How We Roll Out AI Coding Agents in Large Companies: Lessons From the Field

• Adoption is not a technical problem, it is change management. The weakest link is not the technology, it is people's habits. A developer who has written code a certain way for 10 years does not switch overnight.
• Governance has to come before the rollout. Who has access? To which repositories? How do you log what the agent does? At a financial institution, these are not optional questions.
• ROI is not measured only by how fast code gets written. The more interesting metrics are time to first commit for a new developer, the number of round trips in code review, or how long it takes to fix bugs.

• The agent's identity. The agent must have its own account with auditable access, not a developer's shared token.
• Scope limits. The agent works only with repositories it has explicit permission for.
• A review gate. No code from the agent reaches the main branch without a human review.
• Logging and monitoring. Everything the agent does has to be traceable after the fact.

• handles routine refactoring and boilerplate in a fraction of the time,
• writes the first version of tests, which the developer then adjusts,
• helps with onboarding, explaining unfamiliar code better than outdated documentation,
• speeds up code review by catching trivial problems before a human sees them.

• generates code that looks correct but ignores the project's business logic,
• generates API endpoints that do not actually exist,
• cannot judge the architectural impact of its own change,
• needs clear context: the better the prompt or issue, the better the output.

"AI tools dramatically speed up research and understanding a platform. Onboarding onto any new platform is much faster thanks to models and tools like Perplexity, Claude, or Cursor; zero to first draft is technically possible within a single day. It helps most with the monotonous parts that do not need deep thinking, and that speeds up the second part."

"You shift into the role of a reviewer; I have blocks prepared for me instead of writing them myself. But you still have to keep a sense of what is going on, what needs to be achieved, and what the limits are. Paradoxically, AI loads the mind a little, because you no longer have time on the monotonous part of writing boilerplate and you keep your attention the whole time on what changed where."

— David Omrai, developer at DX Heroes

1. Fear of vendor lock-in. "What if GitHub raises prices?" The answer: a strategy that plans for a multi-model approach and does not tie you to a single provider.
2. Compliance concerns. "Where does our code go?" The answer: explain transparently how a specific tool handles data, ideally a self-hosted option.
3. ROI skepticism. "How do I know it pays off?" The answer: a pilot project with measurable metrics on 2–3 teams, not a blanket rollout.
4. Developer resistance. "Our seniors do not want it." The answer: start with early adopters, show concrete results, and let adoption grow organically.

"Clients often ask how the process actually works. We start with a four-hour workshop: an intro, mapping the current state, and above all identifying quick wins. A kind of matrix: what brings a fast result with little effort, and what is a long-term goal. Then we continue through AI ambassadors who reach across teams, so adoption spreads organically."

"What convinces clients the most? That they can start very quickly with professional support. You just need a push: show best practices, set up the environment, and the team can then build on solid foundations. We deliver real value in a short time, and clients value that."

"What matters is that it is not just about taking an AI tool and doing things the same way as before. The work shifts toward more thorough preparation; the analysis and the assignment have to be far better, so the implementation with AI is up to standard. People have to change the way they work."

— Prokop Simek, co-founder of DX Heroes