Clay Account Agents: What Per-Account AI Research Really Costs

September 8, 2026. Clay's pay-on-success guarantee is one of the best things about the product. It does not cover the AI agents that Clay is now pushing hardest. If you are budgeting a per-account research agent across a real list, that gap is the number that decides whether the pattern is affordable.

What Account Agents changed

On August 25 Clay announced Account Agents inside Workflows, in open beta for Launch, Growth and Enterprise customers. In Clay's description, the agent "remembers what it concluded last time and chooses the next step from a list of configurable actions". Clay's own comparison puts it against the older Claygent, which it describes as stateless and one-shot, where the Account Agent is persistent and tracks what changed.

That is a genuine capability shift. One agent per account, carrying state between runs, is a different thing from one prompt run across a list. It is also billed differently, and that is where operators get caught.

The pay-on-success guarantee stops at AI columns

Clay bills in two currencies. Actions cover platform work and start at less than a cent each. Data Credits cover data and AI bought from vendors and start at $0.05 each. Waterfall enrichment, the feature Clay markets under the line "pay for what you find", searches providers in sequence until it gets a match, and Clay's pricing FAQ is explicit: if an enrichment returns no result, you are not charged Data Credits or Actions.

Now read Clay's credits documentation, which lists AI columns under the reasons credits get deducted unexpectedly. Its wording is that every row processed incurs a cost.

Those two statements are both accurate and they describe different products. The pay-on-success economics belong to waterfall enrichment. They do not transfer to an agent. An Account Agent that researches a company, finds nothing useful and concludes as much has still done work, and you still pay for it. On a 5,000-row list with a 40 percent hit rate, the difference between those two billing models is the whole business case.

What a per-account agent actually costs

Clay publishes a per-model credit table, which makes the cheap end easy to price. Fixed-rate web research runs at 1 credit for Clay Helium, 3 for Clay Argon, 1 for GPT-5 Mini, and 0.5 for GPT-5 Nano. So a Helium run costs roughly $0.05 to $0.06 including the action, and an Argon run roughly $0.15. Every frontier model is listed as variable pricing, so those are not budgetable in advance from the table. Clay also states that 75 percent of runs cost less than its estimate, which is a candid way of saying a quarter cost more.

Compare the do-it-yourself route. Google's published guidance for its Deep Research API estimates roughly $1.00 to $3.00 per task, and $3.00 to $7.00 for the Max tier. That is 20 to 100 times a fixed-rate Claygent run for a single account. For most agencies the platform is not the expensive option, which is the opposite of the assumption people usually start from.

Two further cost details worth knowing before committing: Actions do not roll over month to month, and credit top-ups carry a 30 percent premium over plan rates. Clay's plan pages also show $167 and $446 per month on the plan cards while the FAQ on the same page says $185 and $495. The lower numbers are the annual-billed equivalents. If you are quoting a client on monthly billing, use $185 and $495.

Where it goes wrong, in Clay's own words

Clay's published guide to using Claygent for prospect research is unusually honest, and worth quoting to any client who wants the agent pointed at everything. It warns that a prompt demanding an answer for every row guarantees fabrication on the rows where no answer exists, and that the failure mode of a confident model is a wrong answer that looks exactly like a right one. It also warns against running a full list before testing, which turns a prompt bug into a full-list bill.

The data underneath has limits too. Clay's own guide to firmographic data reports that the best providers land in the mid-80s on accuracy, and that the provider with the widest coverage is usually the least accurate. On one revenue benchmark the most accurate provider was right 88 percent of the time while covering only 42 percent of records. Local service businesses are the hardest case of all, because ownership, headcount and service area change faster than any database refreshes, which is why we lean on live qualification rather than stored firmographics for verticals like home services.

What it means for operators

1. Price the miss, not the hit. Estimate cost as rows multiplied by per-run cost, with no discount for rows that find nothing. That is the real number for an AI column.
2. Test on 25 rows before running 5,000. Clay says this itself, and the reason is billing, not just quality.
3. Pin a fixed-rate model where you can. Helium or GPT-5 Mini at one credit is budgetable. A frontier model on variable pricing is not.
4. Give the agent permission to return nothing. Prompts that require an answer manufacture one.
5. Do not assume the DIY build is cheaper. At $1.00 to $3.00 per deep research task, rolling your own gets expensive faster than the platform does.

Per-account research is genuinely better than list-wide prompting when the accounts are worth researching. It stops being better the moment you point it at a list you have not qualified first, which is the part lead generation has to get right before any agent runs.

Sources: Clay's pricing page, its AI pricing documentation, the Account Agents product page, and Google's Deep Research API pricing guidance.