Appearance

Concepts ​

Flowstate runs a closed loop: capture what happens, attribute it, raise signals, enforce the rules. This page gives you the vocabulary and shows how the pieces connect.

Two lenses ​

Flowstate views your AI estate through two lenses.

The human team. Your employees and contractors use copilots: Claude Code, Cursor, GitHub Copilot, Gemini and the rest. Copilot spend loads onto the person who used it, as part of their fully loaded cost. A copilot is a tool, never a worker.

The agentic team. Agents are autonomous AI workers you plan and resource, the way you plan headcount. Each Agent has an accountable owner, a model from the catalogue, and a forecast built from expected runs. See Agents.

The same product can appear in both lenses. A Claude Code session owned by an engineer is copilot usage. A Claude Code session owned by an AI service account is agent work. Attribution decides which, and nothing is counted twice. The full cost model is at The AI cost model.

What "AI" means here ​

Not just coding tools. Flowstate tracks image generation, video generation, audio and voice, sales and marketing copilots, legal research, enterprise search, and the long tail. Coding tools have the deepest coverage today, because that is where the agent and the provider billing APIs give the most signal. See Catalog for what's visible right now.

Core vocabulary ​

  • Agent — a planned, resourced AI worker: the agentic-team analogue of a headcount line. It has a name, a purpose, an accountable owner, and a model reference from the Models catalogue. You assign it to projects and forecast it by runs (expected runs × per-run cost), the way you forecast a person by salary and allocation. Managed on the Agents tab at Resourcing → Agentic workforce. See Agents.
  • Copilot — an AI tool a person uses. Its spend loads onto that person. A copilot never gets its own worker line.
  • AI service account — the production runtime identity: a deployed AI workload (customer chatbot, support agent, internal automation, autonomous coding agent) that calls providers with its own credentials and reports usage through the Flowstate SDK. It is a Subject in its own right, with its own sessions, spend and behavioural baseline. Linking an Agent to its AI service account is the deployment link: the account's actual sessions and spend reconcile against the Agent's forecast runs, plan versus actual. Managed at Settings → AI → AI Service Accounts. See Service accounts.
  • Models — the Flowstate-maintained model catalogue: name, family, vendor, modality, tier, context window, price. An Agent stores a reference to a model, never a copy of its price, so a catalogue price change applies everywhere at once. The Models tab at Resourcing → Agentic workforce. See Models.
  • AI tools — the umbrella tab over the equipment layer (Resourcing → Agentic workforce → AI tools). It holds two kinds of entry: Detected services and Usage providers.
  • Detected services — the AI services Flowstate has observed traffic for in your org, captured by the on-device agent and the Cloud Proxy. Read-only discovery. See Detected services.
  • Usage providers — admin-configured provider billing connections (API key, OAuth or CSV) that pull authoritative usage and spend. Connected from the Integrations marketplace at Settings → Integrations. See Usage providers.
  • Session — an idle-bounded conversation thread by one Subject in one AI service. A Turn is one logical request (prompt and response) inside a Session.
  • Signal — one detection event raised against a Turn or Session: a DLP hit, an abuse pattern, an off-topic production prompt, a brand violation. Severity is INFO, WARN or BLOCK. Signals surface at Alerts and signals.
  • Rules — the policy rules that emit Signals. Admins manage them on the Rules tab at Resourcing → Agentic workforce. See Policy and enforcement.
  • Subject — the actor an AI interaction is attributed to: a user, an employee, a contractor, or an AI service account.
  • AI caps — AI budget caps for the org and each team, managed at Finance → Budgets → AI caps. See Budgets and spend caps.
  • Cloud Proxy key — the org-wide credential the on-device agent presents. One key per org. See Cloud Proxy key.

The loop ​

capture  →  attribute  →  reconcile  →  signal  →  enforce
  1. Capture. The agent and the Cloud Proxy observe traffic from people's machines. AI service accounts report their own usage through the SDK. Usage providers pull authoritative billing.
  2. Attribute. Every session resolves to a Subject and, where possible, a project.
  3. Reconcile. Observed usage is matched against provider invoices, and agent actuals against the planned Agent's forecast.
  4. Signal. Rules evaluate sessions and raise Signals for the things you'd want to know about.
  5. Enforce. Rules can observe, warn, or block at the request line.

Telemetry ​

The Flowstate Agent sits on developer machines and observes traffic to AI providers. It captures one of two payloads, depending on the pathway you've chosen:

  • Telemetry — request metadata only: host, model, token counts, status, duration. No prompt or response body.
  • Enterprise — full request and response, including prompts and tool calls. For orgs on Claude for Enterprise, ChatGPT Enterprise, and similar contracts.

Attribution rides the org-wide Cloud Proxy key plus the user identity carried on each request. There are no per-employee credentials to issue or rotate. Production workloads are different: each AI service account has its own key and reports through the SDK.

The Cloud Proxy routes managed devices through proxy.flowstate.inc and attributes every request to its subject. It does not by itself provide prompt-body capture, prompt-quality scoring, or request-time enforcement.

Attribution ​

A captured session is useless without a name on it. Flowstate resolves each session to:

  • A Subject — a user, employee, contractor, or AI service account. A session owned by a service account carries no person: it is agent work.
  • A project — inferred from working context (open IDE project, git branch, recent commits) or marked manually by the engineer in My AI.
  • A team, cost centre, value stream — derived from the subject's allocations in the workforce model.

Attribution rolls up live in Agent insights. Drill from org to cost centre to team to person to individual session.

Reconciliation ​

Observed usage and provider invoices rarely match exactly. The Spend reconciliation screen works each provider-billing row through five buckets: unreconciled, suggested, confirmed, written off, deferred. You work the unreconciled queue the way an accountant works Xero.

Reconciliation covers both lenses. Person-owned copilot spend reconciles against seat and API billing. Agent-owned service-account spend reconciles the same way, and also against the planned Agent's forecast runs, so you can see plan versus actual per Agent. See Forecast.

Signals and rules ​

Rules evaluate captured sessions and raise Signals in four categories:

  • Content safety — PII detected, secret leaked, API key in prompt, customer-data pattern, banned content.
  • Brand and usage — brand-policy violation, off-topic in production, personal use of a company account, abuse pattern, wasteful prompting.
  • Budget — forecast breach against an AI cap.
  • Operational — monitoring gap, model regression, auth-context mismatch, key misuse.

Each Signal carries a severity: INFO, WARN, BLOCK. Triage happens at Alerts and signals (Insights → Agent insights → Alerts in the app). Rules themselves live on the Rules tab and are versioned and audited; see Policy and enforcement.

In observe mode a rule raises signals without changing behaviour. In enforcement mode the agent intercepts at request time and returns a 403 to the calling tool with the rule that fired. See Enforcement for the rollout sequence.

Fluency ​

Analysed sessions are scored on the 4D framework: Description (clear prompts), Delegation (the right task handed to AI), Discernment (checking the output), and Diligence (verifying before shipping). Each person sees their own habit scores and verdict on My AI. See Fluency.

Where the agent ends and the platform starts ​

The agent captures and (optionally) enforces. Everything else runs in Flowstate: attribution, reconciliation, fluency scoring, signal generation, policy storage, audit, the dashboards. That separation matters. The agent is small, deterministic, and fail-open by design. If Flowstate is unreachable, the agent buffers locally and lets traffic through.

Next ​

Flowstate Documentation

Copyright 2026 Flowstate.