#aibudget

Uber’s CTO admitted it publicly. The budget was blown. The results were real. Both things are true simultaneously. Here is what that means if you are building with AI and managing a budget that has to answer to real numbers.

On April 30, 2026, Uber’s CTO Praveen Neppalli Naga spoke at TechCrunch’s StrictlyVC event in San Francisco. He said two things in the same breath that together form the most honest public statement about enterprise AI investment made this year.

First: Uber “blew through” its AI budget after opening agentic tools to its engineering organisation.

Second: 10% of all code at Uber — across 8,000 engineers — is now generated autonomously. A hotel-booking integration that would normally take a year was done in six months using agentic workflows only.

He did not say the budget overrun was a problem. He said it in the same sentence as the results. That framing is the whole lesson.

The verified facts from this week

Uber CTO Praveen Neppalli Naga at TechCrunch StrictlyVC, April 30, 2026: Uber “blew through” its AI budget after opening agentic AI tools.

10%: share of all code at Uber now generated autonomously, across ~8,000 engineers. Naga: “10% at our scale is huge.”

6 months vs 12 months: Uber hotel-booking integration built using agentic workflows only. Normal timeline: one year. Agentic timeline: six months.

$950M raised by Sierra on May 4, 2026: led by Tiger Global and GV at $15.8B post-money valuation. 8 months after a $350M round at $10B.

$150M ARR: Sierra’s run rate by February 2026, up from $100M in November 2025. $50M added in under 3 months.

40%+ of Fortune 50: now use Sierra’s enterprise AI agent platform. Customers include Prudential, Cigna, Blue Cross Blue Shield, Rocket Mortgage, Nordstrom, and Singtel.

$400 billion: Bret Taylor’s estimate of annual global customer service spend, which he says is now addressable by AI agents at a fraction of the headcount cost.

80% reduction in patient authentication time: Cigna, using Sierra’s platform, 8 weeks to production.

70%+ resolution rate: Singtel AI agent, 10 weeks to production.

All figures from TechCrunch (May 4, 2026), CNBC (May 4, 2026), Brainroad, and Trending Topics Europe. All verified.

Why budgets blow through and why that is not the whole story

The pattern is consistent across enterprise AI deployments in 2026. ICONIQ Capital’s January State of AI report found inference costs average 23% of total revenue at scaling AI companies. BetterCloud found scaling AI from pilot to production typically reveals 500 to 1,000% cost underestimation.

The specific mechanism at Uber: when you open agentic tools to 8,000 engineers simultaneously, usage scales faster than any pilot-based projection. Engineers who discover a tool that accelerates their workflow use it more than the beta data suggests. Inference costs multiply with every code generation cycle, every automated test, every agentic task run across a large organisation.

This is not carelessness. It is what happens when a genuinely useful tool meets a large user base without proper inference governance in place.

The urgency enterprises feel about deploying agentic AI comes with significant upfront costs, even as companies anticipate lower costs and higher revenue in the long term. — TechCrunch, May 4, 2026 (on Sierra’s raise)

What Sierra’s rise tells us about where the return actually comes from

Bret Taylor’s framing at the HumanX conference in April 2026: most enterprise software is not used. Employees interact with systems like Workday only at mandatory moments. The agent-first model replaces navigation with intent. No employee manually navigates a system. Agents do it for them.

The deployments producing documented results in 2026 all share one pattern: narrow scope, high-volume workflow, measurable baseline, clear improvement target. Not broad agent access to everyone. One specific workflow. One specific outcome. Measured.

Deployments that produced results

Cigna: 8 weeks to production, 80% reduction in patient authentication time

Singtel: 10 weeks, 70%+ resolution rate on customer queries

Nordstrom: 5 weeks to deploy voice agent “Nora”

Uber: hotel-booking integration in 6 months vs 12 months standard

What they had in common

Narrow, specific workflow with documented baseline

Measurable outcome defined before deployment

Human escalation path built in from day one

Short time to production — 5 to 10 weeks

The constellation of models — what Sierra does that most don’t

Sierra’s platform runs on 15+ AI models simultaneously, not a single provider. Bret Taylor calls it a “constellation of models” architecture. Each task is routed to the model best suited for it. No single provider’s pricing change or deprecation can break the platform. This is the same multi-model routing architecture that reduces inference costs by 60 to 85% compared to routing everything to a single frontier model. GMTA’s AI development services are built around this architecture for the same reason Sierra uses it: at scale, single-provider dependency is a cost and reliability risk that compounds.

The four things to do differently than Uber

Budget for 95th percentile usage, not average. Uber’s overrun was partly seasonal. Any product in a business with volume spikes needs an inference model that accounts for peak, not average monthly figures.

Build governance before the agent layer. Priority queues, inference caps, cost dashboards, escalation thresholds — these are pre-deployment decisions, not post-launch fixes. They are the difference between a controllable AI cost and an uncontrolled one.

Start with one workflow, measure it, then expand. Every successful deployment in Sierra’s customer base started narrow. Broad agent access to an entire organisation is not a starting point. It is what you earn after proving the model on one workflow.

Set a 30-day cost review. A formal review of inference spend 30 days after go-live catches governance gaps before they compound. Most teams do not schedule this. Most cost overruns run for 60 to 90 days before anyone looks seriously at the bill.

The return is real when the scope is narrow and the governance is designed in. Uber’s story proves both sides of that: the budget ran over, and the results were meaningful. The founders who replicate the results without the overrun are the ones who design the governance layer before they open the door. GMTA’s 

Planning an AI agent deployment and want to get the governance right from day one?

GMTA builds agentic AI systems with cost governance built in — inference caps, priority queues, escalation paths, and cost dashboards. Because the budget overrun is avoidable if you design for it.