI initiatives, up from 17% a year earlier, and the average organization scrapped 46% of its proofs of concept before production. The article does not print the question.

For most companies, AI creates more work than it removes. The work removed is a shorter email week, a faster hour on a queue that already existed, and about 2% of the workweek in the registers and in the hours workers say AI saved. The work created is what the other side can now start: the note, the question after hours, the file someone can finally read, the second pass that throws off something a person at the company still has to sign. The first is measured. The second grows with their adoption. The easy part got cheaper. The signature did not. Firms that spend the early gain on automation can meet it. Firms that return the gain, or lose it in a project that fails, meet it with a human throttle, and the tax keeps coming due.