Companies Are Spending Millions on AI But Few Can Prove It's Working

Companies Are Spending Big on AI - But Who's Checking If It Works?

In December 2025, Uber gave its engineers access to the Claude Code tool and set up internal leaderboards ranking teams by how many tokens (AI usage units) they consumed. By April 2026, the entire 2026 AI coding budget was gone. The real surprise came when Uber's President and COO, Andrew Macdonald, admitted there was no clear link yet between all that increased usage and actually better products for riders and drivers.

The episode gave a name to a broader problem: "tokenmaxxing" - a surge in token consumption without the ROI to justify it. According to Gartner, enterprise AI agent software spending is expected to reach roughly $207 billion this year - up more than 139% from $86.4 billion in 2025. The catch: token pricing doesn't behave like traditional software costs. The same employee, using the same tool, can generate wildly different bills depending on how they used it that day.

Uber now caps spending at $1,500 per employee per month per agentic coding tool. Microsoft questioned the cost of Claude Code licenses and canceled them across one of its divisions, and Duolingo reversed a plan tying AI usage to performance reviews after employee pushback.

Industry experts disagree on the fix. Some argue it's an infrastructure problem - many tools default to routing every task, even trivial ones, to the most expensive frontier model when a cheaper one would do. One emerging solution is "smart routing" - a software layer that automatically routes each task to the most cost-effective suitable model, already offered by vendors like Databricks, AWS Bedrock, and Azure AI Foundry.

In practice, almost no company could put a hard number on the return. The one exception was Everlaw, a legal-AI company: one project spent $3,500 in tokens and cut a task from an estimated 9.5 engineer-months down to just 2.5 - a clear, easily justified ROI.

The takeaway for small and mid-sized businesses: before chasing the widest possible AI adoption, it's worth pausing to ask how you'll actually measure whether the investment is paying off. Tracking usage is only the first step; measuring real business value is the harder challenge.

Read the full story on VentureBeat

Google Launches Gemini 3.8 Flash - A Model That 'Works Harder' But Might Cost More

Google has launched Gemini 3.8 Flash, arriving just weeks after its predecessor. The company says the new model "works harder" by performing more reasoning steps on complex tasks and calling tools iteratively. Introductory pricing matches the prior model ($0.75 per million input tokens, $3.75 per million output tokens), but Google warns the model may use more tokens to maximize performance - meaning it could end up costing more in practice, echoing the exact concern described above.

The model shows improvements on software engineering, finance, and legal benchmarks, and ships with dedicated safeguards against misuse in cyber and chemical/biological domains. Alongside the launch, Google also unveiled Gemini 3.8 Flash Cyber - a dedicated version limited to governments and trusted partners (around 650 organizations, including CrowdStrike), as part of its new "Fairwind" proactive cyber-defense program.

A Plain-English AI Glossary for This Week's New Terms

Following a week full of new AI vocabulary like "opaque recurrence" - the controversial reasoning technique in OpenAI's Astra model we've covered this week - TechCrunch published a plain-English AI glossary explaining terms like AI agents, chain of thought, and MCP. A handy resource for anyone trying to stay current without getting lost in the jargon.

---

Sources for this article: VentureBeat, The Verge, TechCrunch