The software development landscape has shifted dramatically with the use of AI coding tools. Coding assistants and agentic workflows promise unprecedented velocity, yet engineering leaders are finding that this revolution comes with a massive hidden catch: it is incredibly expensive, and engineering output isn’t necessarily getting better. Instead, software organizations are finding themselves caught in a cycle of skyrocketing AI token bills and messy code.

{{cta}}

AI in software engineering: Use more, spend less, and prove it’s working

Engineering teams globally are experiencing this disconnect. In Faros's 2026 AI engineering research, we found that while rapid AI adoption increases the volume of code shipped, it is also leading to quality challenges, such as higher incident rates, bugs, and code churn—a phenomenon known as the AI acceleration whiplash.

Yet, the top-down pressure to adopt AI hasn’t slowed. Leadership wants speed, but they are now flashing three conflicting signals at once: use AI more, don’t blow the budget, and show us what it actually produced.

Without proper guardrails, satisfying these demands is nearly impossible. In our latest webinar, Naomi Lurie, Head of Product Marketing at Faros, and Chase Norton, Head of AI, dive deep into the financial and operational realities of AI-assisted engineering. Naomi notes, “One CTO told us that their best engineer spent $47,000 on tokens in a single month.” Chase adds that when teams build without visibility, “It is very easy to hit footguns that explode the cost and then hear about it the next day.”

Not all AI token spend is created equal

To combat this, engineering teams must move away from brute-force tokenmaxxing toward strategic outcome-maxxing. In the webinar, Chase introduces his comprehensive 6-point framework for productive AI work. Key highlights include:

• Intelligence allocation: Not every task requires the most advanced, expensive model. (“The easiest footgun is a beginner picking the most expensive, highest-reasoning model right away to solve a bug.”)
• Implementation planning: Establishing an evolving conversation with AI before letting it write a single line of code.
• Definition of done: Setting strict guardrails to prevent agentic loops from burning through capital.

Efficient vs. inefficient AI token usage: A side-by-side breakdown

To prove the framework’s value, Chase walked through a live example tackling the same bug twice: once using his efficient framework, and once using the typical “beginner” approach of pasting a bug directly into a top-tier model. The results are staggering:

• The inefficient approach spawned a team of agents that blindly looped, costing $66.37 for a single pull request.
• The efficient framework was nearly 5x cheaper, with a resulting PR quality score 35 points higher—and the developer actually understood every change instead of relying on a black box.

Watch the full webinar now

Are your developers self-aware of their burn rates? Are they using the right models for the right tasks, or are they accidentally spinning up $3,000 bills in ten minutes?

Don’t let your AI coding tools become an unmanaged capital expense. Watch the full on-demand webinar, What is Productive AI?, to get Chase’s complete 6-point framework and learn how to build sustainable, cost-effective AI habits across your engineering teams.

{{cta}}