Every CRM is “AI-powered” now. The trouble is that the label covers everything from a genuine language model reading your deal to a hard-coded rule that's been around since 2015, freshly rebranded. If you're evaluating a tool — or just trying to trust the one you have — it's worth knowing how to tell them apart. So this piece does two things: it gives you a simple test, and then it turns that test on us, naming exactly which of our own features are real AI and which aren't yet.
Why the distinction matters
This isn't pedantry. If a feature labeled “AI” is really seeded demo data or a lookup table, you'll make decisions on output that isn't what you think it is — and once you discover one fake, you stop trusting the features that genuinely are AI. The hype tax is paid in lost trust. The only way out is to draw the line clearly and keep it drawn, including over the parts that aren't ready.
A four-question test
You don't need to read source code to sort real from hype. Four questions get you most of the way there.
1. Does it call a model?
A real call to a language model (an LLM) runs on each use — you can see usage logged and metered.
It's if/then rules, a lookup table, or seeded demo data with no model behind it.
2. Is it grounded in your data?
It reads your actual record — the thread, the deal, the notes — and writes for that situation.
It produces generic output that ignores your data, or pre-baked text dressed up as 'AI'.
3. Does the output vary with context?
Different records produce genuinely different, situation-specific results.
The same canned phrasing appears regardless of what's on the record.
4. Is it honest about its limits?
Features that aren't a real model yet are labeled 'Soon', not sold as live AI.
Everything is 'AI-powered', including the parts that are plainly just rules.
Rule-based isn't a dirty word — but it isn't AI
Worth saying plainly: rule-based and statistical features are good. A transparent lead score you can audit is often better than a black-box model — every point traces to a rule you set. The problem isn't that a deal-risk flag uses rules; it's calling that flag “AI” when no model is involved. Keep the words honest, and rule-based logic is a feature to be proud of, not one to disguise.
Now turn the test on us
Here's where most vendors stop. We'll keep going and apply the four questions to Kudos AI, our own assistant. The honest answer is that some of it is real generative AI today, and some of it isn't yet — and we mark the difference in the product with a “Soon” pill instead of pretending.
Real AI today (Claude)
- Email assistant — compose, reply, rewrite, retone, translate, subject lines
- Record summaries — notes & history into a plain-language recap
- Document extraction into structured fields (you confirm)
- AI form builder — fields drafted from a prompt
- Subtask suggestions for a task; AI tag suggestions
- AI workflow steps — summarize / draft / classify / next-best-action
- Ops AI standup — a grounded daily digest
Not real AI yet (“Soon”)
- Conversational copilot chat panel — a demo router today, not a real model
- In-record predictive scoring — deal risk, win-probability, health (rule-based/seeded)
- Next-best-action as an in-record score (real only as a workflow step, not predictive)
- The pipeline forecast — weighted and rule-based, not an AI prediction
The pattern is consistent: the writing and extraction work — email drafting and replies, summaries, document extraction, the form builder — is where a real model genuinely earns its place, so that's what we built and what we market. The conversational chatbot and the predictive “which deal will close” scoring are harder problems we haven't shipped as real AI, so we don't claim them. That's the whole policy: market only the AI that genuinely calls a model.
What that means for you
When you use Kudos AI's email assistant or record summaries, you're using a real call to Anthropic's Claude, grounded in your data, with usage logged and a human approving every result. When you see a “Soon” pill on the chat copilot or a predictive score, that's us telling you it isn't a model yet. Both messages are deliberate — the point of the test is that you can trust the label either way.