When someone fills in a form on your site, the clock starts. The reply that comes within the hour lands very differently from the one that comes on day three.
The problem is that most inbound is not worth an hour. Someone curious about pricing, someone asking for free work, someone with no company at all. If you read every one with equal care, the one that matters waits in the pile.
Here is a small system I built in Clay to sort that pile, and what broke when I tested it.
The setup
The workflow does four things to each submission:
- Takes the form fill. A name, a company website and a message.
- Enriches the company. Size, industry, revenue, country and description.
- Scores it with AI. It reads the message and the company data together and returns a tier (hot, warm or disqualified), a score from 1 to 10, a one-line reason, and a suggested route.
- Routes it. Hot leads get a short summary written for a team alert. Warm leads go into a nurture sequence. Bad fits are tagged and closed.
For the test I used six made-up form fills, labelled as tests, each pointing at a real company website. In production the form would post to a Clay webhook, which is a paid feature, so I tested with a CSV instead.
What the test showed
| Lead | What it looked like | Result |
|---|---|---|
| Software company asking about starting outbound after its SDR left | Clear need, fits ICP | Hot, scored 9 |
| Software company asking about deliverability after a drop in opens | Clear problem | Hot, scored 9 |
| Software company that had just raised and wanted an outbound engine in 60 days | Clear need, urgent | Hot, scored 10 |
| "Just curious about pricing" from a B2B software company | Vague | Warm, scored 6 |
| A flooring store wanting more customers | Wrong fit | Disqualified, scored 2 |
| A Gmail address asking for a free lead list | No company, asking for free work | Needed a fix |
The AI sorted five of the six correctly. It also gave each hot lead a two-line summary written for the person who would call them.
The one that broke
The Gmail lead got no score at all. The enrichment step has nothing to look up for gmail.com, so the AI column ran on empty inputs and returned nothing.
I fixed it with a formula column that runs after the scoring. It checks whether the domain is a free email provider and, if so, returns "Disqualified, free email" instead of waiting for an AI answer. Everything else passes through to the AI's route.
That is the lesson I would take from the whole build: do not rely on AI to notice the obvious. A one-line rule for a case you already know about is cheaper and more reliable than a smarter prompt.
The rules I would put in the prompt
- Tell it what a good fit is, in plain words
- Tell it what to disqualify: free emails, local consumer businesses, requests for free work
- Ask for a reason, not just a score, so you can sanity-check it
- Ask for the alert text only when the lead is hot, so you do not generate noise
What I would not automate
I would not auto-reply to hot leads. The point of scoring is to get a person on the phone quickly, not to put a template in front of the best prospect. Let the system do the sorting and write you a summary, then answer the top few yourself.
Costs
On my test the enrichment cost about one credit a row and the AI column about a tenth. That is cheap enough to run on every submission. Real costs depend on the models and providers you choose, so read the estimate before you run, as I describe in how to stop burning Clay credits.
The short version
Enrich, score, route. Add a hard rule for the cases you already know about. Read the hot ones yourself. And remember that the speed of the first reply is worth more than the quality of any scoring model. If you are thinking about how the whole reply handling should work, what your cold email agency should report every week covers the numbers to watch.