There is a particular kind of shock that happens when a finance team asks an AI team to justify its spend and the AI team discovers it cannot. Not because the results are not there, but because no one ever measured them against the cost. That is the story playing out inside enterprises across the United States, Europe, and Asia right now. The trigger: CNBC reported in late June 2026 that Uber had burned through its entire annual AI budget by April. Microsoft had cancelled Claude Code subscriptions across multiple product divisions. Meta had quietly taken down the internal leaderboard it had built to track which team was burning the most tokens — a leaderboard that had, until recently, been treated as a badge of honour.
The practice that drove all of this has a name. Tokenmaxxing is what happens when an organisation treats token consumption as a proxy for AI productivity. The logic that got enterprises here was seductive and not entirely unreasonable: the teams burning the most tokens were shipping the most AI-assisted features. Tokens in, features out. But that equation only works if every token is doing useful work, and in practice most are not. Context windows fill with irrelevant prior conversation. Agents re-derive facts they already established. Models are called for tasks that do not require their capability level. Prompts carry boilerplate that could be cached. The result is an AI spend that scales with usage but not with value — a metric that finance departments are now finding impossible to defend to boards.
The numbers behind the reckoning
Enterprises running unoptimised AI deployments typically pay for 3-5x more tokens than the actual task requires. Prompt caching cuts input costs by up to 90% on repeated context. Model routing — sending simple tasks to lower-cost models — reduces spend by 40-60% without accuracy loss. Semantic caching reduces API calls by 30-40% on query-heavy applications. Combined, these three techniques routinely deliver 60-75% cost reductions in the first 90 days of implementation. The enterprises seeing 171% ROI from AI are the ones applying all three. The ones burning through budgets are the ones applying none.
Why the Tokenmaxxing Era Made Sense at the Time
Understanding why enterprises got here is important because it explains why the fix is not as simple as telling teams to use AI less. The original rationale for tokenmaxxing was a legitimate productivity bet. In 2024 and early 2025, the constraint on AI-assisted output was not cost — it was adoption. Organisations that were not pushing their teams to use AI were losing ground to competitors that were. Spending more on tokens was a rational signal of competitive intent. Procurement and finance teams approved large AI budgets because the alternative — under-investing in AI capability — seemed riskier.
The problem was measurement. Most organisations built their token spend before they built their measurement infrastructure. They knew how many tokens they were consuming. They did not know which tokens were producing value. By the time cost became a board-level concern — which for most large enterprises happened in Q1 2026, when the first full-year AI bills landed — the spend was already embedded in dozens of workflows across dozens of teams, and the people who understood those workflows had been optimised for output, not efficiency.
The Three Levers That Actually Move the Number
When Wizeb works with an organisation on AI cost reduction, the diagnosis almost always reveals the same three failure modes operating simultaneously. The remedies are well-understood. The difficulty is applying them without disrupting the AI-assisted workflows that are generating real business value.
1. Prompt caching: The fastest win
Most enterprise AI applications repeat a significant portion of their context on every call. System prompts, policy documents, product catalogues, knowledge bases — these are loaded fresh with each request even when they have not changed. Anthropic's prompt caching charges 90% less for cached input tokens than fresh ones. For an application with a 4,000-token system prompt running 10,000 daily API calls, the difference is the equivalent of making 9,000 of those calls free. Implementation is a single-day change for most applications. The ROI is immediate.
2. Model routing: Matching task to capability
Not every task requires the same model. Summarisation, classification, extraction, and simple question-answering perform at high quality with smaller, cheaper models. Complex reasoning, multi-step analysis, and creative generation benefit from flagship capability. Yet most organisations send every request to the same model — often the most capable and most expensive — because routing logic was never built. A well-designed routing layer (sometimes called an LLM gateway) analyses each incoming request and routes it to the appropriate model tier. The typical result: 40-60% cost reduction with no measurable quality regression on the tasks that matter. For organisations with mixed workloads — support agents, document processing, creative generation, and analysis all on the same platform — the savings are often higher.
3. Semantic caching: Eliminating redundant API calls
The third lever is the least understood but often the most impactful for query-heavy applications. Semantic caching stores the outputs of previous API calls and returns cached results when an incoming query is semantically similar — not necessarily identical — to a query already answered. For support agents and internal knowledge tools, where users consistently ask variations of the same questions, a well-tuned semantic cache reduces API calls by 30-40%. The cache layer sits in front of the AI model, responds in milliseconds rather than seconds, and costs nothing to serve. At scale, it is frequently the difference between an AI support deployment that is cost-positive and one that is not.
What a Realistic Cost Reduction Looks Like
One mid-market B2B software company came to Wizeb in Q1 2026 running an AI-powered customer support agent. The agent was handling 3,200 tickets per month. Monthly AI cost: $38,400. That is $12 per resolved ticket, against a human support cost of $18 — a positive ROI on paper, but barely. The improvement target was to get below $6 per ticket without degrading resolution quality.
The audit revealed three problems. First, the system prompt was 6,800 tokens and loaded fresh on every call. Second, the agent was using the flagship model for every interaction regardless of complexity — and 60% of interactions were simple FAQs that needed no complex reasoning. Third, the application had no caching layer; users frequently asked variations of the same ten questions, and the model was called fresh each time.
- Prompt caching on the system prompt reduced input token costs by 68% on average across calls. Monthly saving: $8,200.
- Model routing sent 58% of incoming queries to a lower-cost model tier. Average call cost dropped from $0.022 to $0.009. Monthly saving: $14,800.
- Semantic caching eliminated 34% of API calls entirely. Monthly saving: $5,100.
- Combined result: monthly cost fell from $38,400 to $10,300. Cost per resolved ticket: $3.22. Resolution quality scores were unchanged.
That is not an outlier. It is what structured optimisation looks like when applied to a deployment that was built without it. The compound savings were available from the first deployment — they just required someone to go looking.
The Governance Question Enterprises Are Not Asking
Beyond the technical levers, there is a governance question that most organisations are either not asking or asking too late: what are you measuring? Uber's problem was not that it spent too much on AI. It was that it had no mechanism to connect spend to outcome. The teams burning tokens were not malicious or careless. They were optimising for the metrics they had — feature velocity, deployment frequency, ticket throughput. Token spend was not a metric anyone was held accountable for.
The enterprises that are pulling ahead in 2026 have added a third metric to their AI programme alongside capability and adoption: cost per unit of business value. Not tokens consumed. Not API calls. Cost per resolved ticket. Cost per qualified lead. Cost per document processed. Cost per decision supported. When cost is measured this way, the optimisation conversation changes. A team burning 50% more tokens than another team to achieve the same business outcome is not more productive — it is less efficient. That distinction was invisible when token spend was not tracked against output. It becomes visible immediately when it is.
Where Wizeb Fits In
Wizeb's token optimisation service exists because the gap between what organisations are spending on AI and what they are getting for it is real, measurable, and fixable. We audit existing AI deployments, identify the three to five highest-leverage optimisation opportunities, and implement them — typically within four to six weeks. We track cost per business outcome before and after. The target is always 50% or more cost reduction without regression in output quality.
The organisations that engage with this now — before the next budget cycle forces the conversation — are the ones that will be able to demonstrate AI ROI to their boards in 2027. The ones that wait will be the ones fielding uncomfortable questions about why their AI spend is still growing faster than their AI-generated revenue. The reckoning has already started. The question is whether your organisation gets ahead of it or gets caught by it.
Ready to audit your AI spend?
Wizeb's token optimisation audit identifies your three highest-leverage cost reduction opportunities and delivers an implementation plan with ROI projections in two weeks. No commitment required for the audit. Contact us at wizeb.com/services/token-optimization to start the conversation.