"We use AI."
You've heard it from your CRM vendor. Your accounting software. Your email marketing platform. The company that handles your payroll. The startup pitching you on a new document management system. At this point, the phrase has become so ubiquitous that it tells you approximately nothing.
Here's the thing: they're probably not lying. Most of these tools do use some form of artificial intelligence. But "we use AI" is like a restaurant saying "we use heat." Technically true. Completely unhelpful for deciding whether to eat there.
The good news? One question cuts through all of it: What does the system actually produce?
The two circles that matter (and the marketing fluff around them)
Picture two circles, one inside the other. The bigger circle is machine learning — systems that learn patterns from data rather than following rules someone wrote. Inside that sits generative AI, the systems that create new content.
Every generative AI system is machine learning. But not every machine learning system is generative AI.
Why does this matter to you? Because the layer you're dealing with determines what data the system needs, how it fails, and how you check its work. Get the layer wrong, and you'll ask the wrong questions, set the wrong expectations, and probably waste money.
The one-question test
Ignore the pitch deck. Ignore the marketing copy. Ask: What does this system actually produce?
If the answer is a score, a label, or a ranking — that's machine learning. The system is looking at something and telling you what category it belongs to or where it falls on a scale.
If the answer is something new — text, an image, a video, a summary that didn't exist a second ago — that's generative AI. The system is creating content.
That's it. That's the test.
When a vendor says "AI-powered," ask them to show you the actual output. The format of that output tells you which layer you're dealing with.
Machine learning in your business: scores, sorts, and flags
Machine learning excels at one thing: looking at data and making a judgment about it. The output is always a label, a score, or a ranking.
Churn scores. Your customer success platform looks at usage patterns, support tickets, billing history, and contract terms. It produces a number: this customer has a 73% likelihood of not renewing. That's machine learning. It's not writing the save-the-customer email — it's just telling you who to call first.
Fraud flags. Your payment processor examines a transaction — amount, location, time of day, device fingerprint, purchase history. It produces a label: legitimate, suspicious, or blocked. Machine learning. It's not explaining why it's suspicious in a paragraph. It's just raising a flag.
Lead scores. Your marketing automation tool looks at website behavior, email engagement, company size, and job title. It produces a ranking: these are your hottest prospects. Machine learning again.
Document sorting. An incoming invoice gets routed to accounts payable. A support request gets routed to the right team. A contract gets tagged by type. The system is reading the document and assigning it to a category. Still machine learning.
In every case, the system is taking something that exists and telling you something about it. The output is a judgment, not a creation.
Generative AI: creating something new
Generative AI produces content that didn't exist until the system created it.
A drafted customer email. You give the system context about a customer issue, and it writes a response your support rep can review and send. That's generative AI. The email didn't exist before.
A product description. You feed in specs and features, and the system writes marketing copy. Generative AI.
A meeting summary. Your AI meeting assistant listens to a call and produces a written recap with action items. The summary is new content. Generative AI.
An image for a presentation. You describe what you need, and the system creates a visual. Generative AI.
The pattern: the system takes input and produces something new — text, images, audio, video, code. Content that didn't exist a second ago.
The trap: the verb isn't the output
Here's where vendors get slippery.
A recommendation engine that "generates recommendations" sounds like generative AI. But look at what it actually produces: a ranked list of products you already sell. It's not creating new products. It's sorting existing ones by likelihood of purchase. That's machine learning.
An email tool that "generates personalized subject lines" might be generative AI (if it's writing new subject lines) or machine learning (if it's selecting from pre-written options based on what performs best for similar audiences).
A chatbot that "generates responses" could be generative AI (writing novel answers) or a rules-based system (selecting from a decision tree someone programmed).
Read what comes out, not the marketing copy. If the output is new content, it's generative AI. If the output is a selection, ranking, or score applied to things that already exist, it's machine learning.
What about rules-based systems?
Here's where the "AI" label gets stretched to the breaking point.
Your if-this-then-that automations? Not AI. Just automation.
A chatbot that follows a decision tree with predetermined responses? Not AI. It's a fancy phone tree.
A fraud detection system that flags any transaction over $10,000 from a new device? Rules-based automation. Not AI.
These systems follow logic a person wrote. They don't learn from data. They don't improve on their own. They do exactly what they're programmed to do, every time.
That doesn't make them bad — sometimes they're exactly what a problem needs. Plain automation, following rules a person designed, often beats a complex machine learning system in terms of cost, predictability, and ease of maintenance. But calling them "AI" is marketing, not accuracy.
When a vendor tells you their tool is "AI-powered" and you discover it's actually a rules engine, you haven't been scammed exactly — but you should recalibrate your expectations. Rules-based systems are predictable and explainable. They're also rigid. They can't handle situations their designers didn't anticipate.
Don't assume you need machine learning or generative AI. Sometimes a well-designed rules-based system solves the problem faster, cheaper, and more reliably. Just don't pay AI prices for automation.
Why the layer matters: different failures, different oversight
The layer you're dealing with determines how the system fails — and that determines how you need to check its work.
Machine learning fails by giving wrong scores. A churn model might say a customer is low-risk when they're actually about to leave. A fraud system might flag legitimate transactions. A lead score might rank tire-kickers above serious buyers. The failure mode is miscategorization. You catch it by sampling outputs and comparing them to reality.
Generative AI fails by creating confident, plausible-sounding wrong content. A drafted email might include details that aren't true. A product description might make claims you can't support. A meeting summary might attribute a statement to the wrong person. The failure mode is hallucination — content that sounds right but isn't. You catch it by reviewing the actual content before it goes out.
Rules-based systems fail by not handling edge cases. They do exactly what they're told, which means they break when reality doesn't match the rules. You catch it by monitoring exceptions and updating the rules.
Different layers, different oversight requirements. A machine learning system might be fine running in the background with periodic accuracy audits. A generative AI system producing customer-facing content needs human review on every output — at least until you've validated its reliability for your specific use case.
Questions to ask before you buy
When a vendor pitches you an "AI-powered" solution, here's your checklist:
What does it produce? Get them to show you actual output, not a demo. Is it a score, a label, a ranking? Or is it new content? Or is it just following rules?
What data does it learn from? Machine learning systems need training data. Generative AI systems need context. Rules-based systems need someone to write the rules. Who provides that, and how current is it?
Who checks the output? For machine learning: how do you validate accuracy? For generative AI: who reviews content before it goes to customers? For rules-based: how do you handle exceptions?
What happens when it's wrong? Every system makes mistakes. What's the process for catching errors, correcting them, and improving the system?
Can you see inside it? Can you understand why the system made a particular decision? For high-stakes applications — pricing, hiring, credit decisions — explainability isn't optional.
| System Type | Output Type | Failure Mode | Oversight Approach |
|---|---|---|---|
| Rules-based automation | Actions based on programmed logic | Misses edge cases | Monitor exceptions, update rules |
| Machine learning | Scores, labels, rankings | Wrong categorization | Sample and validate accuracy |
| Generative AI | New content (text, images, etc.) | Plausible-sounding hallucinations | Human review before release |
Where to start
Here's your homework: make a list of every tool you pay for that claims to use AI. For each one, write down what it actually produces. Score, label, or ranking? New content? Actions based on rules someone wrote?
Then do the same for any AI tools you're considering.
Once you've named what you're actually dealing with, ask yourself: does this match the problem I'm trying to solve?
If you need to prioritize a queue — which customers to call, which invoices to review, which leads to pursue — you need machine learning. Something that scores and sorts.
If you need to create content at scale — customer communications, documentation, marketing copy — you need generative AI. Something that produces new content.
If you need to route things according to rules you can define — approvals, escalations, notifications — you might just need automation. Something that follows the logic you specify. And that's fine. Don't pay for AI when automation will do.
Match the tool to the problem. That's how you cut through the noise and make decisions that actually make sense for your business.
If you're not sure which approach fits your problem, it's worth talking to someone who's done it before. The wrong choice costs more than the conversation.
Entvas Editorial Team
Helping businesses make informed decisions
