Clay AI and Clay.com, explained

Clay.com (formerly clay.io, and often miscalled a CRM) is a spreadsheet IDE for go-to-market teams. You load a table of companies or people, add columns that call data providers, chain them into waterfall enrichment, and push the result to your CRM or outbound tool. It is not a CRM and not a sending tool: it is the layer that builds and enriches the list before those.

The four jobs teams actually buy it for: building target lists, enriching records across many providers at once, qualifying accounts against a custom ICP, and generating personalized outbound copy from what it found. Everything else in the product exists to serve those four.

It is genuinely powerful and genuinely technical. Most teams running Clay seriously have one person whose job is Clay. That is the part the pricing page does not mention.

The AI part is not one feature: it is three distinct things that share a bill.

First, AI formulas: an LLM call inside a cell that rewrites, classifies, or drafts from data you already have. Second, Claygent: an agentic research column that goes out to the open web, reads pages, and writes a structured answer back into the row. Third, Account Agents, in beta and limited to the Launch, Growth, and Enterprise plans, which run continuously against an account list rather than once per row.

The distinction matters because they fail differently and cost differently. An AI formula is cheap and deterministic-ish. Claygent is where the research happens, where the money goes, and where most of the frustration lives.

Claygent takes a prompt you write, a set of input columns from the row, and an output schema. It searches, opens pages, reads them, and returns fields you defined. One row, one run. It is the closest thing Clay has to a research agent, and it is what most teams are actually buying when they say they bought Clay for AI.

You choose the model per column. Clay ships its own managed models (Helium, Neon, Argon) at a fixed credit price, and lets you swap in frontier models such as Sonnet 5 or GPT 5.6, or open-weight models such as Kimi K2.6 and GLM-5.2, at variable pricing. Same prompt, different model, very different bill and very different answer quality.

Clay.com bills on two meters. Data Credits pay for the data and model work; Actions pay for the operations your table performs. Clay's managed models are priced per row: Helium 1 credit, Neon 2 credits, Argon 3 credits. Since March 2026, every Claygent run consumes an Action on top of the Data Credits, so one research task now drains both balances at once. That change is the single most common source of surprise bills in 2026.

Run the math on the Launch plan at $185 a month: 2,500 Data Credits and 15,000 Actions. Research with Argon at 3 credits per row and the plan tops out around 833 companies a month, roughly $0.22 per researched company, before you count the operator running it. Drop to Neon and you get 1,250. Data Credits, not Actions, are what run out first for anyone doing real research. Growth at $495 a month raises the ceiling without changing the shape of the problem.

Two more line items people forget: failed and empty runs still consume credits, so a badly tuned prompt costs money while it teaches you nothing. And the operator salary belongs in the comparison, because Clay is not a tool you dip into once a month.

Clay exposes webhooks and integrations to move data in and out of tables, and it can be triggered from other tools. What it is not is an API you call to get a researched answer back in the same request. The unit of work in Clay is a table row processed by a workflow you built, not an endpoint that takes a question and returns a verdict.

That distinction decides the architecture. If you want research inside your own product or pipeline, on demand, with a typed response, you want a research API, not a spreadsheet you poll. That is the gap teams hit when they try to make Clay the backend of something.

Claygent reads the searchable web well. It struggles where the truth sits behind interaction: JS-heavy careers pages, filtered directories, gated pricing, portals that need clicks before they show anything. A search-and-read agent gets a homepage summary; it does not drive the site.

It is also row-shaped by design. Clay AI enriches a list you already have. Building the list from nothing (discover candidates, dedupe them, qualify them against a brief) is a different motion that Clay expects you to assemble from other columns and providers.

And the output is only as good as your prompt. Teams routinely spend weeks tuning Claygent prompts per use case, which is real work that never shows up on the pricing page.

Stay on Clay if you live in tables, orchestrate many providers, and want one surface where enrichment, AI, and routing all meet. Nothing else combines that breadth, and the community around it is genuinely strong.

Move to a research API when the research itself is the product: you want a plain-English brief in and a typed, cited verdict out, with a fixed cost per company and no second meter. Gumshoe returns exactly that from one call, drives a real browser when a site fights back, and bills one credit per company at $99 a month for 2,000 cases, roughly a fifth of the per-company cost of Argon research on Launch.

The honest framing: Clay AI is a workbench with an agent bolted in. If you want the workbench, keep it. If you only ever wanted the answer, you are paying an operator and two meters for a tool you use one column of.