When somebody asks me where the no-brainer AI wins are, the things I’ve seen work in the real world without turning into a six-month project, I usually tell them to think smaller.
Most people go straight to a whole role. They ask how to automate the role, or they say they don’t have an SDR and want to create an AI SDR.
That’s one of my least favorite conversations. An SDR is often the first human interaction somebody has with your company, and you want to hand that over to a bot before you’ve figured out the day-to-day for that role. I don’t think that makes sense right now.
Start with one task that takes somebody 15 minutes.
The 15 minutes before the meeting
A sales rep spends 15 to 20 minutes getting ready for a meeting. They look at the company’s website, check LinkedIn, and try to learn enough about the person they’re meeting that they can bring up something relevant on the call. If they’re doing it well, it takes a real amount of time before every meeting. And there are a ton of questions that are only relevant to your company and product, so the default features you’re seeing pop-up in HubSpot and Salesforce that do this out-of-the-box are garbage.
When the meetings are back-to-back, that prep usually gets skipped, or the rep gets up a couple of hours early and does nothing but prepare for calls.
AI can handle that piece. You can have it look at the calendar, check the CRM, go through the sources you choose, build a prep note, and drop it into the CRM. That’s the same pattern as pulling CRM fields out of the call transcript, just running before the meeting instead of after it.
I build that exact template in Gumloop for almost everybody I work with because it saves reps a lot of time. The rep still runs the meeting. The workflow handles the research that happens before it. You can build a lighter version of it with just a really good recurring prompt in Claude Cowork that skips all approvals and has access to all of the relevant tools.
That is a small task somebody does before a demo. It has a clear starting point, a clear output, and a person can tell pretty quickly if the note is useful. You can fix the workflow without someone needing to scope a whole project.
The other thing I’ve been doing on Reddit
I’ve been on Reddit for a long time. People post about problems they’re having, and some of those problems are things I can help with. I used to read posts, answer questions, and send a message when I had something useful to add. I was doing that before I was trying to turn it into a way to get meetings.
The problem is the volume. I can’t keep up with 200 subreddits every day, and I never thought of that as a lead acquisition strategy because the manual work was too much.
Now I have an agent scan those 200 subreddits, compare posts to my ICP, and check whether I have a past project that solved something similar. My Notion is full of specific work I’ve done with past clients, so the agent can find a post and connect it to an example I can talk about.
In the morning, I get a Notion task that says this person looks like a nine out of ten fit, explains why, and points me to the project that might be relevant. I still decide whether I want to respond, and I write the final message myself.
I won’t copy and paste the draft the AI gives me (please don’t do that, it’s so obvious). People on Reddit can spot that pretty quickly. But finding the post, choosing the right example, and getting a first draft down used to take too much time for me to do this at any useful scale.
I’ve shared the Gumloop template for that with around 20 people. It works best for an early-stage company trying to get its first 50 to 100 clients. Once the company gets past that point, it becomes more of a steady source of opportunities than a main acquisition channel. Setup some time with me if you’re interested, I’ll shoot it over to you and walk you through Gumloop too.
When the volume goes up
The part that matters once these workflows run more often is consistency.
If you give Claude access to a tool through an MCP server, which is the connection between an AI model and another piece of software, and put that workflow on a recurring schedule, the model will sometimes handle the same process a little differently. It doesn’t have to make up an answer. It might just interpret one step differently than it did the last time.
At 100 runs a month, you might never see the problem. At 2,000 or 5,000 runs a month, the small differences start to matter.
That is when I move the repeatable action into a more deterministic workflow (uses code, works the same way every time). You can do this with Gumloop, vibe code a solution, add a custom-coded action to a HubSpot workflow, etc. The exact way to do it depends on the action, and picking the right tool for the job is most of that decision. It also uses fewer tokens (since it’s not AI performing a complex task anymore).
Text instructions are fine when you’re starting. They let you test whether the workflow is worth keeping. Once the volume goes up, I want the part that needs to happen the same way every time to live in a template instead of a paragraph of instructions.
Start with the task you describe in full
The way I start most engagements is with a two-week audit. I show the team what I would do and how I would do it, including the smaller workflows that are worth building first.
You can do this yourself. I tell people how I’d set it up because the work is usually more about the time it takes than some secret method I need to keep from you. You probably have something more pressing to work on than a workflow I’ve already built a hundred times, and if you’d rather learn how to do it, that’s exactly what my coaching engagements are for.
The reason “think smaller” works is that you can actually describe a 15-minute task. You cannot describe everything a person does all day, which is why scoping the human work first is the step people skip and then regret.
Start with the one 15-minute task someone is doing a few times a day. Build that. Watch what breaks. Fix it. Then move on to the next thing. This is so much more scalable than building an AI SDR or support person. I made the same argument in shorter form in issue 003 of the newsletter.