The backlash against ‘tokenmaxxing’, which equates more AI usage with greater productivity, is growing as token bills escalate. For Claims leaders, a successful AI pilot is only the start. In this blog we explain why costs can spiral in production, where internal builds leave insurers exposed, and how to secure beneficial outcomes with more predictable AI spend.
We’ve all seen the stories of mystery companies blowing $500 million on Claude AI in a single month, of Uber capping employee AI usage after spending their entire AI budget in four months, and of a major healthcare insurer seeing its monthly AI token consumption go from three million to over one hundred fifty million in under a year.
The days of unrestricted AI adoption are passing, and the cost vs. value of deploying AI into your business is creeping up the agenda as exponential token spend hits P&Ls at every enterprise. Finance and Claims leaders are becoming increasingly discerning about where their AI budgets are spent, and experienced CIOs are expecting delivery outcomes supported by predictable AI spend.
Predicting AI costs is an unsolved problem.
Predictable spend is not easy to achieve, and evidence shows that neither humans nor AI are good at forecasting AI spend.
A recent WSJ study found that only 11% of nearly 400 businesses surveyed could accurately forecast their AI spending. More than a third were off by 26–50%. Nearly 1 in 7 were off by more than half.
AI doesn’t perform much better. A 2026 Stanford University paper proved that frontier models fail to accurately predict their own token usage and systematically underestimate real token costs. The same paper showed that token usage is highly variable, almost random, with runs on the same task varying by up to 30x in total tokens with no guarantee of higher accuracy.
The best approach available for now is real-time visibility, with an ecosystem of token-level spend trackers, budget alerts, and cost breakdowns offering Finance teams a live view of what their AI programs are costing at that moment. But none of these tools enable Claims leaders to confidently forecast their next month’s token spend.
Jevons Paradox: as cost-per-token falls, token usage and total spend rise.
To make matters worse, AI systems are consuming more tokens by design, with each new version. And while per-token costs are falling, token usage and total spend is rocketing. For token consumers, this puts control and predictability even further out of reach. Renowned economist William Stanley Jevons’ observation that efficiency gains often lead to greater overall consumption is becoming ubiquitous across the tech world. There is even an AI model named after him. But, at least for now, the truism holds: tokens are getting cheaper, and we’re using more of them.
Today’s frontier models ‘think’ before they answer, and you pay for every word of that thinking, even though you never see it. The messy edge cases and ambiguous documents a model will encounter in insurance claims production are exactly the inputs that will trigger longer thinking, multiple reasoning turns, and (by extension) greater token consumption. In addition, that thinking time isn’t fixed. One model may think for longer than another. One query run within the same model at different times can require very different amounts of thinking. So, the same claim can cost ten times more depending on how and when it is processed..
Agentic systems compound this bloat. Findings in the same Stanford University paper show that agentic tasks consume 1000x more tokens than code reasoning and code chat, with agents calling agents, and multiple reasoning turns quickly bloating token usage.
Tokenizer changes further compound this increase. When a vendor releases a new model with a new tokenizer, your bill can rise even though the published price-per-token stays the same. The price list doesn’t move; the meter does. To illustrate this, one investigation comparing Anthropic’s Opus 4.7 vs Opus 4.6 found 4.7 used 1.46x more tokens than 4.6 on the same system prompt. In real terms, the budget for your pilot can quickly change with each model upgrade, turning the same claim into a third more billable tokens overnight.
What this means for your insurance claims AI pilot
Let’s see this in action. When building an AI pilot, Frontier LLMs will perform well at understanding a single claim, extracting the core components of a document, perhaps even offering a high-level coverage assessment against a single policy document. The inputs are curated, the context window is limited, agentic tasks are short and tidy, and budgets are set based on the model selected at the time of the pilot.
When this pilot is unleashed on real-world production claims, the potential for token usage to climb increases dramatically as AI systems meet more complex documents, have context windows which span years of claims history, and run multiple reasoning turns on those complex claims the AI was deployed to resolve. That doesn’t factor in costs associated with model upgrades, vendor costs, and system maintenance.
What we’re left with is a picture of unpredictable token usage, which directly hits an insurer’s bottom line. A Claims AI pilot might work in a curated trial, and might even appear cheaper at the outset, but exposing it to production volume could lead to unforeseen costs which could grow out of control as tasks become more complex, claims volume increases, and models need re-tuning or upgrading.
Thoughts from leaders in our space
“Insurers routinely face technology buy vs build scenarios to strike the right balance. Although there are several advantages to building AI in-house, insurers must also weigh timelines, expertise, and all related costs. Partnering with purpose-built AI solution providers allows insurers to buy and build, essential for a successful AI strategy.”
Alan Demers, President-Founder, InsurTech Consulting
“Building in-house is rarely a question of whether insurers can, but whether they should. A frontier model can handle a curated pilot. Production claims are a different problem, and that’s where specialist expertise and predictable economics matter. For most insurers, the real question is where to build and where to buy, and how to do it without exposing their P&L to unpredictable token spend.”
Roi Amir, CEO, Sprout.ai
Sprout.ai offers predictability and confidence.
Sprout.ai offers contracted AI outcomes and deep claims expertise on terms you can plan against.
We absorb the risk that unpredictable and invisible token usage poses – LLM usage costs are included in the standard product pricing, so insurers are not faced with the impossible task of managing their token bill.
With that predictable pricing comes delivery guarantees. Models which have been trained on more than 10 million real insurance documents across tens of thousands of different document types, deployable to production in under 3 months, and with AI governance built in at source – confidence scoring, audit trails, and evidence linked to source documents, with performance tracked against agreed accuracy targets.
With a predictable pricing structure based on the volume of claims you process – not your token usage – and contracted outcomes in accuracy and agreed service levels, your AI investment becomes manageable and your bottom line is protected.
Frequently asked questions
AI pilots run on curated inputs, short tasks and limited context windows, so token usage stays low. In production, models face complex documents, years of claims history and messy edge cases. These trigger longer “thinking”, more reasoning turns and agentic workflows, and agentic tasks can use up to 1,000x more tokens than simpler tasks. Model upgrades, vendor costs and maintenance add further spend, so a pilot that looked affordable can become unpredictable at production volume.
Token usage is highly variable. A 2026 Stanford University paper found that the same task can vary by up to 30x in total tokens, and that frontier models underestimate their own token costs. A WSJ survey found that only 11% of businesses could accurately forecast their AI spend. New models can also use different tokenizers, so your bill can rise even when the published price per token stays the same.
The most reliable approach is to move token risk off your P&L. Sprout.ai includes LLM usage costs in its standard product pricing and charges by claims volume, not tokens. It also contracts on outcomes: agreed accuracy targets, service levels and production deployment in under three months. Governance is built in, with confidence scoring, audit trails and evidence linked to source documents. Insurers get AI spend they can plan against instead of a token bill they have to manage.