AI Agents are Dogs

Every vendor selling AI agents right now is selling you the same picture.

Across markets and industries, the framing is: bring in AI, manage AI, scale your AI workers.

I think that’s the wrong picture.

Because if you build your AI strategy on the mental model of having AI agents become drop-in replacements for human workers, you'll make bad investment decisions. You'll end up with AI agents doing stupid things, generating bad outcomes, and having people finger-pointing each other. Most likely, you'll never deploy an AI agent at all.

The better picture in my opinion:

AI agents are Dogs.

Let me explain.

Calling AI agents "digital employees" sounds reasonable at first until you realize that the entire management toolkit does not transfer:

AI Agents can't be held accountable for "giving a customer the wrong information", or "getting a report number wrong". It’s literally just a dumb machine that appears to be pretty smart.

So what else? Robots?

That's the other extreme, and it leads people in a bad direction either way.

Because when most people hear "robot," they picture something like this:

Fully deterministic machines that, given a specific task, behave extremely predictably. Same input, same output. A well-oiled machine. A powerful "tool".

That's not what modern LLM-based agents are.

These "machines" can make different judgment calls even when looking at the same data. They don't fail at edge cases, they solve them in unexpected ways. They confidently report processing 500 invoices when they skipped 50. Not very robot-like.

Both pictures miss what an agent actually is: a bounded-capability tool that operates with some autonomy, requires real ownership, and fails in specific ways you have to plan for.

That's a dog.

I know this article can be difficult for some dog owners (personally, I don't have one). So if you're too emotionally attached, just imagine someone else’s dog.

But dogs have 4 traits that are surprisingly helpful when thinking about AI agents.

If your dog bites someone, you get sued. Not the dog. Not the breeder. Not the trainer. You.

In most jurisdictions, dog ownership comes with strict liability. The owner is on the hook regardless of intent, regardless of whether the dog had a history of biting, regardless of whether you were even there. The law assumes that if you decided to keep this dog, you took responsibility for what it does.

That's the right model for agents.

When an agent makes a wrong call, the deploying organization owns the outcome. In practice, this means every agent in your organization needs a named owner. Not "IT". But a specific person who is accountable for what the agent does and doesn't do. That person is responsible for making sure the agent’s outputs get regularly reviewed, errors get corrected, and it gets shut down when it stops earning its keep. It’s also the person that gets the credit when the agent works. (Which is probably the most important part).

In other words: If you can't put a name on an agent, you don't have an agent in production.

The capabilities of dogs are really admirable. A Pyrenean Shepherd can watch sheep on a mountain alone for days. But when you throw a ball, this dog will just look at you: "Well, that was stupid – now you have to go and get it".

The capabilities of AI agents are jagged in a very similar way.

When an AI agent fails at a given task, it’s rarely because the AI model is "too dumb". It typically fails because the agentic system was not designed for ("trained on") the given situation. A research agent optimized for financial data performs poorly when it gets pointed at legal documents.

So don't ask whether the agent is "good".

Ask:

Vendors will sell you an "agent that does X" and let you figure out for yourself whether your X matches their X. They're rarely the same.

The work of an agent owner is not just monitoring its behavior. It's understanding the agent's actual scope of capability and refusing deployments outside it.

You could trust a dog watching the sheep. But you wouldn’t trust it alone with a toddler.

Technically, the dog would be capable of watching the toddler. Still, you wouldn’t let it. The stakes are just too high.

I trust an invoice-processing agent to read invoices. I do not trust the same agent to write invoices, even if the underlying system could actually do it. Capability isn't trust. Trust comes from having seen the system handle the actual inputs it will face, repeatedly, with predictable behavior. The further you move from that tested scope, and the higher the risk of failure, the less you should trust the deployment.

A shepherd might have two, three, maybe four dogs. It's hard to imagine them sending out 100 dogs to watch 10,000 sheep.

But that is precisely what AI Agent vendors are pitching. "Build your AI workforce at scale."

Just imagine watching 10 agents every day. Monitoring their outputs, making corrections, responding to feedback in real time. Sounds exhausting? It is. (Maximum I tried was 5.)

There's a real ceiling on how many agents one human can meaningfully own. A handful is realistic for most things. Maybe more if the agents are well-instrumented and the domain is low-stakes. Beyond that, you're not owning agents – you're hoping.

And hope is not an operating model.

The dog characteristic changes the math on "AI replaces your workforce" in a significant way. You don't replace 100 people with 100 agents and call it a day.

You simply give humans a higher leverage.

Think of drug-sniffing dogs. There's still a human in charge but that human now got superpowers because they’re able to scan 20x as many passengers without even touching them or sending them through complicated equipment.

This is not the typical AI vendor pitch you’re hearing. It's also not the doomer pitch where AI takes all the jobs. The org chart doesn't get new boxes for agents. The boxes that exist get redrawn around expanded scope. If they don't, you'll hardly see any improved outcomes.

Like any metaphor, the dog analogy of course has its limits. Unlike a dog, an agent can work 24/7. You can clone a working agent the way you can't clone a working dog. You can retrain an agent continuously across all its copies. But if you treat your agents like working dogs instead of digital humans, you'll get the deployment decisions directionally right.

You'll know who owns what, where to put them, where not to put them, and how many you can realistically run.

Keep the leash short at the beginning.

See you next Saturday!

Tobias

PS: The four organizational shifts that come with agent deployment are in Chapter 10 of The Profitable AI Advantage — free preview of Chapters 1-2 here.

To grab the full copy: