Blog · AI & Outreach
February 5, 2025 · 6 min read
For most of the history of outbound sales, list building was a bottleneck. Finding the right people, enriching their contact data, and adding enough context to write a relevant email was slow, manual, and expensive. Clay's AI research agents changed that. Not completely, but meaningfully.
The agents work by running an AI prompt on each row of a Clay table. You can point them at a company's website and ask "what problem does this company solve?" You can ask them to summarise a prospect's most recent LinkedIn post. You can ask whether the company fits a specific set of criteria. They return a populated cell for every row, at the speed of an API call.
The most impactful use case is personalisation at scale. Before AI agents, a researcher could add specific hooks to maybe 20 or 30 accounts per day. With agents, the same researcher can configure a prompt that runs across 500 accounts in minutes. The quality of the output depends on the prompt — but well-designed prompts reliably produce usable hooks.
A second use case is qualification. You can run an agent that reads a prospect's job postings and flags whether they are actively hiring in a role that signals a buying trigger for your product. This kind of dynamic, content-based filtering was nearly impossible at scale before AI agents.
AI agents hallucinate. Not often on well-structured tasks with clear source material, but they do. A cell that says "this company recently raised a Series B" should be spot-checked before it goes into a personalisation line in an email. Sending a confident factual error to a prospect is worse than sending a generic email.
The other limit is that agents produce research — they do not produce a message strategy. You still need to decide what observation about a prospect is actually relevant to the problem you solve. An agent that notes "this company expanded to France last year" is useful only if expansion into France is connected to something you can help with.
The best practice is to constrain the task tightly. "Summarise this company in one sentence focused on their outbound sales challenges" produces better output than "tell me about this company." The more specific the prompt, the more usable the output.
Build in a review step for any AI-generated content that will appear verbatim in an email. Even a quick human scan of the first ten rows of a new table catches the category of errors before they reach a prospect.