Provenance is a governed fleet of eight AI agents — spanning data, action, and governance — that cleans up the records your CMDB can't, then routes every action it proposes through an auditable, four-tier control system before anything actually happens.
The gap between what discovery tools see and what the CMDB says creates a recurring pile of manual work. Provenance turns that leftover pile into a governed, evidence-backed stream of decisions instead.
The model can recommend an action. Whether it's actually allowed to happen is decided by a fixed, deterministic policy — never by the model itself.
Reversible, no-cost fixes — like normalizing a software name or correcting a CMDB field — go through immediately, but only after the evidence and reasoning behind them are recorded.
The model is a triage step for the leftover ambiguous cases — it doesn't replace the core matching logic. That's what makes this work at enterprise scale without turning every single record into an expensive model call.
Candidate pairs
Records compared across sourcesBlocked set
Only the plausible candidates move onResolved by exact key
Serial · UUID · MAC — no model call neededLeft for Gemini
Reasoned through, with the reasoning keptSent to a human
The genuinely contested casesIllustrative volume for a mid-market estate. Deterministic matching still does most of the work.
Each agent is scoped to its own job. All eight are built: identity resolution, software normalization, custody checks, license position, lifecycle management, seat reclaim, renewals, and evidence assembly — all governed by the same four-tier risk broker.
Figures out which hardware records across sources are actually the same device.
Standardizes the long tail of messy software titles.
Checks disposal records and keeps the ownership chain intact.
Keeps effective license position up to date, continuously.
Handles joiner, mover, and leaver decisions.
Spots unused resources and reclaims idle licenses.
Tracks contract and vendor renewal cycles.
Assembles audit-ready evidence — it doesn't submit anything on its own.
A scoped backend stops an agent from reading or writing anything outside its own domain.
The broker isn't optional — it's the only path an agent has to reach a write operation, built into the structure itself.
Untrusted documents get sanitized before they ever reach a prompt, while trusted structured APIs skip that overhead entirely.
Firestore's own security rules and the repository interface both refuse updates and deletions, and every proposal carries an OpenTelemetry trace.
See how a governed agent fleet turns conflicting records into verifiable actions — without asking your team to just trust a black box.
Built with Gemini and Google Cloud for the All Things Agentic Hackathon.