What enterprises do when the AI bill runs past the plan
Over the past week, the heads of the two most prominent AI labs each struck a cautious public note, one on the pace of model progress, the other on the timing of a public listing. We will leave the interpretation to others. What we can add is what enterprises themselves are doing and planning, drawn from three ETR surveys: the September AI Tools Pulse, and the October Macro Views Survey and Technology Spending Intentions Survey (TSIS), both still fielding with preliminary results that update daily.
The short version: enterprises are over their AI budgets, and they are choosing to fund the overrun rather than stop it.
For the first time in three months, the share of companies enforcing hard limits on AI consumption moved up, to roughly one in five. More telling is what companies do once a limit exists. In August, the most common responses were cutting tokens per task and restricting agentic workloads. In September, the top response became shifting routine work to cheaper models or lower tiers, followed by cutting tokens and, newly among the leading responses, raising the budget rather than restricting usage. Restricting agentic workloads fell sharply.
The sample of companies managing against limits is small, so we treat the levels as directional. The direction is not ambiguous. When enterprises hit the ceiling, they route routine work down the price ladder and ask for more room. They do not turn the agents off. One Global 1000 healthcare respondent described the mechanism plainly: select the model behind the scenes so that cheaper models handle the routine work.
The Macro Views Survey, with more than 1,300 technology leaders responding so far, confirms the pattern at scale. More than half of organizations report that AI spending exceeded budget over the past six months, up from the July survey. The most common response was to seek supplemental budget approval, followed by reallocating from elsewhere in the IT budget. Shifting workloads to open-weight models appeared as an option for the first time this wave and immediately outranked pausing. Pausing or reducing the scope of AI initiatives remains the least common response, chosen by fewer than one in five. In July, the ranking looked like below; October's preliminary data reorders the middle of the field.
The funding must come from somewhere. The donor list has not changed much since July: external contractors and consultants, legacy infrastructure and hardware refresh, and non-AI software licenses. What changed is the weight on hardware refresh, which grew as a source of funds even as respondents report paying more for the hardware they do buy. Fewer units, at higher prices, to pay for tokens.
The September Pulse shows where the pressure is concentrating. The AI platform, with the highest user satisfaction in our survey for the fourth month running, now also carries the lowest cost-effectiveness score of the four we track. At the other end of the field, respondents rate one platform identically on satisfaction and cost, and its satisfaction has not moved all summer. Cheap and unloved is a position. So is beloved and expensive.
Cheaper is winning share at the margin. The tool with the largest satisfaction gains this month earned it with respondents citing lower-priced models. Inside the leading bundled productivity assistant, buyers are standardizing on one default model provider over the other by a widening margin. And open-weight usage jumped in a single month to a clear majority of surveyed companies, without a corresponding drop in any single proprietary tool. Enterprises are adding cheaper options alongside the premium ones, not swapping them out.
Preliminary October TSIS data adds one more layer. Spending intentions for the leading model provider remain the highest in our universe and are still climbing among the largest enterprises. Where they have softened, it is among midsize and small accounts. The deepest budgets are deepening. The thinnest are the ones tiering down.
For technology buyers, the playbook emerging in the data is to route, then raise: send routine work to cheaper models and tiers, and go back for more budget for the work that justifies premium pricing. Turning agents off is the last resort, and almost nobody is taking it. For technology vendors, the message is that quality alone does not settle the bill. Buyers are resolving the tension between the tool they like and the price they pay by routing around it, and a credible cheaper tier is now part of the competitive set. And the first read on 2027 IT budgets suggests enterprises plan to grow the budget to fit AI, not absorb AI into the budget they already have.
ETR clients can access the complete September AI Tools Pulse, including tool-level satisfaction, cost-effectiveness, and consumption drivers, along with preliminary October Macro Views Survey and TSIS results as they update, on the ETR Platform. To learn more, contact service@etr.ai.