Field Note · Applied AI and system ownership
What Should You Build, Rent, Own, or Keep Human When AI Enters the Operating Layer?
The model is rarely the moat. The harder question is whether the business still owns the context and method that make the model’s work useful.
A thinking frame by Andrew Moss
The questions I get
Usually some version of these:
- Should we build or buy this AI capability?
- What data and context should remain ours?
- Which decisions should never become fully automatic?
What a lot of people seem to think
Companies often frame the decision as custom software versus an off-the-shelf tool and assume the durable advantage lives in model access or a large pile of data.
How I look at it
Rent the commodity pipes. Own the memory, method, permissions, and accumulated correction layer that make the work distinct. The moat is not every output the system has produced. It is the verified record of what good looks like, which failure modes matter, which exceptions were approved, and how the next use improves.
Why the decision matters
The cost is rarely confined to the line item.
If the sequence is wrong
The business rents away its institutional memory, becomes dependent on a vendor’s abstractions, or automates decisions nobody can explain or correct.
If the sequence is right
Commodity components remain replaceable while differentiated context and learning compound inside an accountable operating system.
How reversible is it?
Mixed. Tools can be replaced; lost context, opaque history, and embedded habits are much harder to recover.
The short answer
Own the correction layer, not every component.
Use replaceable models and tools where they are commodities. Preserve the source-grounded context, evaluation standards, approved exceptions, human decisions, and confirmed learning that should survive a vendor change. Keep consequential judgment human until evidence supports a narrower delegation.
The compounding assetModel access → output. Verified correction → better next judgment.
Raw memory can compound error. A correction layer compounds only when the source, reviewer, reason, permission, and outcome remain visible.
Move fromA stack of AI outputs→Move towardA governed record of verified learning
The order I would use
Take the right steps in the right order.
- 01
Separate commodity from differentiation
List the replaceable models, storage, interfaces, and orchestration; then list the context, methods, relationships, and judgment standards that make the work yours.
- 02
Name what must survive a vendor change
Protect source mappings, decision history, permissions, evaluations, reusable artifacts, and approved learning.
- 03
Design the correction unit
Capture what was wrong, what the source supports, who approved the correction, why it matters, and where it may be reused.
- 04
Test the tenth use
Ask whether verified corrections make later work more accurate, useful, or economical than the first use.
- 05
Keep consequential authority human
Delegate only where the evidence, boundary, reviewer, and reversal path are clear.
Questions worth answering
Before the next irreversible move:
- What would still matter if the model vendor changed tomorrow?
- Which corrections should improve the next use, and who verifies them?
- Does memory preserve provenance and permissions or merely accumulate text?
- Which decision is too consequential to delegate yet?
What not to do
Do not confuse storage with learning.
Do not retain every output, treat every correction as universal, or let an unchecked memory become the source of truth. Do not lock differentiated knowledge inside a vendor whose role should remain replaceable.
Keep the perspective
The moat is what the system has learned responsibly.
Access to a capable model will spread. A source-grounded correction layer tied to the business’s real methods, relationships, approvals, and outcomes can become harder to copy with every verified use.
Independent sources
Useful primary material
These sources support the public frame. They do not replace the private facts or the accountable professional.
Common follow-up questions
Does owning context mean hosting every model?
No. A business can use external models while retaining control of its sources, memory, methods, permissions, artifacts, and evaluation history.
What work should stay human?
Work where consequence, ambiguity, relationship judgment, accountability, or poor reversibility exceeds the evidence that the system can handle it safely.