Your AI Bill Is Coming Out of Someone's Bonus

Ed Keisling stopped trusting the word "ready" a long time ago.

Early in his career, his team deployed enterprise software for large financial institutions worldwide. The plan was to ship the hardware and software and let each bank install it. That took far too long. So the team changed tactics. It bought the Sun machines and the Oracle software, built the whole stack in-house and shipped it finished.

"On paper, it should be pretty easy," says Keisling, now chief AI officer at Progress Software. Install the operating system, then the database, and then the application. "The reality is that in practice that was much, much more difficult."

He sees the same gap in AI, with one difference: this one comes with a meter. Many executives still picture "some big easy button," he says. The reality is "way, way, way more complicated than I think anybody is really sort of expecting."

"As little as nine to 12 months ago, the tokens were virtually free," he says. Then the subsidies went away. "That money needs to come from somewhere. It could be employee salaries, it could be bonuses, it could be travel."

That is the token tax in one sentence. (Tokens are the chunks of text a model reads and writes. Providers bill per million.) The tax is not the price of a token. It is who ends up paying for it.

A disclosure: Progress has a stake in this argument. It now describes itself as providing the "context and control" for AI, including control over cost. It bought agentic RAG startup Nuclia in 2025. Several of Keisling's prescriptions point at things Progress sells.

The price fell, but the bill didn't

The first myth is that the token tax means tokens got pricier. They didn't. Keisling himself says newer small models are "orders of magnitude cheaper" than the mid-tier ones. An Accenture survey of 750 executives in 17 countries found respondents expect per-token prices to fall 19% over two years while consumption grows 78%.

So what changed? The wrapper. Flat-rate seats hid usage. Metering them exposes usage. That's why why GitHub moved every Copilot plan to token-based billing on June 1, 2026. In addition, agents are not sitting still. They loop, retry and call other agents, with every step leading to a bill.

So the CFO's question is not "What does a token cost?" It is "What did that token buy?" Few companies can answer.

Every token comes out of someone's budget

Accenture found that only about one dollar in five of token spend can be tied to a financial return. One in three organizations ran out of its annual token budget before year-end. Fewer than one in 12 use chargeback, the practice of billing AI costs back to the team that ran them.

Uber learned this the hard way. By April 2026, it used up its full-year budget for AI coding tools. So, the company limits each employee to USD1,500 a month in token spending per agentic coding tool. As I wrote in October 2026, the meter's fix looked a lot like a seat.

The invoices don't help. The FinOps Foundation notes that provider bills show spend by API key or project, not by business unit or cost center. The tooling around a production agent (vector databases, orchestration, logging) can make up 40% to 60% of feature spend, and none of it shows up on the model provider's invoice.

"That money needs to come from somewhere. It could be employee salaries, it could be bonuses, it could be travel." — Ed Keisling @ Progress Software

Keisling sees the answer in improving visibility. Employees see their own spend; managers see their team's. Budgets also carry thresholds, and crossing one "forces a conversation" about whether the work justifies the money. It's why Progress wired cost telemetry into every agent before release, and it feeds into FinOps systems that allocate spend to teams.

"You should not deploy anything to a production environment or even a test environment unless you have some cost controls in place," he says.

Progress has also shut down internal agents after comparing spend with output. "There is no ROI there," he says. It is the CFO's favorite sentence, said by a CAIO.

The biggest context window is the most expensive one

Volume is half the bill. The other half is what you feed the model.

As context windows grow, teams are tempted to dump everything in. Keisling calls that "probably the most expensive route." You pay for every input token, the model reasons over all of it, and it slows down.

His answer is retrieval-augmented generation (RAG). By fetching only the slice of company data relevant to each query, you use fewer tokens, which leads to fewer tokens billed. It makes sense on paper, but the real-world setup involves real work on indexing and labeling. Progress sells agentic RAG as a service starting at USD700 a month.

The next myth is that fine-tuning a small language model (SLM) on your own data is the less expensive route. Gartner already predicts companies will use small, task-specific models at least three times as much as general-purpose LLMs next year. Keisling agrees on direction, but not on ease.

This is where the Sun servers come back. Getting the data ready takes months, Keisling says, and governance and observability tooling "simply does not exist" as a complete package. Meanwhile, prices fall so fast that an off-the-shelf model can overtake the project.

Routing helps in the meantime. Keisling pins cheap models to routine tasks and saves frontier models for hard ones. He also points to a new class of "decision models" that classify and choose rather than write. TypeSafe AI's Jev launched with USD5 in starter credit, which the company says buys about 120 million tokens.

Shadow AI is a purchase order nobody signed

Cheap steps, however, can still add up to an expensive mistake.

Shadow AI, meaning staff using unapproved AI tools, is where the tax goes off the books. IBM found that high levels of shadow AI added USD670,000 to the average breach cost. Ramp data reported by Axios showed 28% of corporate OpenAI chat spending in February flowed through personal consumer plans.

Keisling reads that as employees telling you which tools they lack. His answer is to stop approving expense claims for unsanctioned tools, and guide staff to approved ones with budgets attached.

"You should not deploy anything to a production environment or even a test environment unless you have some cost controls in place." — Ed Keisling @ Progress Software

Data sovereignty is a major consideration. Keisling argues that even with every privacy box checked, "there's a good chance" some data sent to frontier models ends up training them. Major providers say otherwise in writing, like OpenAI stating that it does not train on API or enterprise data by default. The bigger leak is the consumer account, which is exactly what shadow AI runs on.

Data location matters in Asia-Pacific, where data laws push companies toward local models. Keisling says running open-weight SLMs locally is "the only way that you can 100% protect your data." Accenture's research suggests only about one-third of workloads must stay sovereign. Air-gap that third. Then rent the rest.

Before the next invoice

• Display the meter: Give every employee and manager a live view of token spend.
• Set thresholds that trigger reviews: A breached budget should start a conversation about value, not just create a freeze.
• Instrument before release: No agent reaches test or production without cost telemetry tied to a team.
• Count the harness: Unit cost must include vector stores, embeddings and orchestration.
• Pin models to tasks: Default routine work to small or decision models. Make frontier use the exception, not the norm.
• Kill on ROI: Review every agent against what it produced.
• Map the process first: Many "AI fixes," Keisling says, are really handoff and communication problems.

The size of the box

Keisling's hardest test is for agents that act, such as automation that can change production systems. He scores the risk on four dials: the access the agent holds (read, write, update, delete), the financial impact of its decisions, how fast a bad change can be reversed and the damage to reputation. Then, he says, "put a dollar on it."

It is the Sun server lesson again. Don't take "ready" on faith. Build the box, price it and test it before the agent ships. What matters, Keisling says, is "how big of a box you were able to sort of put the agent in in terms of the tools and the skills and the budget." Most companies are still drawing theirs.

Tokens got cheaper. "Ready" did not.

Image credit: iStockphoto/koumaru