The Vending Machine Test

Why “Deterministic” Matters More Than “AI”

Everyone’s talking about AI.

Almost nobody is talking about the thing that actually makes AI useful inside a business.

Imagine you walk up to a vending machine.

You press B4.

It dispenses a bag of chips.

Tomorrow you press B4 again.

This time it decides you’d probably enjoy a chocolate bar instead.

That’s not intelligence.

That’s a broken vending machine.

The reason it feels broken is because vending machines are supposed to be deterministic. Given the same input, they produce the same output every time. A calculator is deterministic. Follow the same recipe with the same ingredients and you’ll get (roughly) the same cake.

AI is different.

It’s designed to be.

Ask the same LLM the same question twice and you’ll often get slightly different answers. Maybe it phrases something better. Maybe it notices a different nuance. Sometimes the second answer is even better than the first.

That’s not a flaw. It’s exactly what makes AI so good at understanding messy, human language.

It’s also why AI shouldn’t be making every business decision.

Here’s the framework I use with clients:

Use an LLM where free text exists. Use SQL and rules everywhere a decision has consequences.

Think about a typical inbound web form.

A prospect writes:

“We’re evaluating platforms for about 250 sales reps across Europe and we’d like to move quickly.”

An LLM can read that and infer things no rule engine ever could.

* Enterprise opportunity

* EMEA region

* High urgency

* Likely evaluating multiple vendors

That’s AI doing what it does best: turning ambiguity into structured information.

Now comes the next step.

Should this lead go to Enterprise Rep #4 or SMB Rep #2?

Should it enter Sequence A or Sequence B?

Should it qualify for enterprise pricing?

Should Customer Success be notified?

Those aren’t AI questions.

They’re business policy.

Those are decisions with consequences.

Those should be rules, not vibes.

The AI can absolutely recommend a route. But the actual routing should be deterministic. Auditable. Replayable. Explainable.

Because sooner or later someone is going to ask:

“Why did this lead end up with the wrong rep?”

If your answer is:

“The model thought it was the best choice…”

…you’re already in trouble.

If your answer is:

“Our routing rules assign any opportunity over $100K in expected ARR from EMEA to the Enterprise team…”

…everyone immediately understands what happened.

That’s the difference between intelligence and accountability.

The mistake I see over and over is companies deploying AI end-to-end. Messy input goes in. Business decision comes out. No explicit operating logic in between.

Everything works beautifully until something unexpected happens.

Then nobody can explain why.

The design pattern that consistently works is much simpler.

Use AI wherever humans create ambiguity.

Use deterministic systems wherever your business requires accountability.

Let AI read emails.

Let AI summarize calls.

Let AI draft first-pass responses.

Then let deterministic systems decide routing, prioritization, approvals, pricing, discounts, and escalations.

You get the best of both worlds: AI’s ability to understand the messy reality of human communication, and the predictability that businesses need to operate with confidence.

This is one of the reasons I named my consulting practice Helmur.

A helm doesn’t eliminate uncertainty.

It helps you navigate it.

AI changes the tools we use.

It doesn’t change our responsibility to steer the business.

You’re going to hear a lot of AI vendor pitches this year.

Before you buy into any of them, ask one simple question:

Which parts of my workflow are being entrusted to AI, and which parts remain deterministic?

If the answer is “everything is AI,” I’d keep looking.