AI or Automation? A Two-Question Test for Any Vendor Pitch

Everything is "AI" now

Open any CRM vendor's pricing page. The workflow builder that has existed for a decade now has "AI" in front of it. The same is true of the dedupe tool, the lead router, and the email sequencer. Some of that is real. Most of it is a rename.

That matters to you for one reason. AI and automation fail in different ways, cost different amounts to run, and need different things from your data. If you can't tell which one you're buying, you can't plan for what happens when it breaks.

We use two questions before every build. They work just as well on a sales pitch.

Question 1: Where does the AI do the work?

Ask the vendor to point at the exact step where a model reads something and makes a judgment. Not "the platform uses AI." The step.

AI earns its place when the input is messy and the decision needs interpretation:

• Reading a voicemail transcript and deciding whether it's a new lead, a service request, or spam
• Transcribing a lab PDF into named markers and values without interpreting them
• Deciding whether a support question can be answered publicly or needs a verified dealer
• Classifying a support question so it lands with the right team
• Drafting a first reply that a person reviews before it goes out

Every one of those is judgment under uncertainty. The model might be wrong, so the build has to say what happens when it is. A confidence score. A threshold below which a person looks. A log of what the model saw and what it decided.

If the vendor can't point at the step, there's no AI in the product. That's fine. Just don't pay AI prices for it.

Question 2: Where do rules do the work?

Now ask the opposite. Which steps have to be right every single time?

• Syncing an order from your store into the CRM
• Firing a task when a deal moves to a new stage
• Retrying a failed job and keeping the ones that fail for good
• Calculating a commission from a closed invoice
• Blocking a message to a number on the do-not-call list

None of that needs a model. It needs a rule, a trigger, and a retry. A rule is deterministic. Given the same input, it gives the same output, forever. That's exactly what you want for money, compliance, and record integrity.

Vendors who put AI on these steps are selling you variance you don't want. A model that computes commissions "intelligently" is a model that will compute them differently on Tuesday.

What the answers tell you

Put the two lists side by side and the shape of the product becomes clear.

All AI, no rules. The demo is impressive and nothing is guaranteed. Ask what writes to your CRM and what stops a bad guess from landing there. If the answer is "the model," walk.

All rules, called AI. A workflow tool with a new name. Useful, probably. Priced wrong, definitely. Ask for the automation tier instead. Sometimes all rules is the right answer: a direct-to-consumer brand we work with runs 29 automations at 150 to 190 runs a day with no AI in the production system, because every event had a deterministic destination. We said so, and built it that way.

AI reads, rules write. This is the pattern that ships. The model interprets the messy input. Rules decide whether the result is trustworthy enough to record, and where it goes if it isn't. Every agent we've built looks like this. In a wellness-plan platform we built, the model transcribes lab PDFs and drafts the plan, and a deterministic compliance scrubber with a golden test set gets the last word before anything reaches a clinician. The support bot we audit for a battery manufacturer answers product questions, and a rule stops dealer pricing from ever reaching the public. Pricing leaks after hardening: zero.

Three follow-up questions worth asking

Once you know where the AI sits, three more questions tell you whether it's built to last.

1. What's the confidence threshold, and who sees the cases below it? If there's no threshold, every guess goes through.
2. Can I see what the model saw? A decision without a log can't be audited or corrected.
3. What's the rollback? If the model writes a thousand bad records overnight, how do you undo it?

A vendor who has real AI in the product will have crisp answers. A vendor who renamed a workflow builder will change the subject.

Run the same test on the consultant

The test isn't only for software. Anyone pitching to build AI for you should answer the two questions about their own proposal, in a sentence, before the price comes up. "The model reads the email. A rule writes the record. A person clicks when confidence is low." If the answer is a slide about outcomes, you're buying a demo.

That includes us. It's why every use case and case study on this site carries two rows: where AI does the work and where rules do the work. Buyers can't tell the two apart from the outside, and the difference decides what breaks and how you fix it. It also keeps us honest. If a problem is a rules problem, we'll tell you a workflow rule is what you need. AI is a tool, not a tier.

If you're choosing an implementer, the guide to choosing a consultant has the rest of the questions.

Keep the test

Next time you're holding a pitch, run the two questions against it and see which of the three shapes you get. If you want to see what the split looks like on real builds, the use case library has a dozen, each labeled with both rows.