AI in the SDLC · 11 Jun 2026

AI across the SDLC: every phase of delivery, rewritten.

The interesting story isn't that AI writes code. It's that every phase of the software development lifecycle — from requirements to maintenance — now has an AI-shaped seam running through it.

Ask most people how AI changed software development and they'll say "autocomplete got really good." That undersells it. In the teams we work with, AI has quietly become part of every phase of the lifecycle — and the teams getting the most out of it treat it as a delivery-process redesign, not a tooling upgrade.

Here's what that looks like, phase by phase.

Requirements & planning: from documents to dialogue

Product managers now draft PRDs in conversation with an LLM that has retrieval access to past tickets, support transcripts and analytics. The result isn't that requirements write themselves — it's that the first draft arrives with edge cases, conflicting constraints and open questions already surfaced. Estimation improves too: models grounded in your delivery history are uncomfortably good at flagging the epic that will blow up.

Design & architecture: faster options, same judgment

Architecture is still a senior-human game, but the loop has changed. Instead of one proposal debated in a meeting, teams generate three candidate designs with explicit trade-off tables, then spend their senior attention on choosing and stress-testing. AI is also brutal at consistency checks — "where does this break our existing API conventions?" is now a five-minute question.

Implementation: the agent writes the first draft

Coding assistants have graduated from line completion to bounded delegation: hand an agent a well-scoped ticket and it produces a branch — code, tests and a self-review note. The developer's job shifts from typing to specifying, reviewing and integrating. Two practices separate teams that win here from teams that drown: tickets written with acceptance criteria precise enough for an agent, and a hard rule that nothing merges without human review.

Code review: a tireless first pass

AI review bots now catch the mechanical 60% — null paths, race conditions, divergence from team conventions, missing tests — before a human ever looks. That doesn't replace review; it upgrades it. Human reviewers spend their time on the things models are weakest at: intent, design fit and "should we build this at all?"

Testing & QA: coverage stops being the bottleneck

Generating unit tests, property-based cases and realistic fixtures is now nearly free, which moves the conversation from "do we have tests?" to "do we trust them?" The new QA discipline is curating AI-generated suites — deleting the tautological ones, keeping the ones that encode real behavior — and adding evaluation harnesses for any feature that itself calls a model.

Release & operations: the loop closes

CI failures arrive with a suggested diagnosis. Release notes draft themselves from the merged PRs. In production, AIOps tooling clusters alerts, drafts incident timelines and proposes runbook steps — with a human approving anything that touches state. Postmortems that used to take a week of archaeology start from an AI-assembled timeline in minutes.

What doesn't change

Accountability. The lifecycle still belongs to people: someone decides what to build, someone owns the merge, someone carries the pager. AI compresses the distance between intent and running software — it doesn't absorb responsibility for either. Teams that forget this ship faster for a quarter and then spend two quarters cleaning up.

We've rebuilt our own delivery process around these seams, and we help clients do the same. If you're rethinking how your team ships in the AI era, we should talk.

Rethinking how your team ships software?

We help engineering organizations redesign their SDLC around AI — tooling, process and governance included.