Human intent
State the work and expected outcome in language the domain team owns.
Organisation and ownership
Centralise the road. Keep ownership distributed.
A core AI team should create reusable foundations and help teams adopt them. It should not inherit every domain decision or become responsible for every failure.
How work should flow
Intent enters. Evidence comes out.
An AI can suggest, plan and act. It should not be the only party deciding that its own work is correct.
Bounded task
Define what the agent may do, where the step ends and what is forbidden.
Agent and tools
Give the model enough context and the smallest useful set of actions.
Evidence
Check system state, captured values, tool history or human acceptance.
Close or recover
Pass, retry, stop or escalate with a reason another person can follow.
The prompt guides. Code still owns permissions and hard stops.
The agent acts. Evidence—not confidence—supports completion.
The team owns. AI does not erase accountability.
How I would scale
A core AI team, with shared ownership.
The central team builds a reusable road. Domain teams drive on it and still own where the work needs to go.
Domain teams
Own the meaning
- Workflow and exceptions
- Domain tools and knowledge
- Acceptance criteria
- Final outcome sign-off
AI platform & enablement
Build the shared road
- Agent foundation and tool standards
- Evaluation and observability
- Common controls and cost visibility
- Coaching, patterns and adoption
SRE and Security
Remain independent gates
- Reliability requirements
- Security and data policy
- Production readiness
- Incident and audit standards
What I measure
Usage is a signal. Outcome is the result.
The management view is quality-adjusted productivity: useful delivery after including review, rework, defects, safety and cost.
Task success
Did the requested outcome actually happen?
Review and rework
How much correction made the output usable?
Defect escape
Did speed move defects into a later stage?
Safe execution
Were permissions and approval boundaries respected?
End-to-end time
Did the complete workflow become faster?
Cost per success
What did each accepted outcome actually cost?
The practical foundation