October 13, 2026 
5 Min
As companies invest more in AI, they expect faster delivery, better quality, and less repetitive work. But does spending more on AI mean your engineering teams are getting better results? Not necessarily.
How Does Token Usage Work?
Here’s where it gets interesting: AI doesn’t process only the latest message. When you continue a conversation, the model needs the previous messages to understand what you’re asking. That previous information is sent again as input and added to token usage. The longer the session, the more every new message costs.
Token consumption and spending do not move in a fixed ratio. The model, the mix of input and output tokens, caching, and the product’s subscription or usage allowances all affect what a company pays.
Why Is AI Spending Increasing?
Rising AI spending can have several different causes:
- Broader adoption. More developers are using AI, or existing users are applying it to more stages of the SDLC.
- More demanding work. Analyzing a large codebase or investigating a complex bug may require more context and reasoning than generating a small snippet.
- Different models and plans. A shift toward higher-priced models, additional subscriptions, or usage beyond included allowances can increase spending.
- Underused licenses. Per-user pricing means companies may pay for licenses that are rarely used, increasing spending without a corresponding increase in AI adoption.
- More iterations. Repeated corrections, regenerated code, and failed attempts can add consumption without a corresponding improvement in results.
- More automated work. An agent may perform multiple steps, such as inspecting files, generating tests, and revising an implementation, from a single developer request.
These causes can overlap, but they call for different responses. Expanding useful adoption is different from paying for repeated failed attempts. The bill alone cannot tell you which is happening.
More Tokens Don’t Always Mean More Progress
Consider an illustrative example: a team initially uses AI mainly to generate code. Token consumption is modest, but they still spend considerable time adding tests, correcting edge cases, and responding to review feedback.
Then the team extends AI usage to test generation and checks before review. Token consumption increases. If comparable changes now require less manual rework, pass review sooner, and maintain quality in production, that additional usage may be worthwhile.
Now imagine consumption rises because the assistant repeatedly produces incorrect implementations. Developers spend more time correcting those outputs, and review takes longer. The same increase in usage would tell a very different story.
Neither scenario can be judged from tokens alone. Additional AI spending may reduce the total effort needed to deliver a change, or it may add to it. Developer and reviewer time belong in the evaluation too.
The important difference is between how much AI is used and what that usage helps a team achieve. To understand whether AI is improving engineering efficiency, token usage needs to be considered alongside factors such as delivery time, review cycles, code quality, and the amount of rework required.
What Should You Measure Alongside AI Spending?
Start with the outcomes you expected AI to improve, then connect usage and spending to those outcomes across the SDLC.
| Dimension | What to measure | What it helps you understand |
|---|---|---|
| Delivery | Cycle time and time to production | Is work reaching users sooner, or is faster code generation creating a queue elsewhere? |
| Review | Review turnaround time and revision rounds | Are changes easier to review, or is more effort shifting to reviewers? |
| Quality | Production defects, rollbacks, and rework | Are improvements in speed holding up after release? |
| Effort and cost | AI spending alongside development, review, and correction effort | Is the total resource requirement improving? |
These metrics can help reveal whether AI is improving engineering performance, but no single metric tells the whole story. Changes in delivery time, review cycles, or quality don’t automatically prove that AI caused the improvement. The aim is to look at multiple signals over time and build a clearer picture of AI’s impact.
The goal isn’t to minimize token usage at any cost or to encourage developers to use AI as much as possible. It’s to understand whether that usage is making a meaningful difference to engineering work.
Use those findings to guide decisions: expand workflows that improve delivery and quality, investigate repeated failure loops, and choose models based on the requirements of the task. Optimal token usage depends on the outcome you need and the total effort required to achieve it.
What the Myth Gets Wrong
More AI spending doesn’t automatically mean better engineering results. Higher usage can reflect broader adoption or more complex work, but it can also come from repeated attempts, debugging, and rework.
Understanding the difference requires connecting AI usage to what happens next: how changes move through review, testing, and deployment, and how they perform in production.
