The model may be rented. The business behaviour, memory and evidence around it should remain governed by the organisation.
The invisible asset around the model
An agent is more than a prompt and a model endpoint. Its harness contains instructions, tools, context assembly, memory, permissions, evaluations, approvals and operational traces.
This layer captures how the company wants work to be performed. Losing it creates dependency even when model portability appears easy.
What the organisation must govern
Vendors can provide models, frameworks and managed infrastructure. The organisation still needs clear ownership of the assets that encode its operations.
- business definitions and policies
- critical prompts and workflow logic
- evaluation cases and acceptance thresholds
- identity, permissions and approval rules
- connectors, traces and decision memory
A practical shared-responsibility model
Business owners define outcomes, rules and acceptable risk. IT owns identity, integration and operations. Data and knowledge owners qualify sources. Delivery partners may build components but must document and transfer them.
Every critical component needs an owner, version, export path and replacement procedure.
Test reversibility before scale
Before expanding the agent estate, export prompts, evaluation sets, tool schemas, memory records and logs. Replace one model or connector in a controlled test.
If the workflow cannot move without losing business behaviour, the organisation does not yet own the strategic layer.
CONNECT THE DECISIONS
Place this analysis inside the wider AI system
Durable ownership requires connecting enterprise memory, the components of an AI agent harness, governance of SharePoint sources and contextual architecture connected to core systems.
FREQUENTLY ASKED QUESTIONS
FAQ
Does ownership mean self-hosting everything?
No. It means retaining control, portability and governance of the business-specific assets, even when infrastructure is managed.
Who owns the evaluation set?
The business should be accountable for expected outcomes, with technical teams maintaining the executable evaluation system.
