State of AI4SDLC: How AI Changes Development Processes in Large Companies

Mass adoption has already happened: AI is now an everyday engineering tool, yet its impact is uneven. An individual developer may produce code, tests, or explanations faster, while the gain becomes ambiguous at team and organizational levels. The flow receives more changes, so queues move into intent definition, review, testing, integration, and release. The share of generated lines says nothing about whether work reached users safely or at an acceptable cost.

In a role-based SDLC, every profession can optimize its own stage while creating more work in progress for the next one. The proposed alternative is an agent-based SDLC in which an engineer can own a broader end-to-end scenario and reduce handoffs. Specifications return not as heavyweight documents but as contracts with agents: intent, context, boundaries, acceptance criteria, and a way to verify the outcome. Weak intent does not disappear; AI merely scales the resulting error faster.

At the scale of 10,000+ engineers, thousands of services, and strict fintech controls, agentic development cannot rely on direct model connections and an accidental toolbox. It needs distinct layers: a model gateway for identity, quotas, cost, secrets, and personal data; a tool gateway for controlled calls; and a capability registry with owners, versions, policies, and evaluation sets. Platform-owned domain agents should encapsulate complex semantics instead of forcing a generic assistant to reconstruct them on every task.

The talk maps this design to existing building blocks: unified model access, an MCP Hub and terminal client, the Nessy corporate-context assistant, the Spirit development platform, and domain agents for coding, testing, review, security, and operations. Scenarios have different maturity levels. Read access can open first; recommendations need an inspectable plan; state-changing actions require policy as code, approvals, isolation, audit, rollback, and a kill switch. Autonomy belongs to a specific capability rather than to an agent as a whole.

Measurement must connect adoption to speed, quality, risk, and economics. Instead of AI-generated lines of code, the talk combines DORA, SPACE, and DevEx signals: time to first merge request, pipeline and review waiting, rework, defects, and incidents. Agentic scenarios also need traces from intent to outcome, including the user, agent, model, tool calls, approvals, cost, and result. Evaluation sets and hidden judges make it possible to regression-test whether the agent actually completed the task rather than merely producing a convincing answer.

Teams using identical tools achieve very different results because AI adoption is a sociotechnical change. Strong gains appear where requirements become verifiable contracts, context becomes machine-accessible, and review and checks are redesigned around the new flow. Engineers type fewer lines and spend more time defining intent, orchestrating work across editor, terminal, and background agents, and validating system behavior. Accountability remains human, while domain knowledge, design, and engineering judgment become more valuable.