How AI will influence Agile Development Teams, why it won't replace them and why this is a good thing.

This point of view was co-created together with Lucas Heckmann. Thank you very much for your thoughts, discussions and perspectives!

Disclaimer: This article refers to mission-critical, enterprise-grade software developed by professional teams within large scale organizations. Not the 47th "Workflow designer" or the next social media startup. Or online marketing agencies.

Generative and agentic AI is being marketed as a cure for basically everything in knowledge work: faster code, fewer meetings, instant documentation, instant insights, unbelievable growth with less people. And in narrow slices of work, those promises are real. Controlled studies on AI coding assistants (e.g., GitHub Copilot) found developers completed a defined coding task substantially faster with AI support.

But enterprise software delivery is rarely constrained by “typing code.” It’s constrained by the delivery system around the code: governance, risk ownership, architecture, security/compliance, and operability.

1) The prototype trap: “It runs” is not the same as “It’s shippable”

AI makes it easier than ever to produce something that looks like a product: a UI, endpoints, a flow that works on a happy path. What is often missing is the invisible enterprise foundation: identity and access, observability, data governance, resilience patterns, dependency management, secure-by-default configuration, and an architecture that can evolve without becoming brittle.

AI can generate code quickly. It cannot automatically guarantee that the system is secure, operable, compliant, maintainable, and economically sustainable.

2) What AI will change inside agile teams

AI is most valuable when it reduces friction and waste around the core work:

• Less formatting, more thinking: A Product Owner shouldn’t spend time perfecting the syntax of a story. AI can standardize structure and acceptance-criteria templates, freeing time for problem framing and validation.

• Faster access to specialist knowledge: AI can explain unfamiliar technical concepts on demand and reduce the cost of asking “basic” questions.

• Developer flow improvements: AI can draft tests, summarize unfamiliar code, propose refactorings, and help with documentation—especially strong for boilerplate and routine transformations.

• Lifecycle automation: Summarizing bug reports, clustering incidents, drafting release notes, synthesizing customer feedback—these are realistic, high-leverage enterprise uses.

A broader “augmentation” pattern shows up outside software, too: a large-scale study of generative AI in customer support found productivity gains on average, with the biggest gains for less experienced workers.

3) What AI will not replace (and why agile teamwork still matters)

Enterprise delivery is a socio-technical problem: unclear goals, competing constraints, hidden assumptions, and changing environments. That requires more than generating text or code.

• Delivery speed is mostly a governance and risk problem, not a typing speed problem. “Lead time for changes” in DORA is measured from commit to production—where approvals, release mechanics, environment readiness, and risk decisions dominate. Governance often exists for good reasons (incidents, regulatory findings), but can drift into process theatre that adds burden without reducing risk. AI can help reduce bureaucracy, but it cannot fix unclear risk ownership or fragmented decision rights on its own.

• Developers remain accountable - and must be able to validate AI outputs. AI does not transfer responsibility. Engineering teams still own correctness, maintainability, security posture, and operability. In enterprise settings, that also means being able to review and challenge AI-generated changes like any other contribution, with the same quality bars.

• AI is best at legwork and pattern work - not accountable invention. It can produce outputs that look creative, but research suggests high-novelty creativity is harder and often amounts to recombination of prior patterns. That’s valuable - just not the same thing as making a defensible, accountable decision under real constraints.

• AI is often optimized to be helpful and agreeable - exactly what teams sometimes should not be. Many assistants are tuned to satisfy the user, which can lead to “sycophancy”: agreeing with flawed framing instead of challenging it. That’s a real risk in discovery and architecture discussions where constructive conflict is essential.

4) Local speedups don’t guarantee enterprise outcomes

Even if AI accelerates parts of development, it can also increase instability if it pushes more change through a system that isn’t ready.

DORA’s 2024 research reported that higher AI adoption correlated with reduced delivery stability, and the effect on throughput was small and could even be negative - despite some improvements in code review speed and code quality measures.

That’s not an argument against AI; it’s an argument for integrating AI into a healthy engineering system (platform, testing strategy, review standards, observability, and clear quality bars).

5) Enterprise constraints you can’t „prompt away”

• Security: LLM-based use cases introduce specific risks (prompt injection, sensitive data disclosure, insecure output handling). OWASP maintains a dedicated “Top 10” list for LLM application risks.

• Regulation and governance: In the EU, the AI Act is phasing in with enforceable requirements over time. Enterprise rollouts need governance because “internal productivity tool” can quickly become “customer-impacting system.”

6) The parallel to agile transformations

A decade ago, many organizations tried to “install agile” by renaming roles and meetings. It didn’t work because constraints were structural: funding models, governance, architecture, incentives, and leadership behaviors.

AI is at a similar moment. It won’t be enough to “add an AI tool” and expect step-change outcomes. Organizations will get results when they:

• pick the right use cases (high-volume friction points),

• invest in quality and delivery fundamentals,

• set guardrails (security, data, compliance),

• and upskill teams to use AI deliberately (especially for validation and decision quality).

7) Conclusion – and the good thing

AI will absolutely change Agile development teams - but mostly by making the easy parts easier and the hard parts more visible.

• It will accelerate drafts, automate some bureaucracy, and reduce the cost of accessing knowledge.

• It will also raise the stakes on engineering fundamentals: small batches, test automation, security, and disciplined review.

• Teams won’t disappear. But the center of gravity will move: away from producing text and boilerplate, toward making decisions, managing risk, and integrating change safely across a messy application landscape.

And for enterprise delivery, that’s a good thing.