Martin Harrysson, Prakhar Dixit and their colleague Shilpa discuss patterns they say they see in enterprise software teams, at AI Engineer.
Change the workflow, not just the tool
- They describe shorter, one- or two-day cycles where agents work and humans review, with review points adjusted as confidence grows.
- Some clients are organizing into pods of two to four and combining responsibilities; the speakers say smaller teams do not necessarily mean fewer employees.
- Standardized product requirements and specifications give agents consistent inputs; documenting expert knowledge helps supply organizational context.
- Curate a small set of relevant agent skills rather than adopting every available workflow. Include security, governance and cost in the rollout.
Quality and bottlenecks move
- The speakers say faster generation can coincide with more defects, including in their reported per-developer and per-pull-request observations; they do not provide underlying data in the talk.
- Teach review as well as generation, supply carefully chosen examples of good code, and specify quality checks throughout development.
- Faster build and test work can shift constraints upstream to requirements, downstream to release, and sideways to design, accessibility, risk and compliance.
- A bank’s accessibility and color requirements illustrate why adjacent teams’ rules need to be considered early.
Measure the outcome carefully
- Harrysson says some clients experiment with AI-assisted scoring of work’s value, but admits that reliable impact measurement remains difficult.
- McKinsey’s speakers claim that roughly 20–25% of surveyed companies are seeing 2× end-to-end improvements; the talk does not establish the survey methods, baseline or independent causality.
- Their reported 2×–3× improvements at the end refer to multiple unspecified metrics, not a demonstrated universal productivity gain.
“It’s not changing one thing or putting in a new tool and you’re going to see this happen.” — Martin Harrysson