The intake pipeline
Lesson 3 of 5 in AI Inventory, Use-Case Intake, and Risk Triage.
The inventory captures what exists; intake captures what is about to exist. It is the front door of the operating model you built in the last module — and its design determines whether people walk through it or around it.
The intake form is a triage instrument, not an interrogation. Ask what a triage decision needs and nothing more: purpose and affected population, decision consequence (does it deny, price, rank, or recommend something for a person?), data categories, build/buy/embedded, autonomy level, jurisdictions, underlying model. Fifteen questions a requester can answer in twenty minutes. Every additional page of speculative detail raises the cost of honesty — and the shadow estate grows in exact proportion to the friction of the legitimate path.
Three or four of those questions are gating questions — the ones that can end the process immediately or reroute it entirely: Does it involve a practice on our prohibited list (or the EU AI Act’s)? Does it make or materially influence a consequential decision about individuals? Does it process special-category data? Is it customer-facing generative AI? A yes on a gate skips the queue: prohibited-list hits stop, the rest route straight to elevated review.
Key terms: intake, risk tiering, gating question, conditional approval, proportionality
Intake to decision: the pipeline with its clocks
- Requester submits intake form
Fifteen questions, twenty minutes. Auto-saved into the registry as a draft record — intake and registry are one system, not two.
- Completeness check (SLA: 2 days)
A triage analyst confirms the form is answerable and coherent. Incomplete forms bounce back fast — a silent queue teaches requesters the process is where requests go to die.
- Any gating question triggered?
Prohibited practice? Consequential decision? Special-category data? Customer-facing genAI?
- Stop: prohibited practice
Hits on the prohibited list (internal or statutory) end here, with reasons documented and an appeal route to the executive risk committee only.
- Tier assignment (SLA: 5 days)
Analyst applies the tiering scheme (next lesson) and records the rationale. Precedent library consulted — same-shape cases inherit prior decisions.
- High tier or novel?
Charter thresholds from the operating-model module decide the lane.
- Fast lane: delegated approval (SLA: 10 days)
Checklist review by the delegated approver or spoke lead. Second line samples these quarterly for calibration.
- Full review: assessments + committee
Required assessments commissioned (see the impact-assessments module), then committee decision with quorum.
- Approve / conditions / reject
Conditional approval is the workhorse outcome: proceed, subject to named conditions with owners and due dates.
- Registered, tiered, monitored
The record flips from draft to active; tier-appropriate monitoring and attestation obligations begin.
Two design details carry most of the pipeline’s credibility. SLAs, published and kept: a requester who knows the answer arrives in ten days will use the front door; one who has watched a colleague wait a quarter will not. Track cycle time by tier as a first-class program metric — it is your shadow-AI leading indicator. Conditional approvals with teeth: "yes, if" is the outcome that keeps proportionality honest — approve the pilot with a human reviewing every output, approve production contingent on the fairness retest, time-box the approval to twelve months. But every condition needs an owner, a due date, and an automatic consequence on breach, or conditions become the polite spelling of "yes".
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.