AI, Actually

How to Tell Whether a Product Really Uses AI

Published November 13, 2026 — 4 min read

TL;DR: Most "AI-powered" labels describe a rules engine with a new coat of paint, and three plain questions will tell you which one you're buying. You don't need a technical background to ask them.

Key Insight

A rules engine is not a bad product. If your invoices always follow the same layout, a rule that reads the third column is cheaper, faster, and easier to audit than any model. The problem starts when a vendor charges AI prices for it, or when you plan around capabilities the product doesn't have.

The difference comes down to who wrote the logic. In a rules engine, a person wrote every branch ahead of time: if the subject line contains "refund," route to billing. In an AI product, the system learned a pattern from examples or reasons over text it has never seen, so it can handle input nobody anticipated. It can also get that input wrong, which is why the two behave differently when things go sideways.

Regulators have noticed the gap. In March 2024 the SEC charged two investment advisers over claims about their AI use, with $400,000 in combined penalties. The FTC followed in September 2024 with Operation AI Comply, a sweep of companies making deceptive AI claims. Both cases were about the label getting ahead of the product.

Some of the better-known failures were quieter. Presto Automation sold AI drive-thru ordering while offshore staff handled a large share of the orders. Amazon's Just Walk Out stores reportedly leaned on human reviewers in India to check shopping sessions. Neither is a scam in the cartoon sense. Both are what happens when "AI" describes the ambition and not the system.

Why Teams Miss This

Buyers ask "does it use AI?" and vendors always say yes. It's a question with one possible answer, so it tells you nothing.

The second mistake is judging by the demo. A demo runs on inputs the vendor picked, and rules engines look great on inputs someone planned for. A scripted flow and a real model look identical when the script covers the demo.

The third is treating this as a technical problem you have to hand off. It isn't. The tells show up in behavior, and you can watch behavior without reading any code.

How to Actually Do It

Ask three questions in your next vendor call. None of them needs a data scientist in the room.

1. What happens when I give it something it has never seen?

Bring your own messy input: a badly scanned form, an email in a different language, a request phrased sideways. A rules engine fails in a predictable way. It returns nothing, routes to "other," or asks you to rephrase using a keyword. A real model gives a reasonable answer, and sometimes a confidently wrong one. Both outcomes tell you something. If the vendor won't let you try your own input live, that tells you something too.

2. Does it get better when you correct it, and how?

Ask what physically happens when a user fixes a wrong answer. With a rules engine, someone files a ticket and an engineer edits a rule. With a learning system, there's a feedback path: corrections get logged, evaluated, and either retrained on or used to improve prompts. Ask for the last three improvements the product shipped and what caused each one. "We added a rule for that" is an honest answer. Just know what you're paying for.

3. Which model is in it, and what does it do when it's unsure?

Skip the architecture questions. You need a name or a category ("a hosted large language model," "a classifier we trained on 40,000 of your category's tickets") and a plain account of the uncertain case. Real AI products have a confidence behavior: they hand off to a person, flag the result, or show a score. A system that is never unsure is either very narrow or not doing what you think.

If you want a hard check, run a small test yourself. Take 20 real inputs from your own work, 15 ordinary and 5 odd, and record what the product does with each:

# inputs.csv: id,input,expected
# results.csv: id,output,handled_gracefully (y/n)
# Score the odd cases separately. A rules engine often matches
# on the 15 ordinary rows and collapses on the 5 odd ones.
python3 score.py inputs.csv results.csv

The gap between the two groups is the signal. A rules engine posts near-perfect scores on the ordinary rows and a cliff on the odd ones. A model usually degrades more evenly.

What We've Learned

The honest answer is often "both." The best production systems we see put rules around a model: rules handle the cases that must be exactly right, like totals, permissions, and compliance wording, and the model handles the fuzzy part, like reading intent. A vendor who can explain where their rules end and their model begins usually knows what they've built.

Your next experiment is small. Pick one product on your team's "AI" list and run the 20-input test this week. If it passes, you have evidence for the renewal. If it doesn't, you have a specific question for the vendor, and a better price conversation.

Sources

Have a specific workflow in mind?

Bring it to a Quick Scan — a live working session where we'll tell you honestly whether it should be an agent, a workflow, or left alone, before you spend a dollar building it. You get 3 prioritized recommendations on the call, a one-page summary after, and the $500 credited toward any engagement within 30 days.

Tell us what you'd hand off →

See AI Agent Consulting·Book an intro call

Get new posts + practical agent-ops notes

One email when something new goes up. No nurture sequence, no spam — unsubscribe whenever you want.

Thanks — you're on the list.