ENT-AI-025 / ENTERPRISE TECHNOLOGY
Decision support learned when not to decide
A regulated enterprise introducing retrieval and recommendation tools across operational review teams.
Context
Several teams built useful prototypes with different data sources, confidence language, review steps, and incident paths.
The problem beneath the brief
The organization governed models but not decisions. No shared lifecycle connected use approval, data access, evaluation, human review, monitoring, correction, and retirement.
- 31
- decision uses inventoriedapproved use registry
- 5
- risk/authority tiersdecision-rights model
- 100%
- active uses correction-linkedoperating acceptance query
Risk constraints
What could not be traded away.
- role and data boundaries
- automation bias
- source provenance
- model/provider change
- human appeal and correction
Findings
What inspection changed.
- risk tiers described technology rather than decision consequence
- reviewers could not see source freshness
- provider changes could alter behavior without reopening acceptance
Architecture
The operating system we installed.
- 01use-case and decision registryrisk-tier decision rights
- 02approved-source layersource freshness
- 03evaluation and acceptance serviceprovider/version change gate
- 04human review/overrideabstention
- 05monitoring and correction workflowappeal and correction
Delivery sequence
Four gates. No ceremonial phase changes.
- 01
Frame
Define the decision, outcome, work products, authority, dependencies, exclusions, and acceptance evidence.
A named sponsor and principal approve the bounded charter. - 02
Inspect
Observe the operating reality, trace systems and records, test assumptions, and rank failure modes.
Critical unknowns have owners, evidence plans, and stop conditions. - 03
Build
Implement the smallest coherent change with versioned decisions, controls, and verification attached.
The integrated state meets the agreed evidence threshold. - 04
Transfer
Rehearse recovery, resolve exceptions, accept the work, remove temporary access, and transfer operating ownership.
The receiving owner signs the handoff with open limits visible.
Complications
Where the plan had to become more honest.
- A low-risk summarization feature influenced a high-risk downstream decision.
- Source freshness mattered more than model performance for one operational use.
Outcomes
What changed—and what the record proves.
- Every deployed use resolved to a named decision owner, source set, model/version, evaluation, and correction path.
- High-consequence uses could abstain and route to qualified review.
- Provider and source changes reopened the right acceptance gates.
Lessons
What we would carry into the next system.
- Govern the decision, not only the model.
- Source freshness can dominate model quality.
- A downstream use can raise the risk of an otherwise ordinary feature.
Handoff
The engagement ended with an operating owner.
- 01use-case owner register
- 02evaluation versioning
- 03change-control triggers
- 04human review and appeal
- 05quarterly retirement review
Start with the decision
Bring the priority. We will help bound the work.
If the decisions or constraints look familiar, start with the operating reality—not a preselected solution.
Start a conversation.