Clay has become shorthand for a whole category of work, and the category is poorly explained. Here's what someone is actually doing when they say they build in Clay.
What Clay is
A spreadsheet where every cell can call an API.
You put a list of companies in. Each column runs something against each row — find the CEO, find their email, check the email is real, read their website, ask a model a question about what it found. The output is a list that is enriched, verified and scored rather than scraped.
That's it. The power is entirely in what you ask each column to do.
What the work actually is
Four things, in rough order of how much they matter.
The waterfall. No single data provider has good coverage. A waterfall tries one, and only pays for the next if the first came back empty. Done properly this roughly doubles coverage and cuts cost per verified contact, because you never pay twice for the same row. Done badly it burns credits finding emails you already had.
Qualification logic. Filters can tell you a company has 50–200 staff. They cannot tell you whether it's actually a B2B SaaS company, because no database has a clean field for that. So you have a model read the website and decide. Getting that prompt right, on your specific ICP, is most of the difference between a list that converts and one that doesn't.
Research columns. Anything you'd otherwise have a human check: do they have a careers page, how many sales roles are open, what does their pricing page say, do they mention a competitor. This is what makes a first line specific instead of generic.
The plumbing. Getting the output into the sequencer, and replies back into the CRM, without anyone touching a CSV.
When you don't need an agency for this
Be honest about which of these you're buying.
If your list is under a few thousand rows, your ICP is easy to define by filters, and you already have verified emails — you don't need Clay and you certainly don't need someone to build it for you. Buy a list, check it, send.
If you're paying someone mainly to click around in a tool, that's a training problem, not an agency problem. The interface is learnable in a fortnight.
When you do
Three situations where the work genuinely pays for itself:
Your ICP can't be expressed as filters. "Companies that sell to restaurants" or "agencies that have started offering outbound" don't exist as checkboxes anywhere, and a model reading websites is the only way to find them at scale.
Your coverage is bad. If you're finding emails for under half your list, a waterfall is worth more than better copy.
You're running it every week. Anything that runs once is cheaper done by hand. Anything that runs continuously needs to be built properly, or you'll rebuild it every Monday.
The question to ask before you hire anyone
Ask what happens when a provider's API changes or returns something unexpected.
Anyone can connect four columns on a good day. The job is what the table does when Prospeo returns nothing, when a site is behind Cloudflare, when the model confidently classifies a dental practice as a SaaS company. That's the difference between someone who followed a tutorial and someone who has run this at volume — and it's the only interview question on this topic that reliably separates them.