Agentic AI Case Study - Uber

EXECUTIVE BRIEF: In April 2026 Uber deployed AI experienced unanticipated token usage costs with AI generated code.

When Uber rolled out an autonomous AI coding assistant to boost developer productivity, engineers embraced it heavily, using it to write up to 70% of new code. However, because the agent ran persistent, multi-turn reasoning loops, expensive usage fees accumulated rapidly. By April 2026, Uber had exhausted its entire annual AI budget. The hidden realization was that heavy usage cost up to $2,000 per engineer, per month—at scale, making the compute layer as expensive as a secondary engineering salary

Comprehensive market studies show highly polarized returns depending on the complexity of the agent's task.

A 2025 global study conducted by Omdia found that organizations investing strategically in generative and agentic AI saw an average ROI of 49%** ($1.49 returned for every dollar spent)

MIT Sloan research warns executives against confusing time-savings with direct budget reductions [1.2.4]. For instance, if an agent reclaims 20% of an embedded engineer's time by auto-generating unit tests, that rarely translates to a 20% labor cost savings. Instead, the value is unlocked by reallocating that engineer to tasks that were previously perpetually deprioritized (like deep security edge-case audits or manual refactoring)

So, how do we apply this use case in practice? Look for opportunities where a demonstrable ROI can be realized:

High ROI: Straightforward, repeatable transactional tasks (e.g., document processing, automated customer service routing, localized code lints). Teneo and Google Cloud data show interaction costs plunging from $4.00–$6.00 per human interaction to $0.25–$0.50 per agent interaction, a clear 85%+ cost reduction

Low ROI: Deep, multi-agent reasoning loops (e.g., an autonomous software agent autonomously diagnosing, refactoring, and deploying a complex embedded system fix) are the least cost of effective of these especially for those systems that require deep domain expertise such as functional safety and security. If the generated product is subject to regulatory compliance and certification then these would likely fall into the Low ROI bucket.