Clay's AI agent has a name, and it is Claygent. It is a research agent that runs inside a Clay table, takes a prompt per row, goes and reads the web, and writes back a data point you could not buy from a standard enrichment provider. In 2026 Clay added a second, broader kind: Account Agents, still in beta and limited to paid tiers.
Clay occupies an unusual position. It is a go-to-market data tool with a research agent inside it, which means people arrive looking for "the Clay AI agent" with very different expectations. Below is what each agent does, how the two-meter pricing actually behaves, and where Clay stops.
What Clay actually ships under "AI agent"
Clay's agents are research and enrichment agents. They exist to fill a column you could not otherwise fill, which is a narrower scope than the word agent usually implies.
- Claygent: the row-level research agent. You write a prompt, it researches the web for each record, and it returns a custom data point. Because the prompt is yours, it can produce data points that standard enrichment providers do not offer as fields, which is the main reason teams reach for it.
- Account Agents: in beta, working at account rather than row level, and available only on the Launch, Growth, and Enterprise plans. Clay charges one action per record processed by the agent, and exports to a CRM or data warehouse carry no additional charge.
- What that tier restriction costs: Clay lists Launch from $185 per month and Growth from $495 per month, with Enterprise on custom annual pricing. Account Agents therefore start at the $185 tier, while Claygent is available from Free.
Claygent behaves like a column rather than a worker: Clay describes it as its AI agent with access to web research, used to generate custom data points, and it is billed per prompt. Two consequences follow directly from Clay's published billing rule, which charges one action per AI prompt plus variable data credits:
- An action is charged per prompt, not per useful answer. Clay's stated unit is the prompt, so rows where the research turns up nothing are still prompts. Budget on row count, not on the number of rows you expect to enrich successfully.
- The variable half tracks the model, not the row. Clay charges data credits on the actual complexity and token usage of each request, so the same prompt over the same rows costs different amounts depending on the model selected for that column.
We have not run a large Claygent job ourselves, so treat anything you read about its hit rate, including here, as unverified until you test it on 200 rows of your own data. The free plan exists for exactly that.
The important framing: Claygent is a column in a table, not an autonomous worker. It runs when the table runs. That constraint is what makes it reliable, and it is also the thing people miss when they compare it to a scheduled agent platform.
How Clay's agent pricing actually works
Clay bills agents on two meters at the same time, and understanding the interaction is the difference between a predictable bill and a surprising one.
| Meter | What triggers it | How it varies |
|---|---|---|
| Data credits | The research and generation work itself | Variable pricing for advanced reasoning models; fixed pricing for Clay's own models and standard content generation |
| Actions | One action per AI prompt | For Account Agents, one action per record processed by the agent |
| Exports | Sending results to a CRM or data warehouse | No additional charge |
The consequence is that your model choice is a pricing decision. Running an advanced reasoning model across a 5,000-row table is a fundamentally different bill from running Clay's own model across the same table, even though the action count is identical. That is the mechanism behind a surprise Clay bill: because cost tracks model complexity rather than row count alone, choosing the strongest model as a global default multiplies the variable component across every row, while choosing it per column confines it to the columns that need reasoning.
The free plan is genuinely usable for evaluation: 500 actions and 100 data credits per month, unlimited seats and tables, multi-provider waterfalls, up to 200 rows per table, and the option to bring your own API key. That last option matters, because it changes the cost equation for anyone with existing model spend.
What we found researching this post
The pricing page and the Claygent FAQ disagree in emphasis, and the FAQ is the more useful document. The pricing page sells plans by action allowance, which invites you to model cost as actions multiplied by rows. The FAQ then explains that data credits are the variable component and that they move with model complexity and token usage. If you budget from the pricing page alone, you will get the shape of the bill wrong.
We also want to flag a limit on what we could verify. Clay publishes a model pricing reference in its documentation rather than on the pricing page, so the specific per-model rates are a live document we have deliberately not reproduced here. Copying a rate table into a blog post is how stale pricing spreads, and rates in this category change quietly. Check Clay's own model pricing reference before you build a cost model.
The third observation is about the beta boundary. Account Agents being restricted to Launch, Growth, and Enterprise means the agent most people read about in 2026 coverage is not the one they get on the free plan. If you sign up to test "Clay's AI agent" and land on Free, you are testing Claygent only.
The gap: work Claygent cannot do
Claygent is a research agent bound to a table. That is a precise scope, and the things outside it are predictable:
- Recurring scheduled work. Claygent runs when the table runs. It is not a Monday-morning agent that checks something and reports back on its own.
- Acting in other systems. It researches and writes data. Updating a project tracker, sending a follow-up, or chasing an internal owner is outside its job.
- Work that is not row-shaped. If the output is not a value in a column against a record, Clay is an awkward container for it.
- Cross-tool reconciliation. Answering a question that spans your billing system and your CRM is not a research task and not what Claygent is for.
None of this is a flaw. Clay is a data platform with a research agent inside it, and the agent is scoped to the platform's job. The mistake is buying Clay expecting general automation and discovering it is very good at exactly one thing.
How to decide, by situation
- You need custom data points nobody sells. This is Claygent's home ground, and the free plan is enough to prove a prompt works before you commit to a tier.
- You are enriching lists at volume. Model your cost per column, not per table, and set cheaper models on columns that do not need reasoning.
- You want the account-level agent. Budget for at least the Launch plan, since Account Agents are not on Free, and treat beta as beta.
- You already have model spend. Bring your own API key; it changes the economics enough to be worth the setup.
- You want recurring work done across your tools. This is not Clay's shape. A scheduled agent platform is the right category, and record enrichment in Airtable shows what the adjacent pattern looks like.
Where Gravity fits next to Clay
Clay answers "what is true about this record." Gravity answers "do this work every week." Those are different products and they sit next to each other cleanly in a go-to-market stack.
A realistic pairing: Claygent researches and enriches the account list inside Clay, and a Gravity agent then runs the recurring work that list creates, checking which enriched accounts actually entered the CRM, flagging the ones that stalled without an owner, and posting a short Monday summary to the team channel. Clay does not schedule, and Gravity does not do row-level web research; each is doing what it is shaped for.
On pricing, the models are opposites. Clay meters two dimensions that both move with your model choice, which rewards careful design and punishes defaults. Gravity is a subscription: a free tier with one agent at $0 per month, and paid plans from $20 per month including $20 of usage, with extra usage purchasable beyond your plan. For a broader view of how these models compare, see AI agent pricing explained and the cheapest AI agent platforms.
For go-to-market workflows specifically, see the best AI agents for sales teams and Salesforce data hygiene, which is the work enriched records usually create next. Other vendor explainers in this series: Workday AI agents and Ramp's finance agents.
Frequently asked questions
What is Clay's AI agent called?
Claygent. It is Clay's AI research agent with web access, and it generates custom data points from a prompt you write, running per row inside a Clay table.
Is Claygent available on the free plan?
Yes. Claygent in tables is available on all plan tiers including Free, which includes 500 actions and 100 data credits per month with up to 200 rows per table. Account Agents are not on Free.
How much does Claygent cost?
Clay charges on two meters: data credits, which are variable for advanced reasoning models and fixed for Clay's own models, and actions, at one action per AI prompt. Because cost tracks the model and token usage, the same prompt can cost different amounts.
What are Clay Account Agents?
An agent type in beta that works at account level rather than row level. They are available on the Launch, Growth, and Enterprise plans, charge one action per record processed, and add no extra charge for exports to a CRM or data warehouse.
Can Claygent run on a schedule?
Claygent runs as part of a table rather than as an independent scheduled worker. Recurring work that checks something and reports back on its own is a different category of tool.
Can I use my own model API key with Clay?
Yes. Bringing your own API key is listed as available from the Free plan upward, which changes the cost equation for teams that already hold model spend.
Sources
- Claygent product page and agent pricing FAQ, Clay, checked 7 September 2026.
- Clay pricing, plan names, published Launch and Growth prices, and the allowances for actions and data credits, checked 7 September 2026.
- Clay model pricing reference, Clay University documentation, checked 7 September 2026. Per-model rates change, so read this before building a cost model.