Clay

Agents take authenticated action in outside systems: Clay buys data from more than 200 providers in one place, syncs CRMs, pushes ad audiences to LinkedIn, Meta and Google, sends through its own sequencer or connected email tools, and offers HTTP API integrations and webhooks for anything unlisted.

Sourceclay.comread 2026-10-01

The core product is a table based workflow builder with enrichment columns, AI columns, conditional logic, routing, and trigger driven automations, positioned as the orchestration layer for GTM.

Sourceclay.comread 2026-10-01

Grounding is assembled per request rather than maintained. The customer controls which table columns are included as context on each AI action, a My Business Context setting carries company positioning and brand voice into message drafting, Claygents research prospects on instruction, and some models accept document upload. Data is sent to a model only when the customer explicitly runs an action.

No maintained retrieval structure over the customer's own corpus is documented: there is no documented ingestion of a document set, no index the agent queries at run time and no retrieval interface, and the natural-language company search runs against Clay's own database of fifty million companies rather than the customer's material.

Sourceuniversity.clay.com/docs/ai-in-clayread 2026-09-01

Control is exercised before an action runs rather than by approving its output.

The documentation states that AI features are opt-in through the customer's own actions, that data leaves a table only when the customer explicitly runs the action, that the customer chooses which columns an action may see, and that admins control who in the organization may use AI features, with role-based permissions, function-level permissions on the MCP surface and per-rep credit budgets. Opt-out and do-not-call handling covers outbound compliance.

None of that is an approval mechanism; nothing lets a buyer mark particular actions as requiring sign-off, name who signs off, or define what happens while approval is pending, and the platform is human-operated by construction rather than by configurable gate.

Sourceuniversity.clay.com/docs/ai-in-clay and clay.com/pricingread 2026-09-01

Role based permissions control who can access data, and the Enterprise plan adds sign in with SSO and role based access control. Clay states an active SOC 2 Type II certification and GDPR compliance on the AI in Clay page. Both SSO and RBAC are limited to the Enterprise plan.

Sourceuniversity.clay.com/docs/ai-in-clayread 2026-10-01

Execution is visible at the row level: each enrichment column shows its own result per record, Claygents process row by row with research results written back into the table, failed actions surface for troubleshooting, and a Debug with AI feature analyses error messages and logs against the documentation to suggest fixes. Credit and usage consumption is tracked in detail. No run-history or audit surface is documented that reconstructs a single agent run end to end, showing the steps taken and the tools called in sequence, as against per-column outputs and error logs.

Sourceuniversity.clay.com/docs/ai-in-clay and clay.com/pricingread 2026-09-01

State persists across runs rather than within one: Claygent research results are stored back into the customer's table under their control, tables are continuously refreshed as accounts are monitored for headcount, technographic, hiring and intent changes, and the documentation states a retention position, that the customer can delete individual records, tables or an entire workspace at any time with workspace deletion removing all data after thirty days. What is documented is a data store with a retention policy rather than a memory layer with a stated scope for the agent: nothing describes the agent accumulating learning across runs or reading back a memory distinct from the table it writes to, and Sculptor's chat context is explicitly transient.

Sourceuniversity.clay.com/docs/ai-in-clayread 2026-09-01

A disclosed single region with no choice: customer data is primarily processed in the United States on AWS US-East, the product is cloud only, and no region selection, self hosting, on premises or VPC option is offered; warehouse sync moves data, it does not relocate the service.

Sourceuniversity.clay.com/docs/ai-in-clayread 2026-10-01

A browsable library exists as its own navigation section: use case templates published on Clay University alongside courses and docs, which a buyer can browse and run. Around it sit named packaged components aimed at named jobs: Claygent as a repeatable research agent with its own builder and playground, Sculptor for table building and debugging, the Sequencer for campaigns, Audiences, and waterfall enrichment.

Sourceuniversity.clay.com use-case-templates and docs indexread 2026-09-01

Trigger driven workflows fire on signals like headcount change, technographics, hiring, and web intent, plus webhooks and scheduled refreshes, and push to email tools, ad audiences, and CRMs.

Sourceclay.comread 2026-10-01

Model choice is a documented product surface twice over. Clay publishes a complete table of available models by provider: its own Helium, Neon, Argon and Clay Navigator, OpenAI's GPT 5.5 and 5.4 families through o3 and o4 Mini, Anthropic's Claude 4.8, 4.7 and 4.6 Opus, Sonnet 5, 4.6 Sonnet and 4.5 Haiku, Gemini 3.1 Pro and the Flash line, and image models from BlackForestLabs, Playground, Segmind and Stability. On top of the provider list the customer can bring their own API keys for supported providers, and doing so unlocks MCP server support inside the Use AI action.

Sourceuniversity.clay.com/docs/ai-in-clayread 2026-09-01

MCP is both served and consumed, alongside a public API.

Clay exposes an MCP server at api.clay.com/v3/mcp that lets an external assistant, IDE, agent host or partner product run Clay's find-and-enrich tools and custom functions on behalf of a signed-in user, with OAuth discovery through standard protected-resource and authorization-server metadata and Dynamic Client Registration as a self-service path, plus documented authorize, token, device-authorization, registration and revocation endpoints.

Admins govern that surface with function-level permissions and per-rep credit budgets. Separately Clay consumes MCP servers inside the Use AI action when the customer brings their own keys. A public API, HTTP API integrations and webhooks are documented, and Clay ships as a connector in ChatGPT, Claude and Codex.

Sourceuniversity.clay.com connect-to-clay-mcp and ai-in-clay docsread 2026-09-01

A change can be evaluated before it reaches customers.

The sequence builder is documented as letting the customer write, personalize and pressure-test the emails a campaign will send before a single one goes out, and a separate A/B testing article documents running two versions of campaign copy against each other and then sending the remaining leads to whichever performs better, which is a scored comparison with a decision rule.

Around that sit failed-action troubleshooting, a Debug with AI feature that analyses error logs against the documentation to suggest fixes, and Sculptor for identifying issues in a table.

Sourceuniversity.clay.com build-your-sequence and ab-test-sequence-copy docsread 2026-09-01

Web scraping is a documented capability with its own topic section in the vendor's documentation, and Claygents are described as AI agents that navigate websites, extract information and make decisions about which sources to consult, running row by row with results stored back into the customer's table.

The Use AI action separately performs web research that searches and synthesizes information from the web. That is scraping and headless retrieval wired into the platform. No documented control of a real interface, whether a hosted or local browser, a desktop session or remote computer control the agent drives.

Sourceuniversity.clay.com web scraping docs and ai-in-clay Claygents sectionread 2026-09-01