By Daniel Fallmann, Founder and CEO of Mindbreeze
For the last few years, enterprise AI has run on a comfortable assumption: more context produces better intelligence. If a model can accept a larger context window, feed it more. More documents, longer prompts, deeper histories, bigger retrieval payloads. A better-informed model makes better decisions, right? Not always.
In production, that logic has started sending enterprises a bill they did not budget for.
Once an organization moves from a handful of pilots to thousands or millions of live interactions, every piece of information sent to a model carries an economic cost. Prompts consume tokens. Retrieved documents consume tokens. System instructions, tool descriptions, agent-to-agent chatter, generated responses, all tokens. What looks trivial in a demo becomes a material line on the P&L once it is multiplied across an enterprise. The industry has taken to calling this token maxxing, and at Mindbreeze we have made reducing it an explicit product priority, for a reason I will come to. But I want to be clear at the outset that the deeper issue is not token consumption, it is what runaway token consumption reveals about how an enterprise has architected its AI.
The question executives should be asking is not how much context their AI can process. It is how little context the AI needs to produce the right answer.
Key Takeaways
- Enterprises often assume that more context ensures better AI performance, but excessive context can lead to inefficient token usage and increased costs.
- Token maxxing occurs when organizations send too much data to AI models, which can significantly impact budgets as interaction scales up.
- AI architecture is usually the source of inefficiency, not the model itself; optimizing the design can dramatically reduce token consumption.
- Effective retrieval strategies improve cost efficiency and enhance AI response quality by providing only relevant information to models.
- Executives need a new metric—intelligence efficiency—to track the value generated from AI token usage, rather than just focusing on consumption.
Table of contents
AI Economics Have Moved from Access to Consumption
Enterprise software used to have legible economics. You bought licenses, provisioned infrastructure or cloud capacity, and could forecast the bill with reasonable confidence.
Generative AI broke that predictability. CIO’s analysis of tokenomics describes it as one of the most practical subjects in enterprise AI, precisely because prompts, retrieved context, tool schemas, system instructions and generated responses all feed a usage-based bill, which means AI economics now depend not just on which model you pick but on how the application is designed and how people and agents actually use it. The piece names the recurring culprits directly: context inflation, sending oversized prompt prefixes and retrieval payloads on every call; poor model matching, routing routine tasks to premium models; and response sprawl, letting the model return far more text than anyone needs.
This intensifies with agentic AI. A person submits one request, but an agent may reason across several steps, retrieve repeatedly, call multiple tools and consult other agents before it finishes. One business action can quietly spawn dozens of model interactions.
The people paying these bills have noticed. Reuters reported that Commonwealth Bank of Australia CEO Matt Comyn, whose bank is an aggressive AI adopter, warned that AI costs will rise in less predictable ways as tasks grow more complex, and put the mechanism plainly: as models gain reasoning, tool access and larger context, token costs do not scale on a linear basis. He expects companies to start scrutinizing that spend hard this year.
For a CEO or CIO, this quietly reclassifies AI from an innovation topic into an operating-model topic.
The Problem Is Usually Not the Model
The reflex, when the bill climbs, is to negotiate a lower per-token price or switch to a cheaper model. Both have merit. Both also distract from the larger source of waste, which sits in the architecture around the model rather than in the model itself.
The software harness coordinating an agent, its system prompts, tool definitions and orchestration logic, can dramatically change token consumption even when the underlying model and task are held constant. In the research cited, redesigning the harness alone cut token consumption by 38 percent, cost per task by 41 percent and execution time by 44 percent, at comparable quality. The model was not the variable, the architecture was.
Cheaper tokens will not rescue an inefficient AI architecture. After an adoption cycle that rewarded maximal deployment and experimentation, the piece argues, most companies are simply not built to use tokens efficiently, describing a pipeline that leaks tokens, and dollars, at every stage, with raw uncurated data consuming five to ten times more tokens than necessary. Pouring cheaper tokens through a leaky foundation is not a strategy.
This is why token maxxing is strategically interesting rather than merely annoying. Tokens are not the disease. They are the unit through which an inefficient information architecture finally becomes visible on an invoice.
More Context Can Produce Less Intelligence
There is a second misconception worth retiring: that more information reliably improves an AI response. It does not.
Ask an experienced executive to review a contract. Hand them the contract, the relevant amendments and the current customer history, and you have helped. Hand them every contract the company has ever signed, thousands of unrelated emails and years of unrelated account history, and you have not made them smarter. You have buried the signal.
Large language models suffer the same fate. We examined this directly in the analysis of tokenmaxxing: excessive context routinely carries outdated, duplicated and irrelevant material that adds little to the answer while inflating what you pay to process it. Beyond a point, more context degrades both cost and quality, which is why retrieval strategy matters more than the raw size of the context window.
The architectural goal is not abundance. It is precision. An enterprise AI system should identify the smallest set of authoritative, relevant, permission-appropriate information needed to answer the question in front of it, which requires actually understanding the enterprise’s knowledge before any of it reaches the model.
Retrieval Has Become an Economic Control
This is where enterprise search and knowledge management take on a role most organizations have not yet appreciated.
Retrieval used to be a user-experience concern. Better search helped employees find things faster. In the generative AI era, retrieval quality also governs AI economics, because what you retrieve is what you pay the model to read. This is the conviction behind our recent announcement that Mindbreeze InSpire is engineered to reduce token maxxing: by using hybrid search, semantic understanding and deep content processing to identify the genuinely relevant passages before anything reaches the LLM, the system lowers the tokens required per answer while improving accuracy. Rather than pushing an entire corpus or large volumes of undifferentiated context at the model, the aim is a compact, high-signal context scoped to the specific request.
I put it this way when we announced it, and I will stand behind it here: value does not come from how many tokens you burn, it comes from giving the model exactly the right information at the moment it is needed. When retrieval is precise and grounded, you spend less, you trust the answer more, and AI finally scales the way the business expected. That is not a marketing line. It is the whole architectural thesis, compressed.
Every irrelevant paragraph that never reaches the model is information the organization does not pay the model to process, does not risk the model being confused by, and does not have to govern after the fact. At enterprise scale, precise retrieval stops being a convenience and becomes a financial control.
The C-Suite Needs a New Efficiency Metric
Executives have long measured technology efficiency through infrastructure utilization, licensing, cloud spend and workforce productivity. Enterprise AI introduces another dimension, one most dashboards do not yet track: intelligence efficiency. How much useful business value does the organization generate from the context it asks its AI to process?
That question grows more urgent as agents run longer workflows and AI spreads across the enterprise. A system that consumes twice the tokens to produce the same reliable outcome is not twice as intelligent. It is half as efficient, and at scale that inefficiency compounds into a number a CFO will eventually ask about.
So, I do not think the winners of the next phase will be the organizations with the biggest models or the largest context windows. They will be the ones with the discipline to give their AI exactly what it needs and nothing it does not. That discipline is, in the end, a knowledge problem before it is a cost problem, which is precisely why we have spent years building the retrieval and grounding layer that makes it achievable. The next phase of AI economics will not reward maximizing consumption. It will reward maximizing the value of every token.
