Most leads don’t go cold because they said no. They go cold because nobody followed up. A prospect fills out your form on Tuesday, you’re slammed until Friday, and by then they’ve already talked to a competitor who replied in ten minutes. The fix isn’t “try harder” — it’s an AI agent that watches for new leads and follows up on its own, around the clock, in a voice that sounds like you wrote it.
The good news: you can build this without writing a single line of code. Below is the exact recipe we use to set these up — what the agent actually does, which tools fit which situation, and the honest tradeoffs so you don’t waste a weekend on the wrong setup.
What “an AI agent for follow-up” actually means
People throw the word “agent” around loosely, so let’s be precise. A simple automation sends the same canned email to every new lead. An AI agent does something smarter: it reads the lead’s details, decides what to say, writes a personalized message, sends it, and then keeps following up on a schedule until the lead replies or you tell it to stop.
In practice, your agent will handle four jobs:
- Detect a new lead the moment it lands (form submission, ad lead, booked-then-no-showed call).
- Personalize a message using the lead’s name, what they asked about, and where they came from.
- Send it through the right channel — email, SMS, or WhatsApp.
- Sequence the follow-ups: wait a day, check if they replied, send a different angle, repeat 3–5 times, then stop.
That last part — stopping when someone replies — is what separates a helpful agent from an annoying spam machine. We’ll come back to it because it’s the step people skip.
The tools you’ll actually use
You need three pieces: a place leads arrive, an automation platform to run the logic, and an AI model to write the messages. The middle piece is where you’ll spend your time. Here’s how the realistic options compare.
| Platform | Best for | Honest catch |
|---|---|---|
| Make (formerly Integromat) | Visual builders who want full control over multi-step logic and cheap operation costs | The visual canvas gets busy fast; a 6-step sequence looks intimidating at first |
| Zapier | Fastest setup, biggest app library, “AI Agents” feature built in | Pricing climbs quickly once you have real volume and multi-step “Zaps” |
| n8n | People who want to self-host and avoid per-task fees; strong native AI nodes | Self-hosting means you babysit it; the cloud version removes that but costs money |
| GoHighLevel | Agencies and service businesses that want CRM + follow-up + SMS in one place | It’s a whole platform to learn and pay for — overkill if you just need follow-up |
For the AI writing itself, you’ll connect an OpenAI (GPT) or Anthropic (Claude) account — both have built-in modules on every platform above. You’re looking at roughly a cent or two per message generated, which is noise compared to the value of a recovered lead.
Our honest default: if you’re starting from zero and want it working today, use Zapier with its AI step. If you expect volume or want to keep costs flat long-term, build it in Make. We use Make for most client work for exactly that reason.
Building it step by step
This walkthrough uses Make and email, but the shape is identical on any platform. Total time for your first version: about an hour.
Step 1: Catch the lead
Create a new scenario and add a trigger that fires when a lead arrives. The most common triggers:
- “Watch responses” from your form tool (Typeform, Tally, Google Forms, your website form).
- A webhook — a unique URL you paste into whatever sends leads. This is the universal adapter; almost everything can POST to a webhook.
- New row in a sheet if leads land in Google Sheets or Airtable first.
Test it by submitting one fake lead. You want to see the real data — name, email, the message they typed — show up in the platform before you build anything else. Don’t move on until you do.
Step 2: Let the AI write the message
Add an AI module (OpenAI or Anthropic) right after the trigger. This is the heart of the agent. Your instruction to the model is everything, so be specific. A prompt that works:
“You’re a friendly sales rep at [your company], which does [one sentence on what you sell]. A new lead named {{name}} just submitted this message: ‘{{their_message}}’. Write a short, warm follow-up email — 4 sentences max, no corporate jargon, one clear question at the end to get them replying. Don’t invent details we don’t know. Sign off as [your name].”
The double-curly-brace fields pull in the real lead data. Two rules that save you embarrassment:
- Tell it what not to do. “Don’t invent details” and “don’t promise pricing” prevent the model from hallucinating a discount you never offered.
- Cap the length. Left alone, AI writes three paragraphs nobody reads. “4 sentences max” keeps it human.
Step 3: Send the first message
Add your email module (Gmail, Outlook, or an email-sending service) and map the AI’s output into the body. Send the first one to yourself, not the lead. Read it out loud. If it sounds like a robot or a template, fix the prompt and re-test. Spend real time here — this single message is what the lead judges you by.
Step 4: Build the follow-up sequence with a stop condition
One email isn’t follow-up; it’s a single touch. The money is in steps 2 through 5. Add a delay module (“wait 2 days”), then before sending the next message, check whether they already replied. Two ways to do this:
- Search the inbox for a reply from that email address. If found, exit the scenario — the agent’s job is done and a human takes over.
- Use a status field in your sheet or CRM. When a reply comes in, mark the lead “replied.” Each follow-up step checks this field first and bails if it’s set.
Then repeat: wait 2 days, generate a different angle (a case study, a quick question, a “should I close your file?” nudge), send, check again. Three to five touches spread over one to two weeks is the sweet spot. After the last one, stop. Chasing someone seven times doesn’t win deals; it gets you marked as spam.
Step 5: Add a human escape hatch
Route a copy of every AI-written message to a Slack channel or your email before — or right as — it sends. For the first week, glance at each one. You’ll catch the occasional weird output, and you’ll tune the prompt based on real cases. Once it’s been clean for a week, you can let it run unattended. Never skip this break-in period.
When this is NOT the right tool
We build these for a living, and we still talk clients out of them sometimes. Be honest with yourself:
- You get fewer than ~5 leads a week. Just follow up by hand. You’ll do it better, and the setup time won’t pay off. Automate when the volume genuinely hurts.
- Your sale is high-touch or high-trust (six-figure deals, sensitive services). An AI cold-follow-up can feel impersonal and cost you the deal. Use the agent only to notify you instantly and draft a starting point — then you send it.
- You’re tempted to do SMS without consent. Texting leads who didn’t opt in to texts isn’t an automation problem, it’s a legal one. Email the lead first; only text people who agreed to it.
- Your follow-up needs real account data (their order status, their balance). That’s a bigger integration than a beginner no-code build — start with the simple version and grow into it.
FAQ
Will leads be able to tell the message was written by AI?
Not if you do step 2 properly. The giveaways are length and tone — long, formal, jargon-filled emails read as automated. A four-sentence, casual message that references what they actually asked about reads as a busy human who cared enough to reply. The personalization is what sells it, not the technology. Test by sending drafts to a friend and asking if it feels human.
How much does this cost to run?
For most small businesses, under $30–40 a month all in: the automation platform (free tier to ~$20 depending on volume) plus a few dollars of AI usage, since each generated message costs roughly a cent or two. The real question isn’t the cost — it’s how many deals you currently lose to slow follow-up. Recovering one extra customer a month almost always pays for the whole thing many times over.
What if the AI writes something wrong or off-brand?
That’s exactly why step 5 exists. Run it in “review mode” first — every message hits your Slack or inbox before going out — for the first week or two. Tighten the prompt with explicit “don’t do this” rules as you spot issues. Once you’ve seen a clean stretch with no surprises, switch it to fully automatic. The break-in period is non-negotiable; the unattended payoff is the reward for doing it.
Your next step
Don’t try to build the whole five-step sequence today. Build step 1 to 3 only: catch one real lead, have the AI write one follow-up, and send it to yourself. That single working loop — trigger, AI, send — is 80% of the skill. Once you’ve seen a genuinely good message generate itself from real lead data, adding the delays and the stop condition is the easy part. Set aside one focused hour, grab a free Make or Zapier account, and get that first message sending. The leads you’re losing right now aren’t waiting.