AI adoption problems usually show up as alignment problems before they show up as technical failures.
The model may be good. The agent loop may be competent. The repository may even be reasonably well-instrumented. The adoption can still fail if engineering, platform, security, product, and management believe they are running different experiments.
The stakeholder matrix
| Stakeholder | Decision they need | Misalignment signal |
|---|---|---|
| Individual engineer | Which tasks are safe for AI assistance? | Large unreviewable diffs, hidden failing tests, drive-by refactors. |
| Team | Which repos and task classes enter pilot? | Templates feel like bureaucracy, reviewers see more low-quality PRs. |
| Architecture | Which system boundaries are authoritative? | Agents optimize per repo while violating cross-system invariants. |
| SRE / Platform | What telemetry and rollback evidence is required? | Agents propose operational changes without diagnosis capacity. |
| Security | What data and tool permissions are allowed? | Agent access expands faster than secrets and IAM boundaries. |
| Management / PM | What evidence counts as progress? | AI work becomes invisible in planning or reduced to vanity metrics. |
The common failure mode
The common failure mode is not that an agent cannot write code. The common failure mode is that the organization cannot agree what the code change means.
Is the PR a prototype, a production candidate, a refactor, a compliance-sensitive change, a platform primitive, or a local optimization? If that classification is missing, the agent will proceed as if the classification does not matter.
It does matter.
Decisions before automation
Before increasing autonomy, teams need decisions in writing:
- Which task classes may be agent-assisted?
- Which task classes require a human-only first draft?
- Which files, systems, or permissions are off-limits?
- Which checks are mandatory?
- Which review role owns the final decision?
- Which metrics prove the pilot is making work safer, not merely faster?
If these decisions do not exist, the right move is not a more powerful agent. The right move is to make the operating model explicit.
What to watch
Misalignment tends to produce a few repeatable signals:
- PRs become harder to review even as output volume rises.
- The issue tracker no longer explains the work.
- Agents make plausible changes without system context.
- CI becomes a rubber stamp.
- Rollback is described as “revert if needed” when the system requires sequencing.
- Productivity claims outrun evidence.
These are not reasons to abandon AI engineering. They are reasons to lower the delegation level until the platform, governance, and team workflow catch up.