The standing rule: agents orchestrate, extract, monitor and summarize —
they never author or change a verdict. Every score comes from the deterministic engine (same evidence in, same verdict out), and every status change passes a human. Agents write to alert/staging layers only.
Traced, in your tenant: in the tenant build, every agent step and every AI call is traced into the group’s own telemetry (Application Insights on Azure; OpenTelemetry) — prompt, evidence returned, verdict, and which human acted. Same tenant, same sign-in, no new vendor. Tenant build · the V1 prototype logs locally.
Enrichment Engine● LIVE
Turns a bare app name into a structured catalog record — vendor, policy URL, category, age range, description.
Runs: Runs when a new application is added or requested.
🔒 Proposes the record; a human confirms category and status.
Round-1 Extractor● LIVE
Reads a vendor’s privacy policy and answers the 19 triage questions with verbatim citations (temperature 0).
Runs: Runs once per vendor at vetting; answers are cached.
🔒 Extracts evidence only — the deterministic engine computes every verdict.
Drift Sentinel● LIVE
Re-checks approved vendors’ policies, re-scores with the same engine, and raises an alert when a verdict degrades or a gate flips. The full loop works today; the scheduler that makes it autonomous ships with the tenant build (timer-triggered function), along with live policy re-fetch and Teams alerts.
Runs: On a schedule (weekly by default, a setting) and on demand via “Scan now”. Re-fetches each approved vendor’s live policy.
🔒 Writes alerts only — never patches a catalog verdict. A human triggers the re-vet.
Document Intake◌ PLANNED · build #1
The front door for documents: watches a shared drop folder (SharePoint in the tenant build) per school, reads each new file — contracts, DPAs, SOC 2 reports, completed questionnaires — and extracts the key fields with a verbatim quote behind each. One intake lane feeds both the Contracts board and Round 2.
Runs: Scheduled folder scan + on file drop.
🔒 Extracts and quotes only; a human confirms every field before it lands in the system of record.
Approval Orchestrator◌ PLANNED · build #2
Owns the workflow operationally: routes by cost tier and category, notifies the right approver as a Teams/Outlook adaptive card, nudges stalls before SLA breach, closes the loop with the requester. Runs inside the group’s cloud tenant — the catalog’s storage stays the system of record; the service-desk ticket mirrors, never drives.
Runs: Event-driven on workflow state changes.
🔒 Delivers notifications and tracks SLAs; routing rules stay in code.
Contract Sentinel◌ PLANNED · build #3
Makes the Contracts dashboard proactive: watches renewals and auto-renew cancel-by deadlines and surfaces renew / renegotiate / drop decisions. (Today the dashboard computes renewal traps deterministically; this agent adds the watching — Document Intake supplies the parsed terms.)
Runs: Scheduled weekly + when Document Intake lands new terms.
🔒 Surfaces decisions only; the deadline math stays deterministic.
Governance Brief◌ PLANNED · build #4
A Monday-morning digest per region lead: new requests, pending approvals, the week’s drift alerts, top consolidation opportunities, upcoming re-vets.
Runs: Scheduled weekly.
🔒 Summarizes deterministic output; adds no new judgments.
Regulatory Watch◌ PLANNED · build #5
Monitors privacy-law changes across the group’s jurisdictions and flags which approved vendors now have a coverage gap for an affected country.
Runs: Scheduled weekly scan + on-demand.
🔒 Reports a jurisdiction gap; never re-scores. Gap → human decision.
Lifecycle Sentinel◌ PLANNED · build #6
Watches vendor news the policy can’t show — product retirements, acquisitions, breaches, renames — and drafts a cited override or re-vet proposal for reviewer sign-off.
Runs: Scheduled weekly per catalog vendor + before any re-vet.
🔒 Researches and proposes; only a named human applies an override, and the engine re-scores.
Repository Copilot◌ PLANNED · build #7
Natural-language Q&A over the catalog and reports — “which approved apps have an unclear AI-training gate?” — with citations to app, question and quote.
Runs: On demand.
🔒 Quotes deterministic output; never re-scores.
Portfolio Analyst◌ PLANNED · build #8
Turns the overlap math into leadership-ready recommendations — which duplicate apps to consolidate and what it saves per school per year.
Runs: Folded into the Governance Brief.
🔒 Narrates existing deterministic math; adds no new numbers.
Planned agents are roadmap scope — named, designed and sequenced, not yet running. They ship during rollout, tailored to each region’s workflow.