AI-Assisted Software Development: 5 Lessons we Learned at Tech Week

At Tech Week in San Francisco, we met engineering leaders from different industries for a closed-door roundtable and workshop on engineering drift.

In the room were VPs of Engineering, heads of platform, enterprise architects, and CTOs, all of whom used AI coding tools on large codebases. We asked them one simple question: Now that AI is writing much of your code, what has changed?

The answer was clear. AI has sped up code production far more than the review process, team knowledge, priorities, or metrics around it. In other words, AI-assisted software development and AI-assisted software delivery are moving faster than the systems around them can keep up with, which brings up challenges around control, context, and cost.

Key Takeaways

• AI-assisted software development has made code production 3x faster or more.
• Review has not kept up. Large AI-generated pull requests are hard to check, and few teams do it well.
• “Mostly right” code is not safe for core systems, especially when it touches customer data.
• The rules that matter most are often unwritten, and AI can’t follow rules that only live in people's heads.
• Building fast is not the same as building the right thing. Off-track features slow other teams down.
• AI can’t fill gaps in data it never had, so teams keep going back to customers.

Speed Is Real, But It’s Not the Whole Story

Nobody in the room doubted the gains they get from AI. Attendees said code production was 3x faster or more on their teams, a number that points in the same direction as industry data.

The 2025 DORA report found that AI use now links to higher delivery speed, a change from the year before. However, it also found that delivery is less stable and that trust in AI tool accuracy among developers reaches only 33%.

That second finding came up in almost every talk we had in San Francisco. Leaders would mention AI speed in one sentence, then spend ten minutes on what it cost elsewhere.

Here are the five themes that came up the most.

1. Review Can't Keep Up With Code Generation

AI writes code faster than people can review it. Pull requests are bigger and arrive more often, so reviewers can’t check each one as carefully as they used to. Several leaders admitted their teams now rarely review AI-generated code line by line.

Faros AI calls this the AI Productivity Paradox. Their report, using data across more than 10,000 developers, found that teams with heavy AI use merged 98% more pull requests, while review time went up by 91%.

2. “Mostly Right” Isn't Safe Enough for Core Systems

Code that works nine times out of ten is fine for a dashboard. It’s not fine for payments, customer data, or other safety-critical operations.

Leaders described AI-generated code that passed tests and checks, then failed on the one case that mattered. They’re not alone: in Stack Overflow's 2025 survey, 66% of developers said AI output that was almost right, but not quite, was their top problem.

3. The Rules That Matter Most Are the Ones Nobody Wrote

Every large codebase runs on rules that were never written: which library to avoid; which service you must never call directly; why that table has a strange column that somehow makes sense.

This type of knowledge typically lives in people's heads and leaves when they do. AI only works from the context it gets, which means unwritten rules are invisible to it. And for AI to work well, it needs the right AI organizational context.

4. Building Fast Isn't the Same as Building the Right Thing

When building gets cheap, teams build more. Leaders described features being vibe-coded fast without a clear fit with the organization’s plan. That then created work for other teams who had to integrate, support, or undo them.

Speed at the team level turned into engineering drift at the company level, showing how faster delivery of the wrong thing is still the wrong approach.

5. AI Can't Fill Gaps in Data It Never Had

Several engineering leaders described the same loop: AI makes building a new feature faster, then the team finds a requirement nobody wrote down and has to reach back out to the customer to fill in missing information.

The problem with unwritten rules or knowledge is not only that it doesn’t give AI enough context to work properly, but also that information gaps show up later as rework.

The Pattern Underneath

None of these five takeaways are tool problems. Better models won’t fix a review process built for human output or recover knowledge that was never written down. These are operational problems that AI has made more visible and harder to ignore.

Governance is still catching up. In Stack Overflow's 2026 survey, only 24% of developers said their employer had published approved AI tools or formal guidance on how to use them.

That leaves us with six open questions:

• What work should AI do, and what should humans own?
• How do we verify outputs before they reach customers?
• What organizational knowledge must AI be able to access?
• Same as above, applied to legacy systems.
• Who decides which initiatives get resources?
• How do we measure outcomes rather than activity?

What's Next

Over the next few weeks, we will publish a series of six articles, each of which will dive deeper into what came out of the discussions we had in San Francisco, and answer these questions.

If your team is working through the same challenges, we would like to hear how it looks in your environment. Start by calculating how much engineering drift is costing you and then request a demo to get the conversation started.

FAQs

Does AI-assisted software development actually make teams faster?

At the coding stage, yes. AI makes code production 3x faster or more. But review, checks, and system health often do not keep pace, so end-to-end delivery can stay flat or become less stable.

Why is code review becoming a bottleneck?

AI makes more pull requests, and larger ones, than reviewers can check at the same depth. Review capacity has not grown with generation, so either review time rises or review quality drops.

What is the “almost right” problem with AI-generated code?

The code looks correct and passes most reviews but fails on edge cases, often where it touches business rules or customer data. It’s harder to catch than broken code because nothing looks wrong.

Why does institutional knowledge matter for AI coding tools?

AI generates code from the context it gets. Rules, limits, and past choices that were never written down are invisible to it, so its output can drift from how the company works.

What is enterprise AI governance for engineering teams?

It’s the set of ownership, checks, and metrics that keep AI-generated software aligned with a company's standards and goals.