Agents Are Stupid. That's Why They Work.

I tell people this all the time, and it always lands strangely the first time.

Here I am, CEO of an AI company, telling you agents are stupid. But stay with me — because understanding exactly how and why agents are stupid is what separates the enterprise AI programs that are actually in production from the ones that got canceled.

This VC Has a Theory

I was talking to a venture capitalist in Silicon Valley a few weeks ago. He said something that stuck with me.

There are two types of AI companies, he told me. Model companies, and companies that burn tokens for the model companies.

He's right. Most of the AI vendor landscape right now is optimized to maximize tokens spent. That's how they make money. The more your agents reason, retry, reflect, and call tools, the more they charge. An agent that loops through five reasoning steps to answer a question a deterministic rule could have answered in milliseconds isn't intelligent. It's expensive.

We built Elementum specifically to be in a third category. A platform where the goal is to get you the right answer as efficiently as possible — not to run up your inference bill.

That required us to be honest about something the industry doesn't want to say out loud: agents are not good at everything.

What Agents Are Actually Good At

Agents are strong in a narrow band. They're good at interpreting ambiguous inputs. Reading unstructured documents. Handling exceptions that don't fit a predefined rule. Classification. Extraction. The long tail of weird cases that your fixed logic doesn't cover cleanly.

That's real value. We use agents for exactly those things.

But here's what agents are bad at: running a goods receipt. Processing an invoice. Routing a service request based on defined SLA thresholds. Anything where the answer should be the same every time, because the logic was already worked out by smart people before you ever deployed an AI.

When you process a purchase order with a probabilistic model, you're not getting intelligence. You're getting variance. And variance in a procurement workflow isn't a feature — it's a compliance problem.

The combo that works: identify intent with an agent, then execute with deterministic logic behind it. That's the architecture. It's not glamorous, but it's the one that survives audit.

Why the Legacy Systems Are Losing

We've been replacing enterprise software for two years now — the dominant ITSM platform, the procurement solution everybody bought five years ago, the HR system your team loves to hate, the CRM that costs more per seat every renewal cycle. And I'll tell you what we've found every time we go in.

The deterministic logic inside these systems? A lot of it is actually good. The rules for how a purchase order gets approved, how an incident gets routed, how a time-off request gets processed — that logic was built by people who understood the business. It works.

What doesn't work is everything around it. The user experience is terrible. The per-seat licensing model makes no sense in a world where AI handles most of the volume. And the data is locked in their warehouse, not yours.

That's not a partnership. That's a prison.

The difference between their prison and ours: we don't have a data store. When you deploy Elementum, you point it at your infrastructure — AWS, Snowflake, Databricks, wherever your data already lives. Your data center of gravity increases. You own the data. The day you decide to turn us off, you can.

I say this to every customer: the goal isn't to swap one set of handcuffs for another. The goal is to put you in control of your own fate.

The Transition Isn't a Rip-and-Replace

Here's how it actually works in practice, because I think a lot of enterprise leaders hear "replace the dominant ITSM platform" and picture a giant risky cutover. You don't do it that way.

You sit on top of your existing system first. We put our experience layer in front of whichever legacy platform you're running. Users interact with Elementum. The legacy system becomes a data repository in the background. You build confidence, understand your workflows, and then — gradually — you turn the underlying system off.

At that point you ask: do I actually need that system anymore? Or do I just need the data that used to live there?

Most of the time, the answer is: you just need the data. And once it's in your own warehouse, you're free.

I always tell customers: you have 18 months. Not because of us — because the contracts you signed with these legacy vendors are coming up for renewal, and by then the world will have moved. The teams that start now are the ones who get to decide their own fate. The teams that wait discover their options are narrowing.

One More Thing About Agents

The reason the hybrid model works — deterministic orchestration with agents inside it, not agents as the orchestration layer — is that it keeps control where it belongs. Outside the model.

An agent can hallucinate. A rule can't. An agent can drift. A rule doesn't. When you need to explain to your board or your auditor why a particular decision was made, you can't point to a probability distribution. You need a traceable path through logic that someone approved.

That's not an argument against AI. It's an argument for knowing what AI is actually for.

Agents are stupid in exactly the right way: they're good at the ambiguous, the unstructured, the exceptional. Everything else should be deterministic. The programs that figure this out early compound their advantage fast. The ones that bet the whole process on agent autonomy are the ones Gartner is tracking toward cancellation.

We built our platform on this premise from day one. Agents reason. Engines govern.

If you want to see what that looks like on a real workflow — one of your actual processes, not a demo scenario — come talk to us. We can usually show you something meaningful in 90 days.

Nader Mikhail is the CEO and co-founder of Elementum, the open orchestration platform for enterprise AI workflows.