Approach
You decide at every gate.
Five stages. Each ends in evidence and a decision: stop, reshape or continue.
Five stages, five decisions
Open a stage for what you get and what you decide.
Diagnose
Map the process, owners, systems and today’s performance.
Your decision Does this domain merit investment?
Redesign
Define the target process and who decides what: rules, AI or people.
Your decision Is it worth building, and buildable here?
Prove
Test one hard case, including what fails.
Your decision Does the evidence justify integration?
Integrate
Connect systems with monitoring, limits and rollback.
Your decision Do the accountable owners accept it?
Operate
Train people, release by scope, review results against the baseline.
Your decision What expands, changes or stops?
See it applied to project-information operations. Compare: more bots or a redesign? · Big Four or founder-led team?
Three rules in every stage
- Rules firstKnown checks and status changes run as fixed rules.
- AI where it helpsReads remittances, emails, PDFs and drawings; compares revisions. Doubtful cases go to a person.
- People decideEvery consequential decision has a named owner.
Human approval and a way back for AI
Before go-live, every AI-enabled process needs two things written down: who decides, and how to undo it. Sembium designs both in the Redesign and Integrate stages, not as policy added afterwards. Every route also has an exception path and a named incident owner.
Decision rights and recovery (illustrative)
The rows are examples. You set the thresholds.
| Action | Who or what decides | Stop and recovery |
|---|---|---|
| Check fields, known terms, status changes | A fixed rule | Rule changes are reviewed and reversible |
| Classify or summarize incoming information | A rule checks it; a person reviews uncertain cases | The step can be switched off; cases fall to a manual queue |
| Update a system of record | A rule within a permitted range; a person above it | Reversible; a named owner can revoke the permission |
| Material financial, commercial, engineering or legal decision | A named accountable person | Escalation path agreed before release |
Why this is designed in
OWASP’s 2025 LLM risk list and NIST’s AI Risk Management Framework both call for human oversight and a way to switch off. NIST is voluntary guidance, not a certification.
In a program these rows become acceptance criteria: nothing goes live until each step has a named owner and a tested way back.
Sources
Sources give context. Dates show when each was last checked.
Start at the first gate.
One 30-minute call about one bottleneck. We solve bottlenecks worth $1M or more.