I’m bullish on AI in delivery—but not because it “runs the project for you.” The real win is simpler: it reduces the noise that slows teams down, and it helps people make better decisions with the information they already have.
Used well, AI can make planning and execution feel lighter. Used poorly, it becomes another dashboard that people ignore (or worse, trust when they shouldn’t). Here’s where I’ve seen it genuinely help, where it doesn’t, and how I’d implement it without creating new risk.
Where AI is genuinely useful
The best AI workflows sit close to the day-to-day friction points. They don’t replace judgment—they remove busywork and surface signals earlier.
- Status synthesis: Summarize updates from tickets, PRs, and incidents into a clear “what changed” narrative.
- Risk scanning: Highlight patterns like recurring blockers, long-lived tickets, or build/QA instability.
- Triage: Group issues by root cause candidates (config, environment, regression, dependencies) to speed up response.
- Decision support: Create short options: trade-offs, dependencies, and what you’d give up by choosing each path.
Where AI tends to disappoint
AI struggles when the problem is ambiguous, political, or under-specified. It also struggles when you ask it to “know” things that live in people’s heads.
- False certainty: Confident summaries that hide missing context (the classic “sounds right” problem).
- Shallow planning: It can propose timelines, but it can’t feel the constraints of a real system.
- Process theater: Auto-generated reports that look polished and say nothing new.
- Security/compliance blind spots: If you haven’t defined boundaries, it will happily cross them.
A practical implementation approach
If I’m rolling this out inside an enterprise environment, I start small and make it measurable. The goal isn’t “AI everywhere.” The goal is fewer surprises and faster decisions.
- Pick one workflow (for example: weekly status + risk review) and instrument it.
- Define success: fewer manual hours, fewer missed risks, faster time-to-triage.
- Constrain inputs to approved sources (tickets, docs, runbooks) and log everything it touches.
- Human-in-the-loop: AI drafts, a human approves. Always.
- Iterate: tune prompts, add guardrails, and retire what isn’t delivering value.
My rule of thumb
If AI is producing content that nobody reads, stop. If it’s producing insight that changes decisions (or prevents a future incident), keep investing.