The Weekly Signal · Episode 2 · 12:34

Watch this week’s video

One of the most unpopular things I say as an AI consultant for marketing teams is, “We actually shouldn’t be doing this in Claude.” Probably not what people expect from the person they hired to help them use AI.

I get how it happens. You open a chat box, ask it to do something, and it does something you didn’t even know how to do. I can build apps now. I’m not a programmer, and that is magical.

But being able to do something and finding the best way for your team to do it are different things. Sometimes we get excited about the first one and stop asking about the second. We add a token bill and a bunch of maintenance, and the original problem is still there.

A lot of the automation work I’m asked to do is roughly 90% code and rules and 10% model calls. Moving information, transforming it, following a known sequence. If this happens, do that.

Somewhere in there, we need a model to analyze something, recommend an approach, decide what belongs in an output, or write draft copy. Those are useful contributions. They don’t mean the model needs to run every step around them.

One client had a 20-step workflow that turned a long-form content concept into smaller pieces for five channels, including supporting images. They were running the whole thing in a Claude chat. Every time something broke or the output looked wrong, somebody had to figure out why.

I recommended moving the workflow into AirOps and using model-call nodes where they actually needed something from a model. It ran more reliably. Weekly maintenance went from six or seven hours to one or two. Token usage went from about 12,000 per run to about 1,200.

Those are the results from that client’s workflow, not a promise for every team. The useful change was that the person responsible could spend less time fixing it and more time making it better. AI still had a job. It just didn’t have every job.

I have clients building elaborate content workflows in Claude while an AirOps subscription on another team barely gets used. Another client built an impressive remarketing system with its own asset database and mailer integration. Roughly half of someone’s time was tied up managing it, and they were already paying for HubSpot.

I nervously asked: is there a reason we wouldn’t build this in HubSpot? It’s a nurture workflow. The other nurture workflows are already there, connected to the CRM.

We can still use a model to generate creative assets or write copy. The automation around that work may have a better home in software the team already owns.

Maybe your current tool doesn’t do exactly what you need. Maybe you need a different tool or an upgrade. But look first. An unused subscription and a workflow somebody has to babysit both cost you something.

Imagine you’re very good at making quarterly business review decks. You know the content and the layout. You spend an afternoon trying to build a Claude workflow to do it, keep revising the output, and eventually decide to make the deck yourself. Now you’ve spent the afternoon and still have a deck to make.

That doesn’t mean you should never automate the deck. It means you count the time it takes to get to something you’re actually happy with. The first output in a chat box often isn’t the finished work.

AI lets me build things I couldn’t build before. An experienced programmer using AI may still make something better, faster. AI makes both of us more capable; it doesn’t erase the difference in what we know about the job.

And while I’m doing the programmer’s job, nobody is doing mine.

If you’re a solo founder or a tiny marketing team, that may be the right trade. You might not have a programmer to ask, and getting something working today can be valuable. On a bigger team, I want to know whether we’re using the expertise we already have. Taking a ticket off someone’s plate doesn’t help much if they have to redo it later and you’ve spent the afternoon away from your own work.

Choose something your team spends a lot of time maintaining, or something you’re about to automate. Before putting another afternoon into it, ask:

If it isn’t materially improving time or quality, let the human keep doing it while you find a better approach. You don’t have to keep using a workflow just because you worked hard to build it.

Sometimes the useful next step is a model call. Sometimes it’s automation in a tool you already own. Sometimes you just need to let the human do it.

Have you moved a task out of an AI workflow and back to another tool or person? What made you decide to change it?

Share an example in the comments. I’d love to hear what you tried and what ended up working better.