Just happened: the AI cost story is starting to change. The old pitch was simple: use AI, automate work, reduce cost. But recent reports around Microsoft and Uber show a more uncomfortable reality - AI can become expensive precisely when everyone starts using it.
This is why token management is becoming a practical skill for anyone building AI agents. If you want to learn how to build agents that are useful without burning through budget, our Applied AI Agents workshop is designed around exactly that kind of hands-on decision-making.
The next AI advantage is not who uses the most tokens. It is who gets the most useful work from the fewest wasted tokens.
What happened at Microsoft and Uber?
Windows Central reported that Microsoft was reducing employee use of Claude Code in parts of its Experiences + Devices division and shifting teams toward GitHub Copilot CLI by the end of June. The report framed the move as partly about workflow control, security and Microsoft-owned tooling - but also noted the financial timing around the end of Microsoft's fiscal year.
Uber's example is sharper. Tom's Hardware reported that Uber's CTO had previously said the company's full-year Claude Code budget was already spent by April. The same reporting highlighted an even bigger question from Uber leadership: where is the clear line between more AI usage and more useful products shipped?
Why token prices falling does not solve the problem
Token prices can fall and total bills can still rise. That is the trap. Agentic systems do not behave like a single chatbot response; they can plan, call tools, read files, retry failed steps, keep long context, and ask a model to critique its own output. Every step can add input tokens, output tokens and hidden overhead.
In a separate analysis, Tom's Hardware cited a Goldman Sachs report suggesting agentic AI could push token demand up more than 24 times. Whether the exact forecast changes or not, the direction is clear: agents make token volume a management problem, not a footnote.
Cheaper tokens do not automatically mean cheaper AI if your workflows use dramatically more of them.
The real lesson: design the agent before you unleash it
The wrong lesson from Microsoft and Uber is "do not use AI." The better lesson is: do not let AI usage scale faster than your understanding of the workflow. A poorly designed agent can spend tokens on unnecessary context, weak prompts, repeated tool calls, and tasks that should have stayed deterministic.
Measure activity. Reward usage. Let every task hit the largest model and hope more tokens produce more value.
Measure outcomes. Route tasks carefully. Use smaller context, cheaper models and human approval where they make sense.
How to manage tokens efficiently
Efficient token management is not just about trimming prompts. It is about designing the whole agent workflow. In our Applied AI Agents workshop, we treat cost control as part of agent design, because an agent that is too expensive to run is not production-ready.
- Keep context small: retrieve only the documents, rows or messages the task actually needs.
- Route by difficulty: do not send simple extraction tasks to the most expensive frontier model.
- Limit loops: cap retries, tool calls and autonomous steps before costs compound silently.
- Cache stable work: summaries, embeddings, classifications and reference lookups should not be regenerated every time.
- Track value per run: measure useful outputs, not just how much AI was used.
Learn AI agent cost control hands-on
If you are building AI agents for work, the question is not "can the agent do it?" The better question is "can the agent do it reliably, safely and affordably?" That is exactly the mindset we teach in the Applied AI Agents workshop.
In one practical session, you will build a working AI agent from scratch, learn where tokens get spent, and practise the design choices that keep an agent useful without creating cost overruns. No prior coding experience required.