A practical look at how unaccountable AI use quietly damages the business, and what technology leaders can do about it.
In early 2026, Uber spent its entire annual budget for AI coding tools in four months. This is one of the most financially disciplined companies in the world, and it still did not see the cost coming. Nobody made a decision to overspend. Usage grew faster than anyone was tracking, and by the time finance understood what was driving the number, the money was gone.
Right now, maybe somewhere in your organization, spending on AI is climbing, and no single person can tell you how much or why. Maybe it began with a few engineers trying a coding assistant, a support team adding a chatbot, or an analyst drafting reports with a model. Every choice was small and sensible. Together they have become a cost that grows every month and answers to no one.
The bill that arrives late
Most AI tools charge by usage. You pay for tokens, the small units of text a model reads and writes. The more an organization uses AI, and the more information it sends the model each time, the higher the bill. Unlike a fixed software license, this cost changes every day, and it usually changes upward.
Uber’s own numbers show how this builds. Some engineers were spending between $500 and $2,000 a month, and one ran up $1,200 in a single two-hour session. The company later capped spending at $1,500 per person for each tool. On its own, no single charge stood out. The cost only became clear once it was added up across thousands of engineers.
The pattern is common. A February 2026 survey of 500 finance leaders by DoiT and Sapio Research found that 79 percent of enterprises had experienced AI cost overruns in the previous year. Gartner expects this to continue, projecting that through 2028, at least half of generative AI projects will exceed their budgeted costs, mainly because of weak design choices and limited operational experience.
The cost is also hard to trace. Deloitte documented a large healthcare enterprise whose AI token usage climbed month after month for half a year before the finance team could identify what was driving it. By then it represented about $6 million a year in unplanned spending.
Why the cost climbs
Part of the answer is scale. Newer AI systems, often called agents, do more work on their own. They read, plan, and act across many steps. Research from Microsoft and Stanford’s Digital Economy Lab found that these agent tasks can use around 1000 times more tokens than a single chat message. Gartner places the range lower but still meaningful, at five to thirty times more tokens per task.
The other part is context, the information you give the model to work with. Many teams assume more context is always better, so they send large volumes of documents, code, and past conversations with every request. This raises the cost directly, because the model has to process all of it. It can also lower the quality of the answer. Research from Chroma, which tested eighteen leading models, showed that accuracy tends to drop as the input grows longer, even on simple tasks. A separate and widely cited study found that models often overlook important details placed in the middle of a long input rather than at the start or end.
Sending a model everything, then, comes at a price. It raises the cost of each request and often lowers the reliability of what comes back.
The damage is more than the invoice
Overspending is the visible cost. The less visible ones can hurt more.
The first is wasted effort. Many AI initiatives never reach production. S&P Global Market Intelligence found that the share of companies abandoning most of their AI projects rose from 17 percent in 2024 to 42 percent in 2025. Money and time go in, and nothing ships.
The second is rework caused by wrong answers. When a model returns a wrong answer that looks correct, someone has to catch it, correct it, and redo the work. Gartner estimates that organizations without proper checks on AI output lose 15 to 20 percent of their expected AI return to errors and rework. Under usage-based pricing, a wrong answer often costs more than a right one, because the work has to be paid for twice.
The third is reputation and compliance. In late 2025, Deloitte’s Australian business agreed to refund part of a government contract worth about 440,000 Australian dollars after a report it delivered contained AI-generated errors, including invented citations. The lesson for leaders is direct. When an AI answer cannot be traced back to a real source, the organization carries the risk, not the tool.
What accountable AI looks like
None of this is an argument against AI. It is an argument for treating AI with the same discipline you apply to any other operating cost.
Accountable use rests on a few plain questions. Do you know how much your teams spend on AI, and where? Do you know what information your AI tools are actually working with? Can the answers they produce be traced back to a reliable source? Most organizations cannot answer these questions today, which is exactly why the costs stay hidden until they grow large.
This is where trusted context matters. When AI works from ranked, source-linked context, meaning the specific and relevant information a task needs, connected back to where it came from, spending falls and answers become verifiable. This is exactly the idea behind Ozgar. The aim is simple: AI that leaders can see, measure, and trust.
The technology will keep improving. The price per token will keep falling. But usage will keep rising faster, and these systems will keep taking on more work by themselves. That is why the discipline has to come from leadership, not from the tools.
So before your organization leans harder on AI, it is worth asking one question. Do you know what context your AI is working with? The answer says a great deal about your next AI bill, and about how far you can trust what the technology gives you.
The AI Bill No One Approved
A practical look at how unaccountable AI use quietly damages the business, and what technology leaders can do about it.
In early 2026, Uber spent its entire annual budget for AI coding tools in four months. This is one of the most financially disciplined companies in the world, and it still did not see the cost coming. Nobody made a decision to overspend. Usage grew faster than anyone was tracking, and by the time finance understood what was driving the number, the money was gone.
Right now, maybe somewhere in your organization, spending on AI is climbing, and no single person can tell you how much or why. Maybe it began with a few engineers trying a coding assistant, a support team adding a chatbot, or an analyst drafting reports with a model. Every choice was small and sensible. Together they have become a cost that grows every month and answers to no one.
The bill that arrives late
Most AI tools charge by usage. You pay for tokens, the small units of text a model reads and writes. The more an organization uses AI, and the more information it sends the model each time, the higher the bill. Unlike a fixed software license, this cost changes every day, and it usually changes upward.
Uber’s own numbers show how this builds. Some engineers were spending between $500 and $2,000 a month, and one ran up $1,200 in a single two-hour session. The company later capped spending at $1,500 per person for each tool. On its own, no single charge stood out. The cost only became clear once it was added up across thousands of engineers.
The pattern is common. A February 2026 survey of 500 finance leaders by DoiT and Sapio Research found that 79 percent of enterprises had experienced AI cost overruns in the previous year. Gartner expects this to continue, projecting that through 2028, at least half of generative AI projects will exceed their budgeted costs, mainly because of weak design choices and limited operational experience.
The cost is also hard to trace. Deloitte documented a large healthcare enterprise whose AI token usage climbed month after month for half a year before the finance team could identify what was driving it. By then it represented about $6 million a year in unplanned spending.
Why the cost climbs
Part of the answer is scale. Newer AI systems, often called agents, do more work on their own. They read, plan, and act across many steps. Research from Microsoft and Stanford’s Digital Economy Lab found that these agent tasks can use around 1000 times more tokens than a single chat message. Gartner places the range lower but still meaningful, at five to thirty times more tokens per task.
The other part is context, the information you give the model to work with. Many teams assume more context is always better, so they send large volumes of documents, code, and past conversations with every request. This raises the cost directly, because the model has to process all of it. It can also lower the quality of the answer. Research from Chroma, which tested eighteen leading models, showed that accuracy tends to drop as the input grows longer, even on simple tasks. A separate and widely cited study found that models often overlook important details placed in the middle of a long input rather than at the start or end.
Sending a model everything, then, comes at a price. It raises the cost of each request and often lowers the reliability of what comes back.
The damage is more than the invoice
Overspending is the visible cost. The less visible ones can hurt more.
The first is wasted effort. Many AI initiatives never reach production. S&P Global Market Intelligence found that the share of companies abandoning most of their AI projects rose from 17 percent in 2024 to 42 percent in 2025. Money and time go in, and nothing ships.
The second is rework caused by wrong answers. When a model returns a wrong answer that looks correct, someone has to catch it, correct it, and redo the work. Gartner estimates that organizations without proper checks on AI output lose 15 to 20 percent of their expected AI return to errors and rework. Under usage-based pricing, a wrong answer often costs more than a right one, because the work has to be paid for twice.
The third is reputation and compliance. In late 2025, Deloitte’s Australian business agreed to refund part of a government contract worth about 440,000 Australian dollars after a report it delivered contained AI-generated errors, including invented citations. The lesson for leaders is direct. When an AI answer cannot be traced back to a real source, the organization carries the risk, not the tool.
What accountable AI looks like
None of this is an argument against AI. It is an argument for treating AI with the same discipline you apply to any other operating cost.
Accountable use rests on a few plain questions. Do you know how much your teams spend on AI, and where? Do you know what information your AI tools are actually working with? Can the answers they produce be traced back to a reliable source? Most organizations cannot answer these questions today, which is exactly why the costs stay hidden until they grow large.
This is where trusted context matters. When AI works from ranked, source-linked context, meaning the specific and relevant information a task needs, connected back to where it came from, spending falls and answers become verifiable. This is exactly the idea behind Ozgar. The aim is simple: AI that leaders can see, measure, and trust.
The technology will keep improving. The price per token will keep falling. But usage will keep rising faster, and these systems will keep taking on more work by themselves. That is why the discipline has to come from leadership, not from the tools.
So before your organization leans harder on AI, it is worth asking one question. Do you know what context your AI is working with? The answer says a great deal about your next AI bill, and about how far you can trust what the technology gives you.
Comments
Share this article
Archives
Associated groups