Technology Report

Managing Token Spending without Choking Off Opportunity

Most of the cost risk comes from a handful of workflows; govern those tightly, and get out of the way almost everywhere else.

By Jue Wang, Anne Hoecker, Alec Koh, Chris McLaughlin, and Vito Trinchera

Published on September 29, 2026
min read


At a Glance As AI adoption grows, companies often respond to rising costs with user-level limits, but these controls can curb valuable experimentation and adoption. Companies should manage AI spending by workflow, moderating low-value use cases, unleashing high-value ones, and ring-fencing risky cases. Measuring cost and value by workflow helps companies manage spending more effectively, channeling investment to where it creates the most value. This article is part of Bain’s Technology Report 2026 Explore the report The first time a company receives a startling AI bill, the instinct is almost always the same: Put limits on usage. Many companies set up per-person budgets and implement approval processes to clamp down on costs before they spiral out of control. The reaction seems to make sense, but it solves the wrong problem. Companies can become so concerned about controlling costs that they restrict the very behaviors they’re trying to encourage. Developers ration AI use, sales teams stop experimenting, and leaders worry more about the bill than the value. Our view is that the bigger problem right now in most organizations isn’t runaway token spending, but too much spending in the wrong places. Most companies should be spending tokens across more of their organization to radically change their business. Across organizations, underuse is the bigger problem. One global technology company set a soft cap of $500 per month for each developer to help encourage AI use; the average cost landed around $200. At another tech company, engineering leaders were metered so cautiously that they left their allocated $40-per-person monthly budget unspent, never touching the buffer pool sitting behind it. Leadership eventually told them to stop metering and go spend. But this restraint is often overshadowed by a few big spenders within a small number of AI workflows that are not spread evenly across thousands of users. A lack of transparency limits the ability to measure returns and channel spending to the places where it will generate the most value. For one widely deployed knowledge-assistant tool, nearly all tokens are cached input tokens (embedded and reused in API requests) that a company can’t easily see or forecast. Managing token spending effectively requires precision, not blanket controls. Manage workflows, not users The right question isn’t "how much should each employee spend?" It’s "which workflows deserve more AI, and which deserve less?" Framing the question in such a way leads to three distinct approaches, each with its proper place in the company (see Figure 1). Moderate: Some activities don't require frontier AI models or even an agent. Routine email drafting, document summaries, or straightforward information retrieval can often be handled by cheaper models or without an LLM at all. Open-weight models can reduce costs for such routine tasks since they can be self-hosted or run at significantly lower per-token rates than frontier APIs. The goal is to reduce unnecessary spending without sacrificing quality. But it’s difficult to enforce routing to cheaper options when the frontier model is one slash command away. Unleash: Most day-to-day work falls into this category. Software development, sales preparation, customer service, and many structured business processes generate significant value from AI while costs remain predictable. Here, organizations should encourage adoption rather than constrain it. The goal shouldn’t be minimizing token spending but maximizing productivity. Ring-fence: A small number of workflows require much tighter oversight. They share traits: Open-ended optimization, autonomous research loops, massive security scans, and persistent AI agents can consume large numbers of tokens without clear stopping points. These should be tested in controlled environments, instrumented carefully, and equipped with safeguards before being deployed at scale. Figure 1 Three distinct approaches to token spending play out differently across functions Source: Bain & Company The principle is simple: Govern the risky tail tightly, and get out of the way almost everywhere else. Measure cost by workflow This approach only works if organizations understand what each workflow actually costs and what it’s worth. Most companies track AI spending at the user or business unit level, but that doesn’t show where value is being created. The more useful question is “what does it cost to complete a proposal, resolve a customer issue, review a contract, or generate production-ready code?” First versions of agentic workflows often cost far more than expected because prompts, orchestration, and model choices haven’t been optimized yet. In our experience, a new workflow might cost 10 times more than it ultimately should. Once teams understand the economics, they can often remove most of that excess through better architecture, smarter model selection, and more deterministic workflow design. As token prices declined, companies expected their costs to follow. But most companies typically upgrade to the latest frontier models, agents consume more tokens as they tackle increasingly sophisticated tasks, and adoption expands into more workflows. As a result, lower per-token costs are offset by higher usage, leaving overall cost per task stubbornly high (see Figure 2). The expectation is that these economics will eventually improve, but the anticipated cost inflection point has yet to materialize (for more, read the Bain Brief “How Token Economics Will Change Opex”). Figure 2 Average cost per token is falling, but token use is rising faster Sources: Wells Fargo, October 2025; Bain analysis Even so, the right strategy isn’t just to use cheaper models; it’s to reserve expensive reasoning for the parts of a workflow that genuinely require it while making the rest more structured and predictable. Govern selectively This also changes how organizations should govern AI spending. Most workflows belong with the business. Engineering leaders should manage engineering workflows, and sales leaders should manage sales workflows. They’re best positioned to understand where AI creates value, and they can encourage adoption without unnecessary bureaucracy (for more, read the Bain Brief “Mobilizing the Organization in the Agent Economy”). Only the highest-risk workflows require central oversight. A small governance team can establish standards for testing, monitoring, and cost controls while ensuring that open-ended agentic systems include appropriate safeguards before they are deployed broadly. Where possible, those controls should be built into the core AI platforms themselves, making safe, cost-effective behavior the default rather than relying on users to follow policy. The same principle applies to spending limits. Low caps often create exactly the wrong behavior, discouraging experimentation and slowing adoption. High soft limits, combined with clear visibility into unusual consumption, encourage teams to use AI where it creates value while still protecting against genuinely excessive costs. Prepare for the next wave Most organizations are still in the early stages of AI adoption. As autonomous agents become more capable, AI consumption is likely to increase dramatically. Persistent agents, long-running optimization tasks, and increasingly sophisticated workflows will all drive significantly higher token usage than today's interactive assistants. That makes it even more important to establish the right management approach now. Organizations that rely on blunt cost controls will struggle to capture AI's full value. Those that understand the economics of their workflows will be able to expand adoption confidently while managing the relatively small number of use cases in which costs can escalate quickly. Read the Next Section The Missing Architecture for Agentic Software Development More from the report New Innovation Is Required to Fund AI’s $6 Trillion Buildout Software Investing in the Age of AI and Slower Growth AI Data Center Boom: Can We Build It If They Come? The $100-Billion SaaS Opportunity Hiding in Cross-System Labor Hardware Strikes Back in the AI Era Tech Services: Cracking the Code on AI-led Growth Cybersecurity’s New AI Imperative: Attacking the Backlog of Vulnerability Alerts The AI-Native Enterprise: Absorption Is the New Advantage The Half-Finished Redesign: How AI Reshapes Software Organizations Managing Token Spending without Choking Off Opportunity The Missing Architecture for Agentic Software Development How AI Is Changing Data Monetization Read our Technology Report 2026 EXPLORE THE FULL REPORT DOWNLOAD THE PDF

At a Glance

As AI adoption grows, companies often respond to rising costs with user-level limits, but these controls can curb valuable experimentation and adoption.
Companies should manage AI spending by workflow, moderating low-value use cases, unleashing high-value ones, and ring-fencing risky cases.
Measuring cost and value by workflow helps companies manage spending more effectively, channeling investment to where it creates the most value.

The first time a company receives a startling AI bill, the instinct is almost always the same: Put limits on usage. Many companies set up per-person budgets and implement approval processes to clamp down on costs before they spiral out of control.

The reaction seems to make sense, but it solves the wrong problem. Companies can become so concerned about controlling costs that they restrict the very behaviors they’re trying to encourage. Developers ration AI use, sales teams stop experimenting, and leaders worry more about the bill than the value.

Our view is that the bigger problem right now in most organizations isn’t runaway token spending, but too much spending in the wrong places. Most companies should be spending tokens across more of their organization to radically change their business. Across organizations, underuse is the bigger problem. One global technology company set a soft cap of $500 per month for each developer to help encourage AI use; the average cost landed around $200. At another tech company, engineering leaders were metered so cautiously that they left their allocated $40-per-person monthly budget unspent, never touching the buffer pool sitting behind it. Leadership eventually told them to stop metering and go spend.

But this restraint is often overshadowed by a few big spenders within a small number of AI workflows that are not spread evenly across thousands of users. A lack of transparency limits the ability to measure returns and channel spending to the places where it will generate the most value. For one widely deployed knowledge-assistant tool, nearly all tokens are cached input tokens (embedded and reused in API requests) that a company can’t easily see or forecast.

Managing token spending effectively requires precision, not blanket controls.

Manage workflows, not users

The right question isn’t "how much should each employee spend?" It’s "which workflows deserve more AI, and which deserve less?"

Framing the question in such a way leads to three distinct approaches, each with its proper place in the company (see Figure 1).

Moderate: Some activities don't require frontier AI models or even an agent. Routine email drafting, document summaries, or straightforward information retrieval can often be handled by cheaper models or without an LLM at all. Open-weight models can reduce costs for such routine tasks since they can be self-hosted or run at significantly lower per-token rates than frontier APIs. The goal is to reduce unnecessary spending without sacrificing quality. But it’s difficult to enforce routing to cheaper options when the frontier model is one slash command away.
Unleash: Most day-to-day work falls into this category. Software development, sales preparation, customer service, and many structured business processes generate significant value from AI while costs remain predictable. Here, organizations should encourage adoption rather than constrain it. The goal shouldn’t be minimizing token spending but maximizing productivity.
Ring-fence: A small number of workflows require much tighter oversight. They share traits: Open-ended optimization, autonomous research loops, massive security scans, and persistent AI agents can consume large numbers of tokens without clear stopping points. These should be tested in controlled environments, instrumented carefully, and equipped with safeguards before being deployed at scale.

Three distinct approaches to token spending play out differently across functions

Source: Bain & Company

The principle is simple: Govern the risky tail tightly, and get out of the way almost everywhere else.

Measure cost by workflow

This approach only works if organizations understand what each workflow actually costs and what it’s worth. Most companies track AI spending at the user or business unit level, but that doesn’t show where value is being created. The more useful question is “what does it cost to complete a proposal, resolve a customer issue, review a contract, or generate production-ready code?”

First versions of agentic workflows often cost far more than expected because prompts, orchestration, and model choices haven’t been optimized yet. In our experience, a new workflow might cost 10 times more than it ultimately should. Once teams understand the economics, they can often remove most of that excess through better architecture, smarter model selection, and more deterministic workflow design.

As token prices declined, companies expected their costs to follow. But most companies typically upgrade to the latest frontier models, agents consume more tokens as they tackle increasingly sophisticated tasks, and adoption expands into more workflows. As a result, lower per-token costs are offset by higher usage, leaving overall cost per task stubbornly high (see Figure 2). The expectation is that these economics will eventually improve, but the anticipated cost inflection point has yet to materialize (for more, read the Bain Brief “How Token Economics Will Change Opex”).

Line chart indexed to 100 at December 2024, tracking Azure AI token consumption against per-token cost through a March 2026 estimate. Tokens reported climb steadily to and index of 541 by the final point, while cost per million tokens falls to an index of 42; an embedded annotation on the chart notes that despite the price drop, higher usage has more than doubled the overall bill. — Average cost per token is falling, but token use is rising faster

Even so, the right strategy isn’t just to use cheaper models; it’s to reserve expensive reasoning for the parts of a workflow that genuinely require it while making the rest more structured and predictable.

Govern selectively

This also changes how organizations should govern AI spending. Most workflows belong with the business. Engineering leaders should manage engineering workflows, and sales leaders should manage sales workflows. They’re best positioned to understand where AI creates value, and they can encourage adoption without unnecessary bureaucracy (for more, read the Bain Brief “Mobilizing the Organization in the Agent Economy”).

Only the highest-risk workflows require central oversight. A small governance team can establish standards for testing, monitoring, and cost controls while ensuring that open-ended agentic systems include appropriate safeguards before they are deployed broadly. Where possible, those controls should be built into the core AI platforms themselves, making safe, cost-effective behavior the default rather than relying on users to follow policy.

The same principle applies to spending limits. Low caps often create exactly the wrong behavior, discouraging experimentation and slowing adoption. High soft limits, combined with clear visibility into unusual consumption, encourage teams to use AI where it creates value while still protecting against genuinely excessive costs.

Prepare for the next wave

Most organizations are still in the early stages of AI adoption. As autonomous agents become more capable, AI consumption is likely to increase dramatically. Persistent agents, long-running optimization tasks, and increasingly sophisticated workflows will all drive significantly higher token usage than today's interactive assistants.

That makes it even more important to establish the right management approach now. Organizations that rely on blunt cost controls will struggle to capture AI's full value. Those that understand the economics of their workflows will be able to expand adoption confidently while managing the relatively small number of use cases in which costs can escalate quickly.

Read the Next Section

The Missing Architecture for Agentic Software Development