If AI is writing more of your code, what are you measuring?
If you only read one line
More AI-generated code can speed delivery and even cut costs, but it only helps if reliability, review, and engineering judgment keep up.
Monday move: This week, pick one AI-related metric your team already tracks. Ask whether it measures completed work and reliability, or just more pull requests.

The Spark
I have been watching teams ship more code than they did a year ago, often with AI writing a large share of it. The dashboards look great until you ask a quieter question: is the software any more reliable, and do the people reviewing it still understand what just shipped? Today is about those trade-offs: faster development, lower costs, and the engineering questions that get bigger as AI scales.
Your $60,000 Hire Actually Costs Up to $93,000

When the inbox is overflowing and the books are behind, the obvious fix looks like a full-time hire.
But a $60,000 salary doesn't cost $60,000. Add payroll taxes, benefits, PTO, and recruiting, and it's closer to $81,000 to $93,000.
Fractional support changes the math.
An Executive or Marketing Assistant takes over your calendar, inbox, and campaigns. A Bookkeeper or Fractional CFO delivers clean books and tax-ready financials before year-end. You pay for outcomes, not overhead.
BELAY's free guide, The $0 Hire: How Fractional Support Pays for Itself, breaks down the true cost of full-time hires role by role and shows which fractional hire will pay you back fastest heading into Q1.
BELAY has matched 25,000+ leaders with vetted, U.S.-based professionals who drive growth instead of adding cost.
Download The $0 Hire
AI Is Making Engineers Faster, But at What Cost?
AI is changing how software gets built, from the number of pull requests developers complete to how quickly code reaches production. But higher output does not automatically mean better engineering. The real challenge is understanding whether AI improves productivity, controls costs, and protects software reliability at the same time.
One engineering team's experience offers a useful look at what happens when AI becomes part of everyday engineering work. The results show significant gains in development activity, lower AI costs after a model change, and new questions about code quality, onboarding, and production incidents.
More Code, More Pull Requests, Faster Releases
AI agents have moved beyond suggesting code and are now producing substantial portions of production software. On one team, nearly 80% of pull requests reaching production contain some AI-generated code.
The shift accelerated around June 2025, when Claude Agents and improved worktree tools made it easier for engineers to coordinate multiple AI sessions and work on different tasks in parallel.
The operational changes have been substantial:
Product engineering headcount increased from 18 engineers in 2024 to 50 in 2026.
Pull requests shipped to production increased 3.4 times compared with the previous year.
Pull requests without human reviewers increased from 12% to 43%.
Average pull request size grew from 22 to 111 lines.
Engineering output, measured through estimated complexity points, increased sixfold.
Releases per engineer per day rose 47%, from 1.9 to 2.8.
The team also reached 140 production releases per day in October 2026, compared with 35 releases per day in October 2024.
These numbers show how AI can expand engineering capacity. Developers can delegate implementation work, run multiple agents, and move changes through development more quickly. However, the increase in activity also creates more work for testing, monitoring, and maintaining the software.
Tip: Measure engineering success through completed work, reliability, and customer impact rather than pull request counts alone. More code is valuable only when it solves the right problems.
AI Spending Is Falling, but Usage Remains High
AI adoption can become expensive when developers rely on advanced models for complex tasks throughout the day. On that team, AI spending peaked in July, when monthly costs exceeded $300,000. At that rate, annual spending would have reached approximately $3.6 million.
The team's AI expenses later declined significantly following the adoption of Claude Opus 5.5. Before the change, spending was approximately $4,500 per engineer per month, largely associated with Opus 5. Afterward, the figure fell to around $1,600 per engineer per month.
Current spending is approximately 80% allocated to Claude and 20% to Cursor.
This matters because reducing AI costs does not necessarily require reducing AI usage. Better model efficiency can allow developers to complete similar workloads with fewer resources.
The team also developed an internal measurement called an "effort point." One effort point represents a medium-sized pull request, whether written by a human or generated by an AI agent. The team uses this measure to estimate the cost of producing work of comparable complexity instead of simply counting lines of code.
However, AI costs vary considerably across teams. Infrastructure engineers can spend significantly more tokens than other developers because they use AI for production investigations, benchmarking, analysis, and other tasks that may not produce a pull request.
This highlights an important distinction: not every valuable AI interaction results in code. A tool can save hours investigating a production issue without generating a single line of software.
Tip: Track AI spending against the value of the work completed. A higher cost per pull request does not automatically indicate waste when the work involves complex investigations or infrastructure improvements.
Productivity Gains Are Not Equal for Every Engineer
One of the most interesting findings is that AI does not increase everyone's output by the same amount.
Between October 2024 and October 2026, the median engineer's weekly pull requests increased by approximately 10%, from 10.1 to 10.7. The average increased by around 30%, from 11.1 to 14.8. More experienced engineers saw a much larger improvement, with weekly output rising from 11.1 to 19 pull requests, an increase of approximately 70%.
The difference suggests that experienced developers may be better positioned to use AI effectively. They understand the codebase, recognize architectural problems, and can judge whether generated code is appropriate.
New engineers face a different challenge. The team asks new hires to avoid heavy AI usage during their first month because early dependence on AI can interfere with learning the codebase.
That approach prioritizes long-term understanding over immediate output. New developers need to learn how systems work, where important components live, and why certain technical decisions were made. Without that foundation, AI-generated suggestions can become difficult to evaluate.
AI can accelerate someone who already understands the work, but it cannot automatically replace the knowledge gained through experience.
Tip: Give new engineers time to understand the codebase before encouraging heavy AI use. Strong fundamentals make AI assistance more useful and reduce dependence on generated answers.
Faster Releases Require Stronger Quality Controls
Higher development speed introduces a difficult trade-off: more changes reach production, which creates more opportunities for defects.
On that team, human comments per effort point declined by 80%, while incidents per 1,000 pull requests fell by 60%. This suggests that the team improved its ability to review and ship code with fewer defects relative to development volume.
However, the total number of incidents still increased because the team was shipping substantially more changes.
The causes of production problems also changed. Earlier incidents were frequently linked to incorrect code that could have been caught during review. The team responded by improving its codebase for AI use, strengthening review rules, and relying more on Cursor's Bugbot to identify potential defects.
Newer AI models also helped produce better code.
But another problem became more prominent: missing information about the production environment. Even when generated code is technically correct, it can still cause problems if the AI does not understand how the software behaves in real-world conditions.
This is where the next stage of AI-assisted engineering becomes important. Agents need better feedback from production systems so they can understand the consequences of their changes, identify failures, and improve future work.
Tip: Automate code checks, but do not treat automated reviews as a complete replacement for engineering judgment. Production monitoring and feedback are essential when development speed increases.
The Real Measure of AI Adoption Is Sustainable Output
The team's experience shows that AI can increase engineering capacity while reducing the cost of certain development tasks. Yet the results also reveal why productivity cannot be judged by the volume of generated code alone.
Experienced developers appear to gain more from AI, while new hires still need time to build technical understanding. Model improvements can reduce expenses, but different teams require different levels of AI resources. Faster releases can improve delivery, but they also demand stronger quality controls and better production feedback.
The most effective engineering teams will not simply be the ones using the most AI tools. They will be the ones that know where AI saves time, where human expertise remains necessary, and how to maintain reliability as output grows.
The goal is not to produce more code at any cost. It is to deliver better software, make engineering work more efficient, and keep systems dependable as development accelerates.
Your 5-minute action
☐ Open a note called AI output check v0. Pick one team using AI to write code this month. Write down the metric you trust most today, then add one quality signal you are not tracking yet, such as incidents, review coverage, or how long new engineers take to get productive.
A rough first version is enough. That is your first spark.
Hit reply and tell me: as AI writes more of your team's code, what question worries you most: speed, cost, or reliability? I read every answer.
Keep going, one small spark at a time.
Erwin
Compare Annuities With A Licensed Advisor In Three Steps

Leverage Planning simplifies annuity shopping with advisor led recommendations and quotes from major carriers like Athene, Allianz, Nationwide, And Pacific Life. Get options aligned to your timeline, risk comfort, and income goals.
• Request A Custom Quote
• Review Curated Options Virtually
• Get Help With The Application
Minimum investment: $50,000
Get Your Custom Quote
PROMO CONTENT
Can email newsletters make money?
As the world becomes increasingly digital, this question will be on the minds of millions of people seeking new income streams in 2026.
The answer is—Absolutely!
That’s it for this episode!
Thank you for taking the time to read today’s email! Your support allows me to send out this newsletter for free every day.
What do you think for today’s episode?
How would you rate today's newsletter?
Share the newsletter with your friends and colleagues if you find it valuable.
Disclaimer: The "Tiny Big Spark" newsletter is for informational and educational purposes only, not a substitute for professional advice, including financial, legal, medical, or technical. We strive for accuracy but make no guarantees about the completeness or reliability of the information provided. Any reliance on this information is at your own risk. The views expressed are those of the authors and do not reflect any organization's official position. This newsletter may link to external sites we don't control; we do not endorse their content. We are not liable for any losses or damages from using this information.
