The Real Cost of Agentic AI: Why Token Bills Break Security Budgets

Key takeaways

• Agentic AI does not spend tokens like a chatbot. It runs on its own, in loops, so the bill can climb fast and without warning.
• Gartner expects agentic models to use 5 to 30 times more tokens per task than a standard GenAI chatbot, and warns that total inference costs will rise even as unit prices fall.
• Cost is now a top reason projects get cancelled. Gartner says over 40% of agentic AI projects will be scrapped by the end of 2027, with escalating cost named as a driver.
• On our own platform, up to 62% of tokens have been served from cache in a single company's usage. That is what keeps the running bill in tens of dollars rather than thousands, and the cached share tends to rise as sustained usage grows.

Everyone is buying AI agents. Almost nobody is costing them.

There is a lot of noise about AI agents in security right now. Autonomous triage, agents that write and send training, agents that investigate phishing reports on their own. Most of the conversation is about what they can do. Far less of it is about what they cost to keep running, day after day.

That gap matters, because agentic AI has a very different cost shape from the AI most teams are used to. A model that scores an email or flags an anomaly runs on your own compute and costs roughly the same every month. A chatbot costs a bit more, but a person is typing, so spend moves at human speed. An agent is different. It plans, it calls tools, it checks its own work, and it loops until the job is done. Nobody is sitting there slowing it down. That is the whole point of an agent, and it is also where the bill comes from. Most AI driven attacks still land on a person, which is why most AI attacks are human attacks.

SecurityWeek made the scale of it concrete in its piece on the AI token costs that can break cybersecurity: a single complex, multi stage incident investigation may burn millions of tokens in minutes. When the work is autonomous, the token meter does not wait for anyone.

Why agentic AI costs behave differently

It helps to think about it in three layers, because they price very differently.

• Traditional machine learning. Runs on CPU or GPU you already pay for. No token cost.
• Generative AI (chatbots, copilots). Token cost is real but bounded, because a human sets the pace.
• Agentic AI. Token cost is open ended. The agent decides how many steps to take, and each step spends more tokens.

This is not a rounding error. Gartner puts it in numbers: agentic models require between 5 and 30 times more tokens per task than a standard GenAI chatbot, and can perform many more tasks than a human using GenAI. Same question, many times the spend, because the agent works the problem instead of answering once.

What the analysts are actually saying

If you only take four numbers into your next budget meeting, take these. They are all from Gartner, with the source and date listed at the end.

• By 2028, AI coding costs will overtake the average developer salary, as token consumption keeps climbing. Gartner is clear that costs will keep rising as model pricing goes up, not down.
• Unit prices fall, total cost still rises. Gartner expects inference on a one trillion parameter model to cost providers over 90% less by 2030 than in 2025. In the same note it warns that as consumption rises faster than prices fall, overall inference costs are expected to increase, and that provider savings will not be fully passed on to customers.
• Over 40% of agentic AI projects will be cancelled by the end of 2027, and Gartner lists escalating costs as one of the main reasons, alongside unclear value and weak risk controls.
• At least 30% of GenAI projects were set to be abandoned after proof of concept by the end of 2025, again with escalating costs named as a cause.

The real world numbers are just as blunt. One company ran up a 500 million dollar bill in a single month simply by failing to put usage limits on employee licences. A security vendor burned through more than 1 million dollars in tokens while testing a model against its own source code. Uber's CTO spent the entire 2026 AI budget by April. These are not small, careless teams. This is what happens when autonomous systems are let loose without a cost control in place.

From our CEO

"Everyone is excited about AI agents right now, and they are right to be. But not enough people are asking the boring question, which is what it costs to run them every single day. We built our framework so the cost cannot run away from you. If a security vendor cannot tell you how they keep token spend under control, that is a red flag, not a detail."

Ozan Ucar, Founder and CEO of Keepnet

Why security teams are more exposed than most

A marketing team running an AI agent now and then is one thing. A security operation is another. Security agents are meant to run around the clock. They pick up alerts at 3am, they chase phishing reports while your team sleeps, they investigate without being asked. That is exactly the behaviour that drives token spend, and it is also the behaviour you want. You do not want your night shift agent to stop and wait for someone to approve the next step.

So the risk is not that agentic AI is too expensive to use. The risk is the surprise. A bill with no ceiling forces bad choices later, like switching features off to save credits, or pulling an agent back right when it was starting to help. SecurityWeek makes the same point. Teams end up making operational compromises they never planned for, purely because of cost.

How Keepnet keeps the cost down

We designed our AI Agents framework with this problem in mind from the start, rather than bolting on a cost dashboard after the fact. The single biggest lever is caching.

Agentic systems repeat a lot of the same information on every step. The same policy, the same context, the same instructions, over and over. If you pay full price for all of that on every call, most of your bill is waste. So we cache the parts that do not change, and only pay full rate for what is genuinely new on each step.

Here is what that looks like on our own platform. Across the agents, up to 62% of tokens have been served from cache in a single company's usage, instead of hitting the model at full price. Every cached token is one you do not pay full rate for, and at that share the running bill stays in tens of dollars rather than thousands.

One thing worth being clear about. That 62% is an example of what one company's usage looks like, not a fixed rate and not a platform average. Cache efficiency depends on how the agents are used. Because agents reuse the same context on every step, the cached share tends to climb as sustained usage grows, since more requests hit the same cached context inside the caching window.

The point is not the exact figure, which will move with usage. The point is the shape. When most of your tokens come from cache, the bill stays in tens of dollars where an uncontrolled setup would be in the thousands. It is the difference between a cost you can plan around and one that ambushes you at the end of the month.

From our Lead AI Engineer

"Agentic systems send a lot of the same context on every step, so most of the token bill is waste if you let it be. We cache the parts that do not change and only pay full price for what is actually new. On our own usage that runs up to 62% of tokens coming from cache, which is why the bill stays in tens of dollars instead of thousands. It is not magic, it is just discipline built into the framework so nobody has to remember to do it by hand."

Gurkan Ugurlu, Lead AI Engineer, Keepnet

Keeping cost under control is part of how we run AI responsibly, not a side issue. Our AI management is certified to ISO/IEC 42001:2023, the international standard for managing AI systems, which covers ethics, risk management, transparency and data integrity. You can read more on our compliance and certifications page.

The one question to ask any AI security vendor

If you are looking at agentic AI for your security programme, the demo will look impressive. They always do. So ask the question the demo will not answer on its own: what is your cache rate, and how do you stop token cost running away?

If the answer is vague, assume the cost is uncontrolled and the bill is your problem. If you want the wider picture on how agents fit into human risk management, our pillar on agentic AI in cybersecurity walks through where this is heading, and how attackers are using the same tools is worth a read alongside it. If your interest is training specifically, see agentic AI security awareness training.

Sources

• Gartner: AI Coding Costs Will Surpass Average Developer Salary by 2028 as Token Consumption Surges, 24 June 2026
• Gartner: By 2030, Inference on a 1T Parameter LLM Will Cost GenAI Providers Over 90% Less Than in 2025, 25 March 2026
• Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025
• Gartner: 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025, 29 July 2024
• SecurityWeek: The AI Token Costs That Can Break Cybersecurity, 30 June 2026
• TechCrunch: The Token Bill Comes Due, Inside the Industry Scramble to Manage AI's Runaway Costs, 5 June 2026
• Keepnet: Compliance and ISO/IEC 42001:2023