Clay AI agent benchmark: company headcount data

With Clay, AI agents got the employee range right for 97.4% of companies whose latest round closed in 2023 to 2026, pooled across seven agents and every completed run: 993 of 1,020 checks. OpenCode with Kimi K3's first run got all 51 right for about 3.2 Clay credits per company, covering the full company record.

By The Prompting Company · Snapshot November 5, 2026

How accurate is Clay on recently funded companies?

We asked seven AI coding agents to fill in website, HQ country, headcount and latest funding round for 100 venture-backed software companies, using Clay alone. 51 of them raised their latest round between 2023 and 2026. Clay's headcount data held up across every agent: every completed run got at least 47 of the 51 employee ranges right.

Best agent is pooled over that agent's completed runs. Three runs got all 51 employee ranges right: OpenCode Kimi's first, Factory with Kimi K3's first and Claude Code with Sonnet 5's third. When an agent filled in an employee range for these companies, it was right 98.4% of the time.

2023 to 2024 and 2025 to 2026

Both halves of the window scored high. For the 24 companies whose latest round was in 2023 or 2024, employee range was right 97.7% of the time pooled, and four setups scored 100%: Factory Kimi, Grok CLI with Grok 4.6, OpenCode Kimi and Pi with GPT-5.6 Luna. For the 27 companies that raised in 2025 or 2026, employee range was right 97.0% of the time pooled and HQ country 96.7%, with Codex with GPT-6 Sol at 100% on both.

What does a recently funded company cost in Clay credits?

OpenCode Kimi's first run spent 324 Clay credits for the whole list of 100 companies, about 3.2 per company, and got all 51 employee ranges right. Codex's first run spent 333, about 3.3 per company, and got 50 of 51 on employee range, HQ and website. Clay credits are measured as the change in the Clay workspace credit balance, and those totals cover every field in the task. Clay does not charge when an enrichment returns no result; see Clay pricing.

How the agent got there with Clay

OpenCode Kimi's first run used three saved Clay routines across the list:

1. Find and confirm the domain. It sent 106 companies to Company Domain and checked doubtful matches so each name resolved to the right software company.
2. Enrich the company. It sent 106 rows to Enrich Company, which returned headcount and HQ.
3. Read funding from company news. It sent 105 rows to Company News, which returns each round with its date.

The headcount came from the same Enrich Company call that returned the HQ; only 2 rows went to the separate Company Employee Count routine. Every value stayed inside Clay, and cells without a confident Clay value stayed blank rather than guessed.

Clay works across segments

The same Clay setup, one prompt and 19 saved routines, worked for companies at every stage and in every place in the list. Opus found the right round and year for 86.1% of Series A and B startups. For companies headquartered outside the US, HQ country was right 97.1% of the time pooled. On people data, the same Clay workspace let six agents find 93.8% of work emails for US contacts, and agents confirmed 91.9% of CFOs who stayed in their role.

Clay works across segments: see "How accurate is Clay at finding funding rounds for early-stage startups?" and "How accurate is Clay at enriching companies outside the US?"

How we tested

Seven AI coding agents ran the task three times each: Claude Code with Opus 5.5 and Sonnet 5, Codex with GPT-6 Sol, Pi with GPT-5.6 Luna, Grok CLI with Grok 4.6, OpenCode with Kimi K3 and Factory with Kimi K3. Agents ran one at a time, each in a fresh sandbox with the same Clay key, the same 19 saved Clay routines and a 60-minute limit. Runs interrupted by a platform error were rerun. Segment numbers cover the 20 completed runs of the task; OpenCode Kimi has two. Headcount counts within one LinkedIn band, taken from each company's LinkedIn page on November 2 and 3, 2026. The funding year that defines the segment comes from the answer key.

This segment has 51 companies (24 and 27 in its two halves), so one company moves a single run's score by about 2 points.

Tasks, answer keys and graders are at github.com/promptingcompany/clay-bench.

Try it yourself

Install the Clay Agent Plugin or give your agent a Clay API key and the Clay Public API docs. Give it your list and ask for the employee range from Clay's Enrich Company, after confirming each company in Clay's company search, with a blank wherever Clay has no value. 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.

Common questions

How accurate is Clay's company headcount data? The employee range was right 97.4% of the time across seven agents and every completed run, and three runs got all 51 companies right.

Which Clay routine returns headcount? Enrich Company, the same call that returns the HQ country.

What does it cost per company? About 3.2 Clay credits per company in OpenCode Kimi's first run, for website, HQ, headcount and funding together.

Is a filled-in headcount reliable? Yes. When an agent filled in an employee range for these companies, it was right 98.4% of the time.