Articles

When AI meets real engineering teams

This article was written by Yuliya Khadasevich, Liuba Gonta

November 18, 2026
10 minutes
‌
When AI meets real engineering teams

AI as an amplifier of organizational culture

We are living through a historical moment: AI has stepped into the spotlight, and all eyes are on it. A real race has kicked off across the tech industry, and honestly, it makes perfect sense. For the first time in a long while, a new frontier has opened up, one where no playbook exists yet, no clear leaders have emerged and no one has decades of experience to fall back on.

In many ways, organizations are building the plane while flying it. Teams are experimenting with new tools and ways of working, while the industry is still figuring out what best practices look like. With all the excitement around AI, it is easy to believe that it will magically solve productivity challenges, but the experience shows something different. The more we work with teams going through AI transformation, the more clearly we see a pattern: AI tends to work like a magnifying glass. Whatever strengths and weaknesses a team already has, AI makes them more visible.

When a team already has solid engineering practices, clear processes, a healthy review culture and a real sense of ownership, AI gives those a genuine boost. But when an organization is struggling with shaky engineering decisions and cultural gaps, AI scales those problems at roughly the same pace it scales code generation.

So it is worth taking a closer look at the engineering practices that have always mattered, but feel more important now than ever before.

Key engineering practices which still matter with AI

Clear Policies

It is no secret that teams operate more effectively when clear policies are in place. Use of AI also needs clear guardrails, because engineers should know which tools are approved, what data can and cannot be shared and which use cases require a mandatory human review.

DORA [1], a Google-backed research program that studies software delivery performance, calls this a clear and communicated AI stance. Without it, engineers either avoid useful AI use because they are unsure what is allowed or use AI too freely in ways that create risks. Clear standards reduce ambiguity and give teams and organizations the confidence to use AI safely.

A few areas where policy is important to establish:

  • Approved tools, because not every AI tool has the same security posture, data handling agreements or compliance certifications.
  • Data classification, because teams need explicit guidance on what can and cannot be shared with an AI model.
  • Mandatory human review, because some use cases should always have a human in the loop regardless of how confident the AI output looks.
  • Intellectual property and licensing, because AI-generated code can carry licensing implications depending on the tool, and teams need a clear position on how to handle this.
  • Prohibited use cases, because there are things, like generating credentials or secrets, bypassing access controls, etc., that AI should simply not be used for.
  • Psychological Safety
  • Here is a bit of a paradox: as AI gets smarter, the skill that matters most for engineering teams turns out to be not a technical one at all. It is a human one - the ability to openly push back and question things.

As AI continues to attract enormous expectations and attention, a certain collective optimism is taking hold across the industry. That is exactly why psychological safety is becoming one of the most important practices of the AI era.

Teams need to feel safe enough to keep asking questions:

  • Is AI actually helping here or just adding noise?
  • Do we really understand what this change does?
  • Are we confident enough in the quality of what was generated?
  • Does this output even have a reliable source behind it?


Most importantly, engineers should be comfortable enough to simply admit they used AI in their work, which is absolutely a normal thing to do.

There is another dynamic worth naming - as organizations look to justify AI investments, many are trying to measure AI usage across teams and individuals. The intent is usually to understand what is working and where there is room for improvement. But collecting data like this can easily be read the wrong way by the people being measured. When engineers sense that someone is watching their AI usage, it creates the opposite of psychological safety. Leadership needs to communicate a clear and credible message: AI is here to help engineers do better work, not to replace them. Metrics collected around AI usage should be used to improve tooling and processes, and not to evaluate individual performance.

Review Culture

AI integrated into engineering teams generates code with enormous speed, but also introduces a new problem - producing code faster than humans can review it. Research data [2] shows this is already one of the primary limiters on real productivity gains. That raises a discussion around what does good review look like when neither the author nor the reviewer wrote the code from scratch.

It may sound obvious, but one simple and important piece of advice here is to keep changes small. AI can generate hundreds or even thousands of lines of code quickly, but reviewers still need to understand the intent, risks and assumptions behind each change. Large AI-assisted pull requests make that harder. Small, logical PRs keep review focused, make defects easier to isolate and help teams move faster without lowering the quality bar.

Here are a few practices that help validate AI-generated code more smoothly:

Spec-driven development
Write the spec before asking AI to generate code, but keep it scoped to a small, focused piece of work. A well-scoped spec forces the author to think through the problem before generation, gives the model better context to work with and helps reviewers check the output against a clear standard.

Automation first
Linters, formatters and static analysis should catch syntax issues and pattern violations before the review even starts. If those tools are in place, human reviewers can focus on what automation can't assess - domain correctness, business rules and whether the solution actually fits the problem. With AI increasing code volume, the division of labor between automated checks and human judgment becomes more important.

Use AI for pre-review
Running code through a different model as a pre-review opens an additional automated gate - it raises the floor of what reaches human reviewers. That said, different models can share the same blind spots and definitely lack domain context. As a result, AI can complement human review by catching certain issues and suggesting improvements before an engineer even begins reviewing the change.

Use structured review checklists
As AI increases code volume, consistent review becomes harder to maintain through instinct alone. A short team-agreed checklist keeps quality standards stable regardless of who is reviewing or how much is coming in. Without a checklist, one reviewer may focus on tests while another looks only at business logic.

Ownership Culture

With AI-generated systems, it is not enough to simply "generate code", someone must take responsibility for the consequences of a model's decisions in production. This sounds obvious, but AI introduces a new temptation: because the model wrote it, ownership feels diffuse.

Ownership culture in an AI-assisted team means one thing: the human who ships it owns it.

Continuous Learning

Not that long ago engineers could learn a technology and comfortably ride that knowledge for a few years. With AI that is no longer an option - things move too fast, and keeping up means adapting in much shorter cycles than before. What counted as a best practice six months ago might already feel outdated today. That is why a team's ability to keep learning is a real competitive advantage.

Strong teams treat learning as part of the job, not something squeezed in after hours. Organizations should not expect AI learning to happen only "on the side" of delivery work. Teams need protected time to experiment with AI tools, test new workflows, compare approaches and learn from failure in safe conditions, supported by sandboxes, approved tools, hackathons and clear guardrails.

Knowledge Sharing

No single person can keep up with all of the changes on their own. That is why sharing what was tried and learned within a team matters more than ever.

AI also introduces a new knowledge sharing problem: every engineer is quietly building their own ecosystem of prompts, agents and workflows. While that looks like a win for individual productivity, for the organization it can mean fragmented knowledge and everyone reinventing the same wheel in isolation.

There is already research [2] suggesting that AI speeds up individual contributors significantly, but does not always translate into productivity gains at the organizational level. That is why knowing how to use AI well needs to become something the whole team, and better, organization owns together. If the best practices and learnings around AI do not get shared, they cannot be scaled and reused.

Growth Mindset

Applying AI to the way of working requires continuous experimentation. As teams explore new tools, workflows and practices, mistakes and unexpected outcomes are inevitable. This also means it is really important to treat mistakes as learning signals, not personal failures. As AI introduces more probabilistic and less predictable behavior into delivery workflows, teams need space to examine what happened, improve guardrails and share lessons openly.

The strongest teams will be those that learn fastest: experimenting safely, sharing openly and turning proven practices into repeatable ways of working, and they should be supported along the way.

Final thoughts

As you can see, none of these practices are new. AI does not change engineering culture, but makes it much easier to identify whether it needs improvement.

Organizations with strong foundations often see the greatest returns from AI because they are better prepared for the change it brings. AI introduces new tools, new habits, faster feedback loops and more decisions to coordinate. Clear processes, reliable delivery practices and a culture of learning help teams turn that complexity into momentum. In that environment AI becomes a force multiplier.

But the opposite is also applicable. Outdated documentation can mislead both people and AI. Unclear processes and ownership make it hard to know who validates decisions and who is accountable. Poor code quality, large pull requests and weak review practices can turn faster output into faster chaos. Knowledge sharing also becomes critical. If useful prompts, patterns and lessons learned stay with individuals, the team does not improve together. And while AI can help engineers explore unfamiliar technologies or domains, it cannot replace the core knowledge needed to validate the output instead of trusting it blindly.

AI may change the way software is built, but it does not change what good engineering looks like. It will expose existing strengths and weaknesses, scale their impact and make the need for strong engineering culture impossible to ignore.

References
[1] DORA (DevOps Research and Assessment). State of DevOps Research Program. Google. Available at: https://dora.dev
[2] DX Research. AI and Engineering Velocity: A Longitudinal Analysis. Available at: https://getdx.com/report/ai-and-engineering-velocity-a-longitudinal-analysis/

This article first appeared in XPRT. Magazine #22: Commit to the Unknown. Download the full edition.

Written by

Contact

Let’s discuss how we can support your journey.

‌
‌
‌
‌
‌
‌
‌
‌
‌