What the Top Companies Do Differently With AI Coding - Martin Harrysson & Prakhar Dixit, McKinsey
AI coding gains come from redesigning the full delivery process—not simply adding tools—and grounding agents in trusted context, standards, and human review.
- Replace two-week sprints with short agent-driven loops; set human-review gates based on confidence.
- Standardize requirements and curate golden examples, policies, and enterprise context for agents.
- Measure delivery value and quality, and include release, risk, compliance, and design requirements in workflows.
The talk shares organizational patterns, but offers few detailed implementation demonstrations. · Best for: Engineering leaders scaling AI-assisted development across large organizations.
McKinsey speakers describe why enterprise AI coding gains depend on coordinated changes to workflows, team structure, context, quality controls, and measurement. They report that some companies are seeing substantial end-to-end improvements, while adoption alone has not reliably produced them.
From tool adoption to end-to-end impact1:13
- Many large companies initially reported modest gains despite broad AI use.
- The speakers say a leading group is beginning to see larger end-to-end improvements.
- They frame impact as a coordinated transformation rather than a tooling change.
Redesign workflows and teams3:40
- Shift from two-week Scrum cycles toward one- or two-day Kanban-style loops, with agents working and humans validating outputs.
- Place human review where confidence is low; expand autonomy as loops mature.
- Use standardized artifacts such as PRDs and specifications to supply consistent inputs.
- Organize work in pods of two to four, with responsibilities increasingly spanning traditional roles.
Build context, governance, and skills curation8:03
- Choose and organize tools around the organization's context; the speakers consider available tooling broadly capable.
- Document expert knowledge so agents can use it instead of depending on individual subject-matter experts.
- Curate a small set of relevant skills rather than adopting every available option.
- Plan governance, security, stability, change management, and AI FinOps before scaling beyond pilots.
Protect quality and address shifting bottlenecks11:35
- The speakers report quality declines even as code generation speeds up, and stress training developers to review generated work.
- Feed agents curated golden examples—not simply average code—and encode quality standards throughout the delivery cycle.
- As build and test accelerate, requirements, release planning, and coordination with risk, compliance, governance, and design can become bottlenecks.
Measure value and close16:07
- Pair AI adoption measures with established software-delivery measures; the speakers say many organizations still lack confidence in impact measurement.
- Some clients are exploring AI to assess the value or intent of work and guide resource use.
- Smaller pods and more teams can offer a near-term way to increase throughput, but measurement and operating practices need to mature.
Step by step
- 1. Map the current delivery process and identify where agents can run short, repeatable loops.
- 2. Standardize the requirements and other artifacts each loop consumes.
- 3. Set human-review gates, then expand autonomy as teams gain confidence.
- 4. Document enterprise context and provide agents with curated examples, standards, and policies.
- 5. Encode quality, release, security, compliance, and design requirements across the process.
- 6. Track delivery outcomes, quality, and value—not just tool adoption—and adjust the workflow.
Notes are AI-generated from the video's transcript and may miss nuance.