Flux Intelligence · Agentic, and bounded

Ask your estate a question. Get a defensible answer.

An AI agent that plans its own investigation — choosing which governed tools to call, and in what order — over your real cloud data, without ever touching that data directly. It cites the tools it used, separates fact from inference, and opens with what it could not cover.

Ask Flux

Two ways in, one governed brain.

Ask a question and the agent decides what to retrieve: it calls a governed tool, reads the result, and calls another until it can answer or hits its ceiling. Ask Flux is a contextual panel that follows you across the product and already knows the page and filters you are looking at. The Intelligence Workspace is the full-screen surface for longer investigations. Both require the same application role you already hold, and both answer from the same published snapshot the dashboards use.

agentic retrievalgoverned charts mermaidevidence links follow-up questions
ASK FLUX · READ-ONLY
YOU
Which VMs are the best right-sizing candidates, and what evidence supports it?
FLUX INTELLIGENCE
get_fleet_telemetryget_rightsizing_recommendations
Across the covered fleet, the strongest candidates are sustained-low-CPU VMs with adequate observation windows. Each carries its p95 utilization, coverage percentage, telemetry source, and current actual monthly cost.
Partial coverage: 12 VMs have no telemetry and were excluded rather than assumed idle.
The governance boundary

The model never gets a database connection.

Most "AI for cloud cost" generates SQL against your warehouse. Flux does not. The assistant can only invoke a fixed catalog of bounded, authenticated, read-only tools — the same ones that power the product's own reports.

QUESTION ASSISTANT APPROVED TOOLS ONLY GOVERNED DATA Authenticated user Entra identity role enforced per call Flux Intelligence agentic tool loop call ceiling enforced spend ceiling enforced response contract validated & scored Governed report catalog approved measures & dimensions Cost & FOCUS evidence daily history · charge level Fleet telemetry & right-sizing many resources per call Inventory · opportunities · policy filtered, bounded reads Refused by design raw SQL · credentials · direct cloud access · writes Published snapshot same version the dashboards serve
get_cost_summarygoverned measures
investigate_cost_changehistory + FOCUS
get_fleet_telemetrymany resources
get_rightsizing_recommendationsdeterministic
get_cost_anomalieswith contributors
search_inventoryfiltered
execute_sqlnot approved
any credential accessnot approved
Agentic, deliberately halfway

Autonomous about investigating. Powerless to act.

Most agentic cloud tooling sells autonomy at the point of change — an agent that resizes, deletes, or purchases on your behalf. Flux draws the line somewhere more useful: the agent is fully autonomous in deciding what to look at, and has no capability whatsoever to change anything.

It can chain a dozen governed reads to chase a cost spike across subscriptions, services, and charge-level evidence. It cannot resize a VM, buy a reservation, alter a tag, or write a byte — not because policy discourages it, but because no such tool exists for it to call.

plans its own retrieval chains governed reads no write tools exist no cloud credentials

The AI narrates. It does not calculate.

Right-sizing, anomaly baselines, allocation, and valuation are deterministic engines — statistics and rules, versioned and tested. The agent retrieves those results and explains them. It is never the thing computing the number you act on.

Fleet-scale in a few calls

Governed tools return many resources per call rather than one, so an estate-wide question resolves in a handful of steps instead of exhausting the loop one VM at a time.

Answer discipline

It tells you what it doesn't know.

An assistant that always sounds confident is worse than useless in finance. Flux Intelligence is constrained to disclose rather than smooth over.

Partial coverage is stated first

If a source is throttled, stale, or incomplete, the answer opens by naming the gap and quantifying it. A partial total is never described as estate-wide.

Facts separated from interpretation

Every response splits what was retrieved from what was inferred, lists the tools it used, and carries its own limitations — as structure, not prose you have to parse.

It refuses to do the math it shouldn't

Where Flux has no deterministic engine — commitment purchase quantities, for instance — the assistant is required to say so instead of inventing a number that looks authoritative.

Deterministic quality scoring

Each reply is scored on structure, grounding, partial-coverage disclosure, tables, and completeness — so answer quality is a tracked metric, not a vibe.

Reviewable, then forgotten

Prompts and validated replies are retained for a configurable 30 days for administrator quality review. Model reasoning is never retained.

Answers link back to evidence

Server-generated links take you from a claim to the Flux page that proves it — inventory, cost report, opportunities — so nothing has to be taken on faith.

Administrative control

Bounded on purpose, and visibly so.

Flux Intelligence runs behind a swappable model-service adapter with fast and deep-analysis profiles. Spend is capped with a hard stop below the ceiling. Tool calls are capped per conversation. Every request carries stage timing so slow paths are measurable rather than anecdotal.

spend ceilingtool-call ceiling role requiredtranscript review stage timingswappable provider
intelligence · status
enabled true
authorization Flux.Reader required
data boundary governed Flux APIs only
retention 30 days · reasoning excluded
budget stops before the ceiling
limitations surfaced with every answer

// no write operations are available

Where it sits in the stack

Flux Intelligence is the conversational surface of a larger capability. Flux Signals is the deterministic rule engine that produces optimization findings, and the governed report catalog is the bounded API both experiences read. The assistant narrates and navigates; it never computes the numbers it reports.

Ask it something hard.

The interesting test is not what it answers — it's what it admits it cannot.