AI Catalog

Flowstate sees AI spend across many modalities. Here's the lay of the land — and exactly what we're tracking versus what we're not. This catalog is the reference view of the broader AI market, organised by modality. For the operational view of what's connected and live in your org, see Detected services, in the app under Resourcing → Agentic workforce → AI tools.

The catalog is also what budgets refer to when a budget says providers: "all" — every entry below that we can actually see.

The four coverage states

Every tool on every modality page sits in one (or more) of four states.

Catalog-listed. The tool is in Flowstate's built-in AI tool catalogue, so Flowstate can classify it, price it, and attribute its spend. A tool a person uses is a copilot: its spend loads onto that person's fully-loaded cost, not onto a worker line of its own. Default pricing ships with the entry; you can override it.

Hostname-recognised. The Flowstate Agent passively captures traffic to one of the tool's known hostnames. No admin setup needed beyond installing the agent. Today this covers five vendors: Anthropic, OpenAI, Google, GitHub Copilot, Cursor. Codeium/Windsurf is also on the proxy allow-list.

Billing-connectable. An admin can wire the vendor's billing API from the Integrations marketplace (Settings → Integrations) for authoritative spend; see Usage providers. Today: the Anthropic Console API, Claude Enterprise, the OpenAI Platform API, GitHub Copilot, Cursor, Windsurf, Gemini API, Vertex AI, Azure OpenAI, Amazon Bedrock, Mistral, Fireworks AI, ElevenLabs and Intercom Fin. Every connectable tile has a working pull behind it — there are no placeholder entries.

Invisible. Flowstate cannot see this tool today through agent traffic or a billing connector. It's still in the catalog for reference, but there is no per-user attribution. Track its cost as contract SaaS spend on the finance side. If the workload is an autonomous production system rather than a tool a person uses, model it as an Agent.

A single tool can hold multiple states. Cursor, for example, is Catalog-listed AND Hostname-recognised AND Billing-connectable.

Copilot, Agent, or Model?

Three distinct things, never interchangeable:

• A copilot is an AI tool a person uses (Cursor, Claude Code, ChatGPT). Its spend loads onto that person's fully-loaded cost. It never gets its own worker line.
• An Agent is an autonomous AI worker you plan and resource. Its usage is its own cost line, attributed to no person. See Agents.
• A Model is an entry in the Models catalogue (vendor, tier, price) that an Agent references. See Models.

How spend lands on each is defined in The AI cost model.

Modalities

• Coding — IDE assistants, coding agents, vibe-coding app builders. Our strongest coverage.
• Chat — General-purpose assistants used across every department.
• Image — Generative image tools used in marketing, design, content.
• Video — Generative video and avatar tools.
• Audio — TTS, voice cloning, music generation, transcription.
• Sales — Enrichment, call intelligence, embedded CRM AI.
• Marketing — Content production, agent workflows, research.
• Legal — Legal research, drafting, contract review.
• Enterprise search — Internal RAG, meetings, ITSM, customer service.
• Foundation APIs — Raw model APIs and gateway providers.

Honest coverage today

What Flowstate is good at today: coding tools — full coverage across catalog entries, agent hostnames, and billing connectors for the major vendors.

What we're catching up on: image, video, audio, sales, marketing, legal, enterprise search, foundation APIs. Most are catalog-listed (so they appear in your roster and forecast), but few are hostname-recognised or billing-connectable yet. The Coverage column on each modality page tells you exactly what's tracked today and what isn't.

If a tool you care about is Invisible, the workaround is the same: track its contract value as SaaS spend on the finance side. The cost still rolls into your budgets and forecast, just without the per-user attribution you get from agent telemetry. If the tool is an autonomous production workload rather than something a person uses, model it as an Agent instead; see The AI cost model.