Redesigning the AI engineering team structure before unclear roles slow everyone down
Engineers working with AI tools now spend more hours reviewing generated code than writing new code. Digital Applied’s Q1 2026 survey of 2,847 developers recorded the crossover, with review overtaking writing as the largest AI-assisted time sink after writing held a four-hour lead as recently as 2024.
Over the same period engineering managers have moved in the opposite direction. They are now more technically hands-on than they have been in a decade as Agentic AI lowers the barrier to direct code contribution.
Both trends meet in the middle of the org chart. The division of responsibilities between engineer and EM was doing structural work that few leaders ever named, and coding agents are dissolving it with nothing arriving to replace it.
The Job Descriptions Stopped Matching the Work
Look at how a senior engineer on an AI-heavy team actually spends a Tuesday. She kicks off two agent runs before standup, reviews a stack of generated pull requests mid-morning, fixes a prompt configuration that started producing flaky tests, and switches contexts across three tools before lunch.
The Pragmatic Engineer’s 2026 survey of over 900 engineers and engineering leaders captured exactly this. Engineers orchestrate more and context-switch more often, managers can be more hands-on, and the survey’s authors flagged the conclusion themselves. The engineer and manager roles are becoming similar.
Managers are converging from the other side. An EM can now ship a fix or prototype a feature between one-on-ones, and plenty of them do. The technical distance that used to accumulate after two years in management has stopped accumulating.
The Old Division Was the Long Pole
The classic split assigned implementation quality to engineers and gave managers allocation, priorities, and people. AI broke this in both directions.
When agents produce the majority of a feature’s code, the person directing them is making allocation decisions. Which tasks go to the machine, which stay human, how much scrutiny each output deserves. That used to be manager territory. And when a manager merges their own AI-assisted fix, they have re-entered the codebase their role was designed to stay out of.
What Role Confusion Costs
One client I worked with last year had three teams improving the same AI workflow at the same time.
Engineering upgraded the model to reduce latency. The AI team refined prompts and retrieval settings to improve answer quality. Operations updated the business rules the agent was expected to follow.
Each team shipped good changes. Each team achieved its own goals. A month later, overall accuracy had dropped.
No single change caused the problem. It was the interaction between all three. Everyone was optimizing their part of the system, but no one owned the system itself.
We uncovered it during a retrospective and established a single owner for end-to-end evaluation, along with shared metrics across the teams.
That problem happened to be visible. Most aren’t.
A Tilburg University study of Copilot adoption in open-source projects found core developers reviewing 6.5% more code while their original output dropped 19%. Stack Overflow’s 2025 survey found 45% of developers citing time-consuming debugging of AI-generated code as a top frustration. And in a March 2026 SmartBear survey of 273 software leaders, 70% said application quality had already degraded as AI accelerated development.
Those numbers tell a consistent story. Code production is accelerating faster than organizational ownership.
When an engineering manager merges agent-generated code and a production incident surfaces two weeks later, who owns the postmortem? The engineer who approved the pull request? The team that tuned the prompts? The platform team that selected the model? The product manager who defined the workflow?
Teams without a clear answer pay for AI twice. Once for the tokens, and again for the coordination overhead of figuring out whose responsibility the output became.
A Redesign That Fits on One Page
The fix requires less machinery than most reorgs. In my work with engineering teams adopting agentic workflows, four decisions cover most of the confusion.
Name one accountable reviewer per code surface. Agent-generated pull requests get a single human owner, assigned by code area and written into the CODEOWNERS.md file. That owner can be an engineer or an EM. What matters is that exactly one name appears, so accountability for quality survives the increase in volume.
Give manager code contribution explicit rules. If an EM ships code, it goes through the same review path as everyone else’s, and its scope stays bounded. Prototypes, internal tooling, and spikes work well. Critical-path features do not, because a manager who owns production code has become an engineer with a reporting-line problem.
Put orchestration in the engineer job description. Hours spent directing agents, writing evals, and maintaining prompt configurations should count as engineering work in performance reviews. If promotion criteria still reward hand-written lines, engineers will optimize for the old job while the actual work goes unmeasured.
Rebuild the EM role around what AI left behind. Stakeholder negotiation, cross-team decisions, career development, and the judgment calls agents consistently fumble. Those responsibilities gained value as everything around them got automated, and a manager whose calendar reflects that is doing the redesigned job instead of competing with their own engineers for the review queue.
Then revisit the whole arrangement quarterly. The tools are changing fast enough that a role definition written in January describes a different workflow by June.
An org chart is a claim about how work gets done. Each quarter the chart goes unedited while the work underneath changes, the claim gets a little less true. The teams outperforming with AI-assisted delivery share one habit that costs nothing to copy; they wrote down what changed. Engineers who manage agents, managers who touch code, and one name on every review.
Role convergence turns out to be a design problem, and design problems reward the leader willing to name them.
…
Nick Talwar is a CTO, ex-Microsoft, and a hands-on AI engineer who supports executives in navigating AI adoption. He shares insights on AI-first strategies to drive bottom-line impact.
→ Follow him on LinkedIn to catch his latest thoughts.
→ Subscribe to his free Substack for in-depth articles delivered straight to your inbox.
→ Watch the live session to see how leaders in highly regulated industries leverage AI to cut manual work and drive ROI.
Top comments (0)