On this page
Can an AI agent do lead research and enrichment?
Yes, with a person sampling the output. An AI agent for lead research can take a new lead's name, email or company domain, look the company and person up in data providers and on the open web, answer the questions you set ("Does this company publish a security page?"), and write the answers into your CRM. Clay's documentation describes agents that navigate websites, extract information and make decisions, and HubSpot documents smart properties that answer a prompt per record from web research, the company website or call transcripts.
Teams run this in production, according to Clay's customer stories. Clay reports that OpenAI's RevOps team used its agent to summarize earnings reports and company websites for every relevant account and more than doubled enrichment coverage, "from low 40% to high 80%". Clay reports that Vanta reached 80%+ enrichment coverage, and that Verkada's growth team, which used to rely on external contractors to research prospects, says tasks "that would take a week now take minutes or hours". All three customer stories on this page come from the same vendor.
The agent should not decide who your customer is, contact anyone, or decide how personal data is used. Treat it as a supervised hire. This page covers research and enrichment only; for writing and sending outreach see outbound prospecting, and for keeping CRM records clean see CRM updates and data hygiene.
What the agent does and what stays with you
Human
- Define the ideal customer profile and the research questions
Agent
- Find the company's domain, size, industry and location for each lead
- Find or validate work emails, phone numbers and profile links
Agent with review
- Answer custom research questions from the web (for example, does this company have a security page)
- Summarize recent company news, filings and changes for a rep
- Score or tier leads against the ICP
- Write results into CRM fields and route the lead
Human
- Decide the lawful basis and privacy notice for data collected about people
8 tasks: 2 agent, 4 agent with review, 2 human
| Task | Who does it | Why |
|---|---|---|
| Define the ideal customer profile and the research questions | Human | Clay's builder pulls the company description, ICP and buyer personas from workspace settings that a person fills in; the agent works to that definition, it does not set it. |
| Find the company's domain, size, industry and location for each lead | Agent | Vanta's customer story describes Clay gathering job titles, company size, revenue and location for each inbound lead; firmographic fields can be checked against the company website by sampling. |
| Find or validate work emails, phone numbers and profile links | Agent | Vanta's story says the workflow validates email addresses, adds phone numbers and cross-checks LinkedIn profiles; bounce rates show errors quickly. |
| Answer custom research questions from the web (for example, does this company have a security page) | Agent with review | Clay documents agents that navigate websites and extract information, and HubSpot documents smart properties that answer a prompt from web research; answers are judgment calls, so a person samples them. |
| Summarize recent company news, filings and changes for a rep | Agent with review | OpenAI's RevOps team used Clay's agent to summarize earnings reports and company websites; the rep reads the summary before using it in outreach. |
| Score or tier leads against the ICP | Agent with review | Clay gives a lead scoring agent that outputs a 1–100 score with rationale as an example; scores decide who gets a rep's time, so a person reviews the rules and a sample. |
| Write results into CRM fields and route the lead | Agent with review | Vanta's story describes enriched signals routed to the account owner in Salesforce; a person decides the field mapping and routing rules and checks overwrites. |
| Decide the lawful basis and privacy notice for data collected about people | Human | The UK ICO says organizations that obtain personal data from another source must tell people the categories and the source; that is a controller's decision, not an agent task. |
The rule behind the split: the agent fills facts a reviewer can check against a source, and a person owns decisions about targeting and about people's data. Firmographics and contact validation go to the agent, because a bounce or a wrong headcount shows up quickly in a sample. Custom research answers, summaries and scores get review, because they are judgment calls that reach a prospect through the rep. The ICP and the privacy decisions stay human.
Clay's own builder reflects this. Its agents read the company description, ICP and buyer personas that a person enters in workspace settings, and Clay suggests turning web search off when the agent works from data already in your table, "to keep runs faster and more consistent".
A worked example: one lead
Assume, for this example, that you sell compliance software and a lead fills in your demo form with a work email at a 200-person logistics company.
The agent takes the email domain, finds the company website, and fills industry, employee band and headquarters country. It checks the email and finds the person's title. Then it answers your three research questions: does the company show a SOC 2 badge or a security page (no), is it hiring for a security role (yes, one posting found, with a link), and has it announced funding in the last 90 days (no result, recorded as "unknown"). It writes a two-line note with the links and scores the lead 72 against your ICP.
Your sales ops owner sees the lead in the weekly sample, opens the job posting link and confirms it is real and current. The rep reads the note before the call and opens with the open security role, not with a guessed revenue figure. The agent sent nothing and touched no other field.
What you need to provide
The agent touches personal data (names, work emails, phone numbers, job titles) and company data, and it needs write access to your CRM. Keep the access narrow:
- A written ICP and a short list of research questions with fixed answer options. Vague questions get vague answers.
- An integration user in the CRM that can edit lead, contact and company records, limited to the fields the agent fills.
- A rule for overwrites: which fields reps own and the agent must leave alone.
- Accounts for any data providers you already pay for.
- Your privacy notice and whoever decides data protection for prospect data.
Rules on data you collect about people vary by jurisdiction; ask counsel which apply to you. For the UK, the Information Commissioner's Office says that when you obtain personal data from a source other than the person, you must tell people the categories of data and where you got it, and if it came from a publicly accessible source you must say this. Separately, site terms limit where an agent may look: LinkedIn's User Agreement prohibits crawlers, scripts or other software that scrape profiles, and HubSpot notes that prompts which use LinkedIn are filtered out of its smart properties.
On cost, both sides of the comparison are not available per lead, so this page gives the units instead. Clay's pricing page lists a free plan with 100 data credits a month, Launch at $167 a month with 15,000 actions and 3,000 data credits a month, and Growth at $446 a month with 40,000 actions and 6,000 data credits a month. Those headline prices equal the annual-billing add-ons on the same page ($54 + $113 for Launch, $185 + $261 for Growth); billed monthly, Launch's add-ons come to $60 + $125 = $185. How many credits one lead uses depends on the providers and prompts, and Clay publishes no fixed figure. HubSpot charges HubSpot Credits per smart property fill, even when no value is found. For a human benchmark, BLS data on O*NET OnLine put the 2025 national median wage for sales representatives of services, the closest occupation, at $69,990 a year ($33.65 an hour); research is only part of that job. Measure credits per enriched lead in your trial.
Job brief template
Give this brief to a vendor, an operator or your RevOps team. Replace the bracketed values with your own.
Copy-ready brief
- Goal
- Every new lead in the CRM has its company, role and contact fields filled and the agreed research questions answered within one business day of creation, with a source for each answer.
- Deliverable
- Filled CRM fields on each new lead and its company, a short research note with source links, and a weekly list of leads the agent could not resolve.
- Inputs
- A written ideal customer profile and 3–8 research questions with allowed answers (for example yes / no / unknown)
- API or integration access to the CRM, limited to lead, contact and company records and to named fields
- A list of the fields the agent may write and the fields it must never overwrite
- Accounts or API keys for any data providers you already pay for
- Your privacy notice and the data-protection rules that apply to prospect data
- Acceptance criteria
- On a 50-lead sample, company domain, industry and employee band match the company's own website or filing with at most 3 errors.
- On the same sample, each research answer has a source link, and at most 5 answers are wrong or unsupported by that link.
- Work emails the agent marks as valid bounce at a rate of at most 5% in the first send.
- No field outside the agreed list is written, and no value entered by a rep is overwritten.
- Every lead created on a business day is enriched or listed as unresolved by the end of the next business day.
- Systems
- CRM (for example HubSpot or Salesforce)
- enrichment or research platform (for example Clay or HubSpot's data agent)
- third-party data providers
- Constraints
- Leads are personal data about named people; collect only the fields the brief lists.
- No scraping of sites whose terms forbid it; LinkedIn's User Agreement prohibits scraping profiles.
- The agent may not send messages, create sequences or contact anyone.
- Your retention and data-location rules apply to copies the provider stores.
- Volume
- Example: 500–5,000 new leads a month
- Cadence
- Example: on lead creation, with a weekly sample review
- Human checkpoints
- Before launch: RevOps approves the field list, prompts and overwrite rules.
- Weekly: a sales ops owner reviews a 50-lead sample against the acceptance criteria.
- Before outreach: the rep reads the research note before using any fact in a message.
- Privacy: the data-protection owner approves the sources and the notice given to people.
Keep the source-link requirement. An answer without a link cannot be checked, and an unchecked fact is the one a rep will repeat on a call.
Risks and human checkpoints
The agent states a wrong fact (revenue, headcount, a tool the company uses) and a rep repeats it to the prospect.
How you notice it
Require a source link per answer and check a weekly sample; OpenAI's story notes the agent showed its reasoning, for example how it determined revenue from an earnings report.
Enrichment overwrites good data that reps entered, or several tools write conflicting values to the same field.
How you notice it
Lock rep-entered fields and log every overwrite; Vanta's RevOps director described five tools enriching the same data point before consolidation.
Personal data is collected without the notice or basis your privacy rules require.
How you notice it
List every source and field in a register before launch and have the data-protection owner sign it off; the UK ICO expects you to name the source, including publicly accessible ones.
The agent collects data from sites whose terms prohibit automated scraping.
How you notice it
Review the sources in the agent's notes; LinkedIn's User Agreement bans crawlers and scripts that scrape profiles, and HubSpot says prompts that use LinkedIn are filtered out.
Spend runs ahead of value because every run consumes credits, even empty ones.
How you notice it
Track credits used per enriched lead weekly; HubSpot states a smart property consumes credits even when no value is filled.
Human checkpoints
- Before launch: RevOps approves the field list, prompts and overwrite rules.
- Weekly: a sales ops owner reviews a 50-lead sample against the acceptance criteria.
- Before outreach: the rep reads the research note before using any fact in a message.
- Privacy: the data-protection owner approves the sources and the notice given to people.
The visible risk is a wrong fact in a rep's opening line. The quieter one is data quality drifting in the CRM. Vanta's RevOps director described, before consolidating on Clay, "five different tools trying to enrich the same data point, each with their own field in our CRM". An agent added without overwrite rules can create the same mess. Decide which source wins for each field before the first run.
How to run a trial
Run the agent on a slice for four weeks before it writes to your main CRM fields.
- Pick one lead source, such as demo requests, and 3–5 research questions.
- Have the agent write to new, separate fields so nothing a rep relies on changes.
- Each week, check 50 leads against the acceptance criteria: firmographics against the company website, each research answer against its link, and email bounces.
- Record every "unknown" and every wrong answer. A confident wrong answer matters more than an honest "unknown".
- Track credits or actions used per enriched lead, and the reviewer's minutes per lead.
- Ask two or three reps whether they used the notes, and which answers they ignored.
Pick one lead source and 3–5 research questions
For example, demo requests
Have the agent write to new, separate fields
Nothing a rep relies on changes
Each week, check 50 leads
Firmographics against the company website, each research answer against its link, and email bounces
Record every "unknown" and every wrong answer
A confident wrong answer matters more than an honest "unknown"
Track credits or actions per enriched lead
And the reviewer's minutes per lead
Ask two or three reps
Whether they used the notes, and which answers they ignored
If two weekly samples in a row meet the criteria, map the fields into the ones reps use and widen the slice. For structuring the test, see how to run a paid trial with an AI agent; for writing the brief, how to write a job description for an AI agent. Related: the SDR role task by task, competitive intelligence and the sales overview.
Sources
- Claygent Builder
ClayPrimaryAccessed Nov 14, 2026
- AI in Clay
ClayPrimaryAccessed Nov 14, 2026
- Automatically populate records using smart properties
HubSpot Knowledge BasePrimaryAccessed Nov 14, 2026
- What privacy information should we provide?
Information Commissioner's Office (UK)PrimaryAccessed Nov 14, 2026
- 41-3091.00 - Sales Representatives of Services, Except Advertising, Insurance, Financial Services, and Travel
O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026
- LinkedIn User Agreement
LinkedInPrimaryAccessed Nov 14, 2026
- How OpenAI is scaling their GTM motion with Clay
ClayVendorAccessed Nov 14, 2026
- How Vanta uses Clay to streamline RevOps and scale signal-based prospecting
ClayVendorAccessed Nov 14, 2026
- How Verkada's growth team uses Clay to launch targeted ABM campaigns
ClayVendorAccessed Nov 14, 2026
- Compare plans, features & costs
ClayVendorAccessed Nov 14, 2026
What we could not verify
- All production evidence on this page (OpenAI, Vanta, Verkada) is published by one vendor, Clay, as customer stories, and the story pages carry no publication date. Clay's Anthropic story was left out because search listings date it to more than 24 months ago. No independent study of research-agent accuracy on lead data was found; an 81.5% accuracy figure for Claygent seen in search results came from a review site with no stated method and was dropped.
- How many Clay actions or data credits one enriched lead consumes is not published as a fixed figure, so no per-lead price is given and the cost comparison block is omitted.
- How many HubSpot Credits one smart property fill uses was not checked; HubSpot's article links to a separate page for it.
- Clay's pricing page shows $167 a month for Launch without stating the billing term; that it is the annual-billing price is our inference from the page's own add-on figures ($54 + $113).
- Notice and lawful-basis rules outside the UK (EU GDPR, US state privacy laws) were not researched for this page; EUR-Lex was not fully available on 2026-11-14.
Related
- AI agents for sales: which jobs to hand over first
- Can AI replace SDRs? The sales development job, task by task
- AI SDR: an AI agent for outbound prospecting
- AI agent for CRM data hygiene: clean and dedupe records
- AI agent for competitive intelligence: what it does
- AI agent job description: how to write one, with template
- How to test an AI agent before buying: run a paid trial