Clay AI agent benchmark: what an agent needs to use Clay well

With Clay, an AI agent does its best GTM data work when it has API access, saved Clay functions with known inputs, a rule for free search versus paid enrichment, and permission to leave a cell blank. In our benchmark of seven agents across 105 runs, the agent that searched Clay first and paid only for the person it chose found 82 of 100 CMOs for 285 credits, under 3 per company.

By The Prompting Company · Snapshot November 5, 2026

The checklist

1. Access to Clay through the Clay Agent Plugin or the Public API
2. Saved Clay functions (Clay's API names each one's required inputs)
3. A rule for when to use free search and when to pay for enrichment
4. A check that each company match is the right company
5. Permission to leave a cell blank instead of guessing
6. A check for subsidiaries and headquarters on list-building tasks
7. Enough time for the list size

How should an agent connect to Clay?

The Clay Agent Plugin bundles the clay CLI, the Public API and Clay skills, and installs with one prompt in about 15 minutes on a Mac. Its skills give your agent Clay's methods up front, including the habits that keep credit use down. It works on all plans, including free and trial, at no additional cost. Our sandboxes could not complete the CLI's browser login, so every agent used the Clay Public API with an API key, which was enough for all five tasks.

Why do saved functions and their inputs matter?

Each agent had the same 19 saved Clay routines, and Clay adds no surcharge for running a function. On work emails, six of seven agents found 93 of 100 in every run with the Work Email routine.

Clay's API tells an agent exactly what each routine needs. In the work email runs, the response to a call missing a field named the required inputs, full name, company domain and company name, and agents self-corrected in one call. The list had domains but no company names, and passing the domain as the company name found 93 of 100. Codex with GPT-6 Sol went straight to Work Email in one run and spent 335.5 credits, about 3.4 per contact.

What to give your agent: the routine IDs and, for each, which inputs it can fill from what it already has. The Clay Agent Plugin's skills cover the rest.

When should an agent search, and when should it pay?

Search first when the agent has to choose a record; sourcing a list with Find People uses 0 Data Credits, and the agent pays only for the records it keeps. Pay for enrichment when the record is already known, as with a LinkedIn URL to re-check: the best three agents left at most 1 person undecided.

How does an agent confirm the right company?

Many software companies share a name with an older business, and Clay's company search shows the candidates side by side. Runs that compared the candidates by industry, size and country scored 93 to 100 on domains. More in "How do I stop an AI agent from matching the wrong company in Clay?".

Should an agent leave a cell blank?

Yes. Clay's own guidance for email is direct: "Leave the email blank if it can't be verified, no guessed addresses… A bounced email costs more than an empty cell." The same holds for names. On the CMO task, Opus left a row empty when no one held the title, and found 74 to 82 of 100 CMOs with title and work email.

Does a target list need extra checks?

Yes, for ownership and headquarters. Clay describes each company the way the business presents itself, which lets the agent confirm that each company is listed in its own name. Codex and Opus filtered out subsidiaries and non-US headquarters before adding a company and listed 82 to 94 qualifying companies per run.

How much time should an agent get?

Measured from run start to grading, every completed run took under 49 minutes, and 78 of 102 under 15. Opus's job change runs that used one enrichment per profile finished in under 15 minutes. For a list of 100, a 60-minute limit left plenty of room.

How did the agents rate the experience?

Each agent rated its run from 1 to 5. Every one of the 100 runs that wrote a rating gave its Clay session 3 or 4 out of 5, and 59 gave 4. Work emails rated highest, with 20 of 21 runs at 4.

How we tested

Seven agents ran each of five tasks three times: Claude Code (Opus 5.5 and Sonnet 5), Codex (GPT-6 Sol), Pi (GPT-5.6 Luna), Grok CLI (Grok 4.6), OpenCode (Kimi K3) and Factory (Kimi K3). Each run had a fresh sandbox, the same Clay key and the same 19 saved Clay routines, and a 60-minute limit; a timeout counts as 0. Agents ran one at a time; Clay credits are measured as the change in the Clay workspace credit balance. Platform failures were rerun. Tasks, prompts and graders are at github.com/promptingcompany/clay-bench.

Try it yourself

The Clay Agent Plugin ships Clay's skills to your agent up front, including the habits that keep credit use down; setup is one prompt (about 15 minutes), and there is no additional cost to use it. Install it, or give your agent a Clay API key and the Clay Public API docs, then put items 2 to 7 of the checklist in your prompt. Clay's pricing page explains Actions and Data Credits.

Common questions

What does an AI agent need to use Clay well? API access, saved Clay functions with known inputs, a search-first rule, match checks and permission to leave blanks; with these, the best agent found 82 of 100 CMOs for 285 credits, under 3 per company.

Does the Clay Agent Plugin cost extra? No. There is no additional cost to use the Plugin, and an enrichment that returns no result is not charged.

How quickly did agents get productive with Clay? Fast: Clay's API response named each routine's required inputs on an early call, and agents self-corrected in one call. 78 of 102 completed runs finished in under 15 minutes.