Not Everything Needs an Agent

I build AI agents for a living. So take this in the spirit it's intended: most of the problems people bring me described as agent problems are not agent problems.

The single most satisfying thing I ever shipped in marketing operations had no model in it anywhere. It was a spreadsheet, a sync, and a tag. It took a fraction of the time an agent would have, and it fixed the thing completely.

The stress I remember from every conference

I spent years in marketing operations, and a lot of that was events. Conferences, booths, meetups. There's a specific feeling that shows up on the flight home from every single one, and if you've run events you already know it.

It's the quiet certainty that you're losing leads right now.

Somebody has business cards in a jacket pocket. Someone else has names in their phone notes. There are scribbles on the back of a printout. Everyone had great conversations and nobody has a system, and every day that passes, more of that context evaporates. Two weeks later somebody asks whether the event was worth the budget and the honest answer is that nobody can tell you.

The obvious fix is to make everyone fill in a form. Scan the badge, log the conversation, tag the lead.

I've watched that fail everywhere I've seen it tried.

You cannot form your way out of this

Here's what I eventually accepted: any solution that depends on a salesperson doing data entry at a conference is not a solution. It's a wish.

They are standing at a booth. They are mid-conversation. They are running to the next meeting. The person in front of them is worth more than your form, and they are correct about that. Every time you put a field between a rep and a conversation, the conversation wins, and it should.

This is where I see people reach for an agent. If the humans won't do the capture, build something clever that does it for them — parse the emails, listen to the calls, infer the context, reconstruct who they met. It's a genuinely interesting build. I've been tempted.

But the interesting build was solving a problem I'd invented. The reps weren't failing to capture leads because capture is hard. They were failing because I'd put the capture in the wrong place in their day.

What I actually built

Names go in a Google Sheet. Whatever way is easiest in the moment — badge export, a row typed on a phone, a colleague dumping a list in after dinner. No fields to argue about. No app to open.

From there it's plumbing. The sheet syncs into the CRM. Every lead gets tagged to that specific event, so attribution is automatic and nobody has to reconstruct it later. And then the whole list goes straight into a sequence, so the follow-up starts while the conversation is still warm in someone's memory.

At one event we ran it end to end — sheet, sequence, event attribution — and the follow-up was out fast enough that people replied referencing the actual conversation they'd had at the booth.

There was no agentic solution here. It was purely using the technology that already existed, arranged correctly.

And the post-event stress just… stopped. Not because the tech was impressive. Because the capture step now sat where the humans already were, instead of where I wanted them to be.

The industry is about to learn this the expensive way

I'd normally leave a post like this at "here's my experience." But this one isn't just my taste, and the numbers are worth putting on the table.

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 — driven by escalating costs, unclear business value, and inadequate risk controls. Notice what's not on that list: the models being bad. The projects don't die of technical failure. They die of having been the wrong shape from the start.

The same analysis names the other half of the problem: agent washing, where existing chatbots and RPA get rebranded as agents. Gartner reckons only around 130 of the thousands of vendors claiming agentic capability are the real thing.

So there's pressure from both sides. Vendors are incentivized to call everything an agent, and buyers are incentivized to be seen doing agents. Somewhere in the middle, a lead-capture problem that wanted a spreadsheet gets a six-month AI project instead — and lands in that 40%.

An engine is not the same thing as an agent

I've argued elsewhere that you should stop buying tools and start building engines, and this might sound like the opposite. It isn't. It's the same argument, and this is the part people skip.

An engine is a system with a defined flow, gates, escalation paths, and a learning loop. Nothing in that definition requires a language model. The event system was a real engine: a lead enters, what happens next is decided by the system rather than by whoever remembers, and attribution comes out the other end without anyone reconstructing it.

Zero intelligence in it. Completely deterministic. Still an engine, because the flow existed and nothing depended on someone's memory.

Reach for a model when the step genuinely needs judgment — reading an ambiguous reply, deciding what a prospect is actually asking, writing something in a voice. Those are real. But judgment is expensive, non-deterministic, and needs QA forever. If a step can be a rule, make it a rule. Save the model for the steps that are hard for a reason.

The principle

The goal was never to build an agent. It was to stop losing leads. An agent is one implementation, and it's the most expensive, most fragile, most fun one on the list — which is exactly why it's the one you have to justify hardest.

Next time something lands on your desk framed as an AI project, try to solve it with a sync and a tag first, and make the agent earn its way in. If a spreadsheet fixes it, the spreadsheet wins. Ship that, and go spend your model on a problem that actually needs one.