The Agent Is Only as Good as the Process You Wrote Down

We went back to basics with a company earlier this year. Pre-seed, strong people, real potential, no operational foundation.

No clear funnel. No definition of what a lead is. No alignment between marketing and sales. No visibility anyone could rely on.

So we stopped and rebuilt the order. Define the flow — what is a lead, when does it qualify, what happens next. Put everything in one system, no side spreadsheets. Create clear ownership so every lead has a path. Only then, add automation.

Nothing fancy. Just the right foundation.

I am telling that story again because the same mistake is about to get much more expensive.

Building agents got easy

At UNBOUND, HubSpot talked about Agent Builder and the custom agents teams are building on it, across deal management, RevOps workflows, prospecting, support and internal automation. No engineering required.

This is genuinely good. The gap between "we should automate this" and "we automated this" used to be a development project. Now it is an afternoon.

And that is exactly where the problem starts.

What an agent actually is

An agent is a written process that executes itself.

That is the whole thing. Every agent you build is a specification: when this happens, check that, decide between these options, take this action, hand off to this person.

Which means the question was never "can we build it."

The question is whether anyone has written down what it should do.

In most companies, nobody has. The process lives in the head of the person who has been doing it for three years, and it includes forty judgment calls they make without noticing. Ask them to explain it and you get the clean version, not the real one.

You cannot build an agent from the clean version. You will get an agent that works on the easy cases and does something confident and wrong on the rest.

What we see working

The companies getting value out of custom agents are not the ones with the best prompts.

They are the ones who already did the boring work. They know what qualified means, in writing, and everyone agrees. They know who owns each stage. They have a process that a new hire could follow without asking anyone.

Those companies build an agent in an afternoon and it works, because they are not designing the process while they build it. They are encoding one that already exists.

The companies without that spend three weeks building, discover their team disagrees about what step four is, and conclude the technology is not ready.

The technology was ready. The specification was not.

Where to start

• Pick the task your team does the same way every time and hates. Not the most strategic one. The repetitive one with a clear trigger and a clear output.

• Write it down as steps, with a real person. The one who actually does it. Ask what they do when something is unusual — that answer is the part that matters.

• Mark the judgment calls. Every point where a person decides rather than follows a rule. Those stay human, at least to start.

• Measure the before. Time taken, volume handled, error rate. Something.

• Then build it, narrow. One task, one trigger, one output, one owner.

The order is the same as it has always been. AI did not change it. It just made the cost of skipping it higher, because now you scale the mistake instead of just making it.

Questions we get asked about this