Rémi KokabiNovember 27, 20256 min readLast update: Oct 7, 2026

I Rebuilt lemlist's GPT-3 Reply Workflow With GPT-4 in 4 Steps (Here's What Changed)

Run a campaign at any real volume and your inbox turns into a sorting problem. Interested leads, out-of-office bounces, unsubscribe requests, polite “not now” replies. All of it lands in the same place, and a human has to read every line to decide what happens next. That human is usually your best rep, and they’re doing it instead of selling. Meanwhile, the window for responding to a hot lead is tiny: lemlist’s own developer documentation sets the bar at warning an SDR no later than two minutes after someone interested replies. Nobody hits that by hand.

In this post, I’ll show you exactly how to rebuild the old lemlist x GPT-3 reply-categorization workflow on a model that still exists, step by step, in n8n or Make. I’ll also show you the part most readers won’t need to build at all, because lemlist now does reply classification natively. One quick warning before you start: the original version of this guide ran on GPT-3, and OpenAI retired the entire GPT-3 generation on January 4, 2024, as documented on OpenAI’s model deprecation page. If you tried to follow the 2023 instructions today, the API call would fail on the first node. This version uses GPT-4o (or whichever current OpenAI chat model you prefer).

Here’s what the finished workflow handles:

  • Slack alerts when a lead is interested

  • task creation when a lead is out of office

  • unsubscribing leads who ask to be removed

  • pushing the outcome into your HubSpot CRM

Before you build this: lemlist’s native AI reply detection

I’ll concede the obvious thing first. A lot of people land on this guide and don’t actually need it.

lemlist AI already reads incoming replies and marks leads as interested automatically. Positive reply, the lead is marked Interested. Negative reply, Not Interested. Either way the sequence stops for that lead, so nobody gets a follow-up after they’ve already answered. You turn it on, you handle the rest from your unified inbox, and you’re done in a couple of clicks.

So why build anything at all?

Because native classification gives you two buckets, and some teams need five. If you want out-of-office replies to create a task dated three weeks out, referrals to route to a different rep, and unsubscribe requests to hit a suppression list in your CRM, you need your own logic. That’s what the rest of this guide is for. Think of it as the power-user layer on top of a feature you’re already paying for, not a replacement for it.

If you want the broader version of this idea (AI agents acting on signals across your whole pipeline, not just replies), look at lemAgent before you start wiring nodes together.

Step #1: Get your OpenAI API key

→ If you don’t already have an OpenAI account, go to platform.openai.com to sign up.

→ Once you have an account, it’s time to get your API key:

  1. log in to platform.openai.com

  2. open the API keys section in the left-hand menu

  3. create a new secret key

When you set the model in your workflow later, pick a current chat model. GPT-4o is the safe default for classification work like this: it’s fast, cheap per call, and accurate enough that I’ve never needed anything heavier for sorting replies into buckets. GPT-4.1 works too. Anything starting with text-davinci is dead, so don’t copy old blueprints blindly.

A note worth reading before you create a key at all. If all you want is model flexibility without managing your own API key, billing and rate limits, lemlist’s Smart messaging AI now lets you choose between Claude, GPT and Perplexity directly inside the product. AI variables run on lemlist credits, which removes the external API key from the equation entirely. Grab a key only if you specifically need a model or a prompt structure that lives outside lemlist.

Step #2: Go to your lemlist account to get your API key

  1. If you don’t already have a lemlist account, go to lemlist.com and sign up for a free trial

  2. In the lemlist app, click on the Team settings drop-down menu

  3. Go to Integrations

  4. Under the API section, click the Generate new key button (if you don’t already have an API key)

Store this one somewhere you’ll find it again. You’ll paste it into several nodes.

Step #3: Create an n8n or Make account

You can use either of these no-code workflow automation platforms to continue your integration process:

→ Option 1: if you don’t already have an n8n account, start a 30-day free trial at n8n.cloud

→ Option 2: if you don’t already have a Make account, start your free trial at make.com

I use n8n when I want to see and edit the prompt as plain text, and Make when I want something a non-technical teammate can audit later. Both get you to the same place.

Step #4: Build your sales automation workflow

Before you touch either tool, bookmark lemlist’s developer guide to Slack alerts on positive replies. It’s the officially maintained version of this exact pattern: a small AI agent built on OpenAI that classifies a lead’s reply, chained to the lemlist trigger node. It gets updated when the API or the node changes. The blueprints below are the fuller build with CRM and task handling layered on, and the terminology matches the developer guide so you can move between the two.

Option 1: Build the automation workflow in n8n

  1. Log into your n8n account and choose to create a blank workflow

  2. Go to this link and click on “copy the workflow”

  3. Paste the new workflow into your blank workflow

Now, your workflow should look like this:

4. For each node, you’ll have to set up your credentials.

For the OpenAI node:
→ Choose Header authentication
→ Name = Authorization
→ Value = Bearer YOUR_API_KEY

Set up lemlist Auth:
→ create a new credential
→ paste your API key

P.S. you can reuse the same authentication process for all lemlist nodes except “lemlist – Mark as interested”. For that one, choose “Predefined Credential Type” → Lemlist API → the same credential you created.

5. When all nodes are set up, activate your workflow.

If you want to add data to your HubSpot CRM automatically, you can do it through n8n:

  1. In the HubSpot app, go to Settings

  2. Under Integration, choose to create a private app

  3. Allow the following scopes:

4. After creating the app, copy the token
5. Choose an App Token for Authentication
6. Create a new credential and paste your token

You can reuse the same authentication process for all HubSpot nodes.

Option 2: Build the automation workflow in Make

  1. Download the blueprint JSON here

  2. Log in to the Make app and click on Scenarios to create a new scenario

  3. Choose “Import Blueprint”

4. For each module, you’ll have to authenticate yourself:

  • lemlist → add your API key

  • HubSpot → use OAuth

  • OpenAI → add your API key

5. Save your module and activate it.

One thing I’d change in the imported blueprint before you switch it on: open the OpenAI module and check which model it calls. Old blueprints still point at retired GPT-3 endpoints. Swap it for your current model, then run a single test reply through the scenario and read the classification output before you let it touch live leads.

P.S. You can build the same integration with Slack, Pipedrive, Salesforce and plenty of other apps. Here are all the options.

Final thoughts

Keeping your outreach replies organized is important to improve efficiency and keep your campaign data accurate.

But the real reason to automate reply classification is speed. The faster your reps answer an interested lead, the higher their odds of converting that lead, and lemlist’s developer guide puts the target at alerting an SDR within two minutes of a positive reply. A human reading an inbox twice a day cannot do that. An AI agent reading every reply the second it lands can.

Start with the native option, because it costs you nothing to switch on and covers interested versus not interested out of the box. Build the n8n or Make version when your routing logic outgrows two buckets. That’s the whole decision.

If you want to see what the rest of the platform does before you commit, lemlist holds a 4.6 out of 5 rating on G2 across more than 1,400 verified reviews. And if you want more workflows like this one, join the lemlist family, where 5,000+ sales and outbound professionals share tips and workflows daily.

Start a 14-day free trial and let the AI read your replies for you.

Launch AI outbound campaigns

Automate prospecting and personalize AI-powered sequences to book more meetings.

Rémi Kokabi

Rémi Kokabi

Co-founder @GTM Club

Share this post