Stop Asking the Model to "Figure It Out"

Prompt < Code
(A statement backed by multiple, frequent tests.)

The team is complaining about the data I get from some of our Agents.
And no, I’m not kidding.
Some of them run purely on OpenAI or Anthropic models.

No custom code.
No deterministic logic.
No hard-coded rules.
Just prompts + AI.

Why?
Because I configured the process the way most people do when they use AI, with prompts... with prompt engineering even...
And it works, in many aspects... but there is always that one data point that isn’t accurate: An email gets missed, A call isn’t captured, A message disappears, A piece of information is interpreted incorrectly.
Not every time. Just often enough to have to review all the time if you want the exact number.
And that’s the problem.

Can we fix it?
Yes. And that’s exactly what we do with our clients.
When something needs to be reliable, we stop asking the model to “figure it out” and start engineering the process around it.
If a rule can be deterministic → we code it.
If a validation can be automated → we code it.
If a data source needs UI → we RPA it.
No “let’s see what the model decides today.”
Just a script doing exactly what it was designed to do.

So why don't I do this?
Because I kept hearing all the promises: “AI can do it all in 15min.”
And I had the feeling I was always missing the last model... or being an old grumpy dude.
So now, deliberately, I leave some of our internal processes fully AI-driven.
And I see them, every day... Even Astra.
And honestly, it has been incredibly useful.

Because it gives us something much more valuable than another AI demo:
First-hand experience of where AI actually breaks , and what many CFOs and finance teams are experiencing right now.

We get to see them ourselves and showcase the gap:
Prompt < Code.

Not because AI isn't powerful.
But because the best systems know when to let AI think and when not to let it think at all.

Have a great weekend!
Alexis

P.S. Here our invoicing process, triggered when the Agent understands it needs to process invoices, once that triggered, no thinking, no mistakes.

—

Originally posted on LinkedIn:

https://lnkd.in/p/d5apQEYy