Uber spent four months proving a point nobody at the company meant to make.
By April, the ride-hailing giant had exhausted its entire 2026 budget for AI coding tools, after engineers adopted Claude Code and Cursor far faster than finance teams had modeled. Uber’s chief technology officer confirmed the overrun to The Information, with monthly costs per engineer running as high as $2,000 against an original per-seat estimate of a few hundred dollars. Uber responded with a $1,500-per-tool monthly cap on employee AI usage, and Walmart, Amazon, and Cisco followed with similar controls of their own. Then its president and chief operating officer went on a podcast and asked, out loud, whether any of it was working.
Uber COO Andrew Macdonald said the link between rising AI spend and better products for riders isn’t there yet, and that it’s genuinely hard to draw a straight line from AI usage stats to shipped features customers notice. Roughly 70% of Uber’s committed code now originates from AI tools. The company still can’t draw a clean line from that figure to a faster, better product.
Uber is an early data point in a pattern about to become an industry-wide reckoning over what AI spend is actually buying.
Token Spend Is Rising Faster Than the Question Gets Answered
Gartner now projects that AI coding costs will overtake the average developer’s salary by 2028, driven by the shift from seat-based software licensing to consumption-based token pricing. Gartner’s Nitish Tyagi put the mechanism plainly: developers optimize for speed and convenience, not cost, and without a governed operating model, spend outpaces the productivity it’s meant to deliver.
Microsoft has begun canceling direct Claude Code licenses in favor of cheaper internal tooling. Walmart capped usage of its internal coding assistant after employees started sending near-identical requests. The pattern repeating across enterprise engineering organizations is the same one Uber hit first: adoption outran the budget model, and nobody had built the instrumentation to tell whether the spend was earning its keep.
Dovetail Software, the customer intelligence platform used by teams across the Fortune 500, sees a structural reason this keeps happening. “Token consumption used to be treated as a proxy for how good you were at using AI,” the company says. “Now it’s the opposite. It’s about how efficiently you use it.”
That reversal is the real story sitting underneath Uber’s budget line. For years, heavy AI usage read as a signal of sophistication, a team moving fast, embracing the tools, shipping more. Gartner’s forecast and Uber’s public accounting suggest the opposite is becoming true. The metric that mattered has flipped, and most engineering organizations haven’t noticed the pole has turned.
Every Prompt Starts From Zero
The mechanism behind runaway token spend is not exotic. A model produces output only as good as the context it’s given, and most teams are feeding it the same context, from scratch, every single time.
Dovetail Software describes two common workarounds, and both are expensive in a way that doesn’t show up until the invoice arrives. One is dumping a spreadsheet of customer feedback directly into a prompt, an approach constrained by the model’s context window and by the sheer clumsiness of asking an LLM to understand a customer relationship from a cold start. The other is stitching together a dozen MCP connectors, so the model burns tokens simply working out which data source it needs before it can even begin reasoning about the actual question.
Neither approach is cheap, and neither compounds. A team using either method pays the same reconstruction cost on the hundredth prompt as it paid on the first, because nothing about the previous 99 prompts made the hundredth one easier or cheaper.
“What those teams needed up front was a durable, structured understanding of the customer, not a fresh guess every time they prompt,” Dovetail Software says.
That distinction, structured understanding versus repeated reconstruction, is where the Uber story stops being a headline about one company’s budget and starts looking like a preview of what’s coming for every organization treating AI spend as a volume problem rather than a context problem.
Building Fast Is No Longer the Hard Part
Uber’s COO framed the dilemma as a missing link between AI spend and product outcomes. That’s a fair description of the symptom. The cause sits one layer beneath it: building has gotten radically cheap and fast, while knowing what to build, and for whom, hasn’t gotten any easier.
An engineering team with unlimited token budget and no structured customer intelligence can ship features quickly. It has no reliable way of knowing whether those features solve a real problem, because the model generating the code has no memory of what customers actually said last quarter, no record of which complaints are trending, and no way to distinguish a genuine pattern from a single loud account.
This is the gap Dovetail’s Sun’s Out Summer Launch ’26 is built to close, giving product and engineering teams a persistent layer of customer signal that AI agents can reason over without rebuilding context on every call. Instead of a raw CSV or a chain of connectors, an agent works from data that has already been classified, themed, and updated continuously as new feedback arrives.
The difference compounds. A support ticket processed today doesn’t just answer today’s question. It sharpens tomorrow’s answer too, because the underlying structure of what customers are saying keeps accumulating rather than resetting.
Efficiency Is the New Scoreboard
Enterprises spent the better part of two years treating AI adoption rates as the scoreboard. Percentage of code written by agents. Number of engineers using Copilot or Claude Code. Volume of tokens burned per sprint. Those numbers were easy to report and looked good in a board deck.
Gartner’s 2028 forecast, paired with Uber’s public reckoning, suggests that scoreboard is about to get replaced. The organizations that come out ahead won’t be the ones spending the most on AI. They’ll be the ones spending the least to get the same or better output, because they’ve solved the context problem instead of paying to re-solve it every prompt.
Token spend for individual developers has climbed sharply over the past nine months as agentic coding tools spread across engineering organizations, and the gap between spend and measurable output is exactly what Uber’s leadership is now questioning in public. Every enterprise running the same experiment, just without a COO willing to say so on a podcast, is asking a version of the same question.
The teams that can answer it will be the ones that stopped treating customer understanding as a document to consult and started treating it as infrastructure that compounds. Dovetail Software’s bet is that this shift, from reconstructing context to reasoning over a structured, always-on layer of it, is what separates efficient AI spend from Uber’s four-month burn rate.
Uber capped its budget because the bill got too large to ignore. The organizations paying closer attention now are trying to make sure they never get a bill that size in the first place.
