How to Use Clay to Research and Score Target Accounts

The short answer: for a marketing lead with 200 to 500 target accounts and no analyst, Clay's Launch plan at $185 a month on monthly billing (pricing page, checked on 26 October 2026) covers the job: import the accounts without running anything, test one AI research question on ten rows in sandbox mode, score the answers with a formula, find two buyers at each qualified account through the Work Email waterfall, and send them to lemlist or a CSV.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

Most target account lists are a column of company names that someone agreed on in a planning meeting. Nobody has checked which of them sell to the customers you serve, are hiring for the problem you solve, or run the tool you replace. Clay turns that check into a column. For plans and how credits work, see the Clay glossary entry.

Before you build:

1. Price the run. For example, AI research on 300 accounts uses 300 actions, one per row. If 120 qualify and you find 2 people at each, Clay's own worked example prices an email find at 0.5 data credits, so 240 x 0.5 = 120 data credits if every search succeeds, per the credits guide. The AI model's data credits come on top and vary by model.
2. Start from the CRM. Export the account list with a company domain column, plus closed-lost and current customers to exclude. If you already pay for a data provider or an AI model, have the API key ready; your own key skips data credits.
3. Agree the measure with sales first. Ask how many accounts from your last list the sales team accepted as worth working. That share is the baseline the scored list has to beat.

Step 1: Import the accounts without running anything

From your workspace home, click + Create new, search for CSV and upload the file, then match your columns to the table fields. Choose Save and don't run, per the CSV import guide, so nothing spends credits before the table is ready. Then switch Auto-run off in the table toolbar while you build, which Clay's credit-saving guide recommends.

Step 2: Write one research question

Click Add column, then Use AI. In the Generate tab, describe the question in plain English: "Visit the company's website and say whether it sells to B2B customers in Europe. Return yes or no, a one-line reason and the URL of the page you used." Clay builds the prompt, suggests a model and sets the output fields, per the Use AI guide. Check the Configure tab: the use case should be Web research (Claygent), and the outputs should be separate fields, so the formula in Step 4 can read them.

Step 3: Test it in sandbox, and read every answer

Click Sandbox Mode in the toolbar. Clay copies the top 10 rows into a test table of at most 50 rows, keeps your main table read-only, and disables outbound actions such as exports, per the sandbox guide. Run the AI column on these rows.

This is the human check, and it does not move to later. Open the evidence URL for each of the ten answers and decide whether you agree. Where the agent guessed, tighten the prompt, for example by naming which page to read first, and run again. When nine in ten answers hold up, publish with Publish and run; rows already run in sandbox are copied back with their results, so you do not pay for them twice.

Step 4: Score with a formula, then gate the spending

Add a formula column that turns the answers into a score: the B2B answer, the employee count from a company enrichment and any signal you care about. Formulas use no actions, per the credits guide. Then give the next, paid columns a run condition, such as score above your threshold, so they only run on accounts that qualify, as the conditional runs guide describes.

Step 5: Find the buyers at qualified accounts

Add a Find People source with the Companies filter pointed at this table, a job title filter on "is similar to" for the roles you sell to, and Limit per company set to 2, per the Find People guide. Sourcing the list uses no actions; enriching it does.

Step 6: Run the Work Email waterfall carefully

Add the Work Email enrichment and turn on Infer Email as the first step, a free guess from a naming pattern that stops the waterfall if it validates. In Full configuration, choose the Conservative validation strategy, which the waterfall guide recommends when bounce rates affect sender reputation. You pay only for the provider that finds a match. Treat catch-all, unknown and role-based results, as defined in the verification statuses guide, as not ready to send.

Step 7: Send the qualified buyers on

For lemlist, add the Lemlist enrichment, connect your API key, and use Add Lead to Campaign with the campaign ID of a campaign that already exists in lemlist, per the lemlist integration guide. Give it a run condition: only rows with a valid email. Each record exported to an outside sequencer uses 1 action. CRM sync to HubSpot or Salesforce needs the Growth plan; on Launch, export a CSV for the CRM import.

How to know it worked

Two numbers. The share of accounts sales accepts from the scored list, against the baseline you agreed. And the cost per accepted account: data credits from Settings and then Usage, divided by the accounts sales took. Clay earns on every row that runs, which is why its own advice to turn Auto-run off while you build, and to test on about ten rows first, deserves to be taken at face value.

What Clay's guides warn about

• Running a new prompt on everything. The credit-saving guide says to test about ten rows first, because prompts take several attempts.
• Leaving Auto-run on while building. Rows added by mistake get enriched and charged.
• Loose validation for cold email. Aggressive validation accepts more risk; keep Conservative when sender reputation matters.
• Expecting exports in sandbox. Outbound actions are switched off there; test an export on the main table with a filtered view.

Later, if one table grows into agents across inbound, outbound and the CRM, the job has become designing an agentic system rather than using a tool; the agentic marketing work I sell covers that.

FAQ

Does the Claygent column cost the same on every row?

Not always. Most models charge a flat number of data credits per task, while token-heavy models are billed on the tokens used and show an estimate with a tilde, per the pricing FAQ.

Can I use my own OpenAI or Anthropic key?

Yes. Each run still uses 1 action and no data credits, and the provider bills the tokens, per the credits guide. Clay shows only the actions, so check spend in the provider's console.

How many accounts can Find Companies return?

Up to 10,000 per search, per the Find Companies guide, with exclusions of up to 300,000 companies across three sources.