How to Use Claygent for Prospect and Account Research

Most enrichment answers a question with a fixed answer. What is this person's email, how many employees does this company have, where are they headquartered. Those are lookups, and a database can serve them.

Research is different. Did this company just open a second office, does their careers page suggest they are building a sales team, is their pricing page hiding an enterprise tier. Nobody sells that as a field, because the answer only exists once somebody goes and looks.

Claygent is Clay's answer to that gap. It is the part of the product that goes and looks. This guide covers how to add one, how to write a prompt it can actually follow, how to test without spending, and what it is genuinely not good at, which matters more than the marketing on either side of the argument suggests.

What is Claygent?

Clay describes Claygents as agents added to a Clay table that take inputs, follow your instructions on a task, and write a structured output into the column you choose. The important word there is structured. A Claygent is not a chat window, it is a column that has to return the same shape of answer for every row.

Clay's documentation names four main things people build them for: lead scoring, where an agent evaluates a prospect against your ideal customer profile and outputs a score with rationale, outbound copywriting, account research covering company activity and recent news and hiring signals, and persona classification that sorts contacts by tier or buying role.

Those four have something in common. Each one is a judgement call rather than a fact retrieval, which is precisely the boundary between where a Claygent helps and where you should have used a normal enrichment for less money.

How do you add a Claygent to a table?

It goes in through the same enrichment panel as everything else. The flow below follows Clay's documentation and independent walkthroughs published this year. Labels move, so treat the shape as the durable part.

1. Open the enrichment panel — In the table holding the accounts or people you want researched, click Add enrichment in the top right corner.
2. Find Claygent — Use the search bar in the enrichment menu to find Claygent and select it.
3. Write the prompt — In the setup window, describe the research task clearly, referencing the input columns the agent should use. Independent walkthroughs describe a Generate step that turns your description into a configured agent.
4. Save, then run a small batch — Click Save, then choose the option to save and run the rows in the current view. Filter the view down to a handful of rows first, so the first real run is a sample rather than the whole table.
5. Open a row and read the full answer — The column shows a short result, and clicking into a cell shows the full detail. Read three or four in full before you trust the column, because a truncated answer can look right while the reasoning behind it was wrong.

Clay has also built a separate home for these agents. Their documentation describes Claygent Builder, reached by clicking Agents in the left navigation, as a hub for creating, testing, and deploying agents from one place. You can build from scratch, describe what you want in plain language, or start from a pre-built template for prospecting, account scoring, contact scoring, or copywriting.

The two entry points connect. Clay's docs note that a prompt you wrote inside a table's AI column can be saved as a reusable agent, and that agents built in the builder can be pushed into a table, with variables mapping to columns automatically. If you have written the same research prompt three times, that is the signal to promote it.

How do you write a prompt Claygent can follow?

Say what to find, say where to look, and say exactly what shape the answer should come back in. Independent guides consistently land on those three, and the third one is the one people skip.

A column is a database field, not a conversation. If you ask for a summary you will get a paragraph, and paragraphs cannot be filtered, scored, or sorted. If you ask for a specific set of named fields you get something the rest of your table can actually use, which is why practitioners recommend asking for structured output such as JSON when the result feeds another step.

• Name the inputs explicitly. Reference the columns the agent should use rather than assuming it will infer them. A domain and a company name is a different research task from a domain alone.
• Point it at a source when you know one. Careers page, pricing page, a specific site. Independent guides recommend specifying where to look rather than leaving it to a general search, both for accuracy and for speed.
• Define the output fields by name. A score, a short evidence string, a source URL. Naming the fields is what turns a research column into something you can filter a campaign on.
• Ask for the source. Requesting a URL alongside every finding is the single cheapest quality control you can add. It gives you something to spot-check and it makes a fabricated answer visible.
• Tell it what to do when it finds nothing. Without an instruction to return an explicit empty value, a language model will often produce a plausible answer rather than admit the page had no information.

Clay also gives the agent context beyond your prompt. Their documentation describes a business context section where company descriptions, ideal customer profile, and buyer personas populate from workspace settings, plus the ability to attach tone guides, messaging documents, PDFs, or CSV files directly to an agent. A copywriting agent with your messaging document attached writes in your voice rather than a generic one.

Test before you run, because testing is free

This is the most valuable thing in Clay's Claygent documentation and it is easy to miss. They state that you can test up to ten inputs at a time for free, and that test runs do not cost credits. Prompt iteration is where a lot of Clay budgets go, and this removes that cost entirely if you iterate in the right place.

Clay's docs describe importing test data from existing table rows, generating test inputs with AI, saving reusable test suites, and comparing outputs across versions and models. That last one matters: it lets you check whether a cheaper model gives you an answer that is good enough for this particular question rather than defaulting to the most capable one everywhere.

What Claygent is not good at

Being clear about this is not a criticism of the product, it is what stops you designing a campaign around an answer that was never reliable.

• Pages it cannot reach. Independent guides note that paywalled content and heavily script-driven sites are where web research agents struggle. If the fact you need lives behind a login, no prompt will fix that.
• Questions with no evidence trail. Asking whether a company is likely to buy is asking for a guess dressed as research. Ask instead for the observable signals you would use to make that call yourself, then score them.
• Confident wrong answers. Every practitioner guide says the same thing: ask for sources and check them. A model that cannot find the answer will frequently produce a reasonable-sounding one instead of stopping.
• Anything a normal enrichment already sells. Using an agent for a field a data provider has is slower, less consistent, and usually more expensive.

Keeping agents consistent across a team

Clay's documentation describes skills, which are reusable methodologies stored at the workspace level rather than on an individual agent. They carry a name, a description, and a set of instructions, and agents load the relevant ones during execution. This is how a team encodes one way of scoring an account instead of five people writing five slightly different prompts.

There is one versioning detail worth reading carefully. Clay's docs note that enabling or disabling a skill creates a new version, and that updating a skill does not retroactively change agents that were already deployed. So improving your scoring method does not automatically improve the agent running in production. You have to redeploy it, and someone has to remember.

Agent you configure, or employee you brief

Claygent is a strong piece of engineering and Clay has clearly thought hard about it. Free testing, version comparison, workspace-level methods, attached context documents: these are the things you build when real teams are depending on the output. If you want an AI research step you can inspect, version, and control precisely, it is a good place to put it.

Sistava works the other way round. You hire an AI Employee for a role, describe what you want to know in a sentence or two, and it decides how to find out. Neither is universally right. If your value comes from owning a research method nobody else has, build the agent. If you just want the research done, brief the employee and get on with the day.

Sources

• Clay documentation, Claygent Builder
• Clay, Claygent product page
• Clay documentation, Enrichments
• Clay documentation, Ways to save Clay credits
• Arcade, How to Use Claygent AI Research in Clay
• ColdIQ, How to Use Clay's AI Agent (Claygent) Feature