The headline examples are easy to dismiss as big-company excess. Microsoft reportedly pulled back most internal Claude Code licences after costs climbed. A Priceline employee told TechCrunch that a routine Cursor renewal came back four to five times more expensive. GitHub said the pricing behind Copilot had become unsustainable as short chats gave way to long, autonomous coding sessions.
These are not isolated procurement mistakes. They expose a structural problem that is even more dangerous for SMBs and scale-ups: AI converts software from a mostly fixed per-seat cost into a variable production input, but most companies still budget, price and govern it like ordinary SaaS.
That mismatch creates a leaky P&L. Some leaks show up in cost of goods sold. Others hide in R&D, support, cloud infrastructure or duplicated software. A few appear only when gross margin misses plan. By then, the business has already scaled the wrong economics.
The leak starts when a fixed seat meets variable work
Traditional SaaS trained buyers to think in seats: headcount multiplied by a monthly price. AI agents break that logic because two users with the same licence can create radically different costs. One asks for a code explanation. The other launches an agent that reads an entire repository, calls tools, retries failed steps and works for hours.
GitHub acknowledged this directly. It said a quick chat and a multi-hour autonomous coding session could cost the user the same amount under its old model, while GitHub absorbed the escalating inference expense. The company moved Copilot toward token-based AI credits, describing the previous premium-request model as no longer sustainable.
The lesson runs both ways. Vendors that sell variable-cost work at a flat seat price attract the heaviest users and inherit their risk. Buyers who treat an AI seat as a fixed licence miss the variable spend that appears in credits, overages, API calls, actions and human review.
Sharp insight: The seat is no longer the economic unit. The unit is a completed, accepted piece of work. If finance cannot see the cost of that unit, the budget is only pretending to be fixed.
Leak 1: Premium intelligence becomes the default
Teams routinely send every task to the strongest available model. It feels safer: nobody gets blamed for buying too much intelligence, while the person who selects a cheaper model owns every edge-case failure.
Salesforce found the cost of that habit in production. Its Agentforce stack originally relied on one rented frontier model for everything, including intent detection, safety checks, retrieval ranking and evaluation. Salesforce says its token bill grew linearly with traffic. Instead of passing the bill to customers, it split routine tasks into specialised models and kept a frontier model for the reasoning that actually required one.
Its intent classifier became roughly 55 times faster than the general model it replaced while improving reported safety-topic accuracy. That is not merely an engineering optimisation. It is a margin redesign.
Sharp insight: Model routing belongs in the operating model, not in a later cost-cutting sprint. The financially mature default is the cheapest model that clears a measured quality threshold, with escalation for the minority of work that needs more capability.
Leak 2: Agent loops multiply invisible work
A chatbot usually produces one visible answer. An agent may plan, search, retrieve, call tools, critique its own output, retry and ask another model to evaluate the result. The user sees one task. The invoice sees a chain of consumption.
The economics deteriorate quickly at the margin. Research cited by TechCrunch found that the heaviest AI users among developers were about twice as productive as lighter users but consumed roughly ten times as many tokens. The same reporting said per-developer consumption had risen about 18.6 times in nine months.
More usage can still be rational, particularly when it ships valuable work faster. But usage is not value, and productivity does not rise in a straight line with token consumption.
Sharp insight: Every autonomous loop needs a financial stopping condition as well as a technical one. Cap retries, tool calls, context size and wall-clock time. Escalate expensive runs to a human before the agent is allowed to keep spending.
Leak 3: Adoption is measured before economics
AI programmes often celebrate seats activated, prompts sent and weekly active users. Those are adoption metrics. None proves that the income statement improved.
McKinsey's latest global survey found that AI use was widespread, yet only 39 percent of respondents reported any enterprise-level EBIT impact. Most of that group attributed less than 5 percent of EBIT to AI. Smaller companies also lagged in moving beyond pilots.
This gap matters because experiments are cheap and forgiving. Production adds monitoring, security, integrations, fallbacks, support and human review. A pilot can appear efficient precisely because it excludes the costs required to make it dependable.
Sharp insight: Do not ask whether employees are using AI. Ask which P&L line moved because they used it. Adoption without an economic counterfactual is activity, not return.
Leak 4: The human cost never leaves
Many AI business cases price the model and assume labour disappears. In reality, labour often moves. People review outputs, correct failures, handle escalations, maintain knowledge bases and redesign broken workflows.
Bessemer Venture Partners' AI pricing playbook warns founders that AI economics do not automatically resemble SaaS economics. It points to gross margins around 50 to 60 percent for some AI businesses, compared with 80 to 90 percent for traditional SaaS, because every query can carry compute, inference and human-in-the-loop cost.
Those figures are not universal benchmarks, but the mechanism is universal: AI can create a second labour line before it removes the first. If the business case counts the theoretical hours saved but not the review time retained, the P&L is flattering the technology.
Sharp insight: Measure verified labour released, not minutes theoretically saved. Time has financial value only when it increases throughput, avoids hiring, reduces contractor spend or creates additional revenue.
Leak 5: AI cost is booked in the wrong place
The same invoice can mean different things. AI used by an internal finance team is operating expense. AI embedded in a customer-facing product is closer to cost of revenue. AI used to build software may sit in R&D. Bundling all three into a single innovation budget obscures the economics that management actually needs.
For a scale-up, the most dangerous version is customer-facing inference hidden outside gross-margin reporting. Revenue rises, usage rises faster, and the product looks healthy until cost per customer expands with every renewal.
Sharp insight: Classify AI spend by economic purpose before classifying it by vendor. The CFO needs to know whether a token supports product delivery, internal productivity, experimentation or infrastructure. A vendor total cannot answer that.
Leak 6: Pricing ignores the cost curve
Founders face a difficult choice. Charge by consumption and customers struggle to predict the bill. Charge per seat and the vendor absorbs heavy-user risk. Charge per outcome and the economics align with value, but the vendor owns delivery risk.
Intercom prices its Fin AI agent by outcome, charging for a resolved conversation or other completed result rather than for every model call inside the workflow. GitHub moved in the opposite direction, exposing more of the underlying consumption through AI credits. Both choices reveal the same truth: the charging unit determines who carries the variability.
For most scale-ups, the answer is not a pure model. A base platform fee can cover fixed delivery costs, while usage or outcome tiers capture variable work. The important step is to test pricing against the customer at the 90th and 99th percentile of usage, not only the average account.
Sharp insight: If your highest-usage customers are your least profitable customers, you do not have a growth problem. You have a pricing architecture problem.
Leak 7: Finance responds with a blanket cap
When visibility fails, the first shocking invoice triggers an indiscriminate response: freeze new tools, revoke access or impose the same cap on every team. That contains cash but destroys information. High-return workflows are cut alongside low-value experimentation.
The industry is now building the missing management layer. The Linux Foundation announced the Tokenomics Foundation to develop common standards and benchmarks for AI infrastructure economics, extending FinOps thinking into token-based spend. Its existence is a signal: AI cost control is becoming an operating discipline, not a procurement footnote.
Sharp insight: The alternative to unlimited spend is not universal restriction. It is differentiated capital allocation. Fund workflows with strong cost-per-outcome economics, constrain unproven experiments and shut down systems that have lost their owner or purpose.
Build an AI P&L before AI builds one for you
A founder and CFO do not need a perfect token-accounting system to regain control. They need a shared economic model that connects technical consumption to business value.
- Separate the spend. Tag customer-facing inference, internal productivity, experimentation and shared infrastructure instead of hiding everything inside one AI budget.
- Name the unit of value. Use a resolved ticket, accepted code change, completed review, qualified lead or verified hour of capacity released - not tokens or prompts.
- Measure cost per successful outcome. Include model calls, retries, tools, vector storage, observability and human review. Track the median and the expensive tail.
- Route deliberately. Use rules, caching and smaller models for repeatable work. Reserve frontier reasoning for tasks where evaluation shows it changes the result.
- Put boundaries inside the product. Set alerts, soft caps, retry limits and human approvals before broad rollout. Do not wait for finance to police the invoice after the fact.
- Review and expire. Every agent needs an owner, an outcome target and a review date. If nobody renews the business case, the workload should stop.
- Price for the tail. Model economics using heavy accounts and peak workflows. Average usage hides the customers most capable of destroying margin.
The number that matters
The fashionable metric is cost per token. It is useful for engineering, but it is insufficient for management. Cheap tokens can still produce expensive failure; expensive tokens can be rational when they produce a high-value result.
The number that belongs in the management meeting is cost per accepted business outcome, compared with the value of that outcome. It gives the CFO a financial denominator and the technical team a design constraint. It also makes the decision obvious: scale, redesign, reprice or stop.
AI spend is not inherently unpredictable. It is granular, variable and easy to detach from ownership. What looks like a runaway model bill is usually the accumulated result of decisions the company never forced itself to make explicitly.
The winners will not be the companies that spend the least on AI. They will be the ones that can see exactly where AI earns its place on the P&L - and where it leaks.
Sources and further reading
- GitHub: Copilot moves to usage-based billing
- Salesforce: Cutting inference spend by right-sizing models
- TechCrunch: The industry scramble to manage runaway token costs
- Axios: AI sticker shock hits corporate America
- McKinsey: The State of AI
- Bessemer Venture Partners: The AI pricing playbook for founders
- Intercom: Fin AI Agent outcome pricing
- Linux Foundation: Tokenomics Foundation announcement
