The AI budget meets the budget process

Samsara's CIO, Stephen Franchetti, has capped AI usage for some non-technical employees while leaving research and development more room to experiment, and built an internal system that tracks all AI expenses daily ^[1][2]. Set that against global AI spending projected at $2.5 trillion this year, a 44% rise on the prior year ^[3], and you have the defining tension of the moment: the tools got approved before anyone built the meter.

The costs are not mysterious once you look at where tokens go. Docusign's chief technology officer, Sagnik Nandy, told Fortune that every engineer at the company uses AI tools and that 75% of the code they produce is initiated by AI ^[4], but that the coding agents were designed to read the entire code base for context before doing anything ^[5]. Changing the default so agents pull only context relevant to the narrow task cut token usage by almost half ^[6]. That is not a model problem or a licensing problem. It is a configuration default that nobody costed, multiplied by every engineer, every day.

The second lever is model tiering. Jim Dausch, chief digital and technology officer at Yum Brands, says token usage is not yet a material number but that the trajectory is worth watching, having noticed usage and expense both rising earlier this year ^[7]. His contention is that perhaps 95% of the tasks AI is asked to perform can be handled by cheaper, more basic models ^[8], which is why Yum is pushing model-selection training and telling business leaders to manage digital spend the way they budget headcount, so the cost does not sit as a single IT line item ^[9]. Cigna has authorised more than 70 models internally; its chief data, digital and AI officer, Katya Andresen, says compute and token usage have risen while total spending has not ^[10], and that the way to run up a bill is to use the most expensive models with no guardrails ^[11].

Gartner's Will Sommer frames 2026 as the year everyone finds out AI is hard, and warns that firms can spend thousands of dollars per head on output that is junk ^[12][13]. In June, Gartner projected that AI coding costs would overtake the average developer's salary by 2028, driven by rising token consumption and the shift to consumption-based pricing ^[14]. Some companies have already blown past their 2026 AI budgets with no corresponding business value, and hyperscalers have responded with cheaper models and price cuts ^[15][16].

Caps only work if you know what you are buying. American Banker's framing of the measurement problem is the most useful correction available: institutions either treat activity such as access, prompts and pilot counts as value, or they demand a fully attributable financial return from a pilot too immature to produce one, killing it early ^[17][18]. The proposed fix is stage-appropriate evidence across experimentation, workflow improvement, business outcomes and financial outcomes ^[19].

There is a cost that no telemetry dashboard captures. Mark Ma of the University of Pittsburgh, writing in The Conversation, cites an Atlanta Fed study in which about 90% of executives said AI has not yet boosted productivity at their companies ^[20]. His own research found AI investment announcements track AI-attributed layoff announcements, that market reaction to those layoffs averaged close to zero and was negative or near zero in more than half of cases ^[21][22], and that AI-related comments in Glassdoor reviews skew more negative than reviews overall ^[23]. Employee sentiment toward AI, he argues, is among the strongest predictors of productivity when AI is used ^[24].

Watch whether caps become permanent chargeback, and whether anyone publishes a return figure rather than a token-reduction figure.