AI systems /
Forward deployed engineers are useful. Keep the capability.
AI projects often fail away from the model.
The data is messy. The actual workflow is undocumented. The team disagrees about who can approve an action. A seemingly small exception turns out to protect a customer promise.
Forward deployed engineers help because they can work across that gap. They can understand the business process, build the system and challenge a request that would create risk.
That makes them useful. It does not make them a substitute for internal ownership.
The risk is rented understanding
An external team can build a capable agent. It can also become the only group that knows why it behaves as it does.
They know which prompt changed. They know the failure cases. They know why an approval exists. They know where the data comes from.
The company uses the system but cannot evaluate or safely change it.
That arrangement can look successful until something changes. A model update causes errors. A team wants to extend the workflow. A compliance question arrives. The internal champion leaves.
Then the company finds that the judgement needed to run the system lives outside the business.
What has to stay inside
Specialists should accelerate the work and transfer knowledge as they build. They do not need to disappear after launch, but the company needs enough understanding to make informed decisions.
Five things should remain under internal control:
- The workflow: the steps, handoffs, approvals and failure paths.
- The knowledge: policies, facts, examples and exceptions the system relies on.
- The evals: examples that show whether the output is good, unsafe or incomplete.
- The permissions: what the system can read, write and escalate.
- The evidence: logs, citations, review notes and reports of failures.
These are not implementation details. They are the operating instructions for the system.
Ask about the handover before the build
Before hiring an AI team, ask what your company will be able to do when the engagement ends.
Can an internal operator add new knowledge? Can someone read a trace and understand a failure? Can the business pause the workflow? Can the team test a model or prompt change against its own examples?
If the answer is no, the engagement may produce a working demo. It will not build lasting capability.
The rule
Hire specialists to move faster.
Keep the understanding needed to run and improve the system inside the company.