AI in CRM is not a block you introduce or don’t introduce. It is a kit of parts at very different levels of maturity. Some have been reliably earning money for years. Others are impressive demos - and have been for just as many years.
To sort the parts, a simple double question suffices in practice: how well can the result be measured? And how expensive is a mistake? Where both come out favorably (measurable, mistakes cheap), AI carries its weight today. Where the mistake is expensive and success hard to measure, you are buying a bet. Bets are not forbidden. You should just know you are placing one - and not write it into the roadmap your revenue hangs on.
When I sort the kit today, it looks roughly like this.
The most mature part is scoring. Propensity and churn models have run reliably for years; they are measurable, correctable, scalable. The only reliable way to ruin them is to roll them out without a control group. Let a churn model compete against the existing rule logic, and often only the direct comparison shows whether it really catches noticeably more at-risk contracts early enough. Without that comparison, everyone ends up believing it works.
Next best action is the next step up: the idea of suggesting the most fitting action to every customer in real time. Strong, but fragile. It works as long as the action space stays manageable and every single action makes sense on its own. It tips over the moment the model has to choose between hundreds of options nobody curated. So my advice is to start with five good actions rather than five hundred mediocre ones.
On text generation I would draw a clear line. As a drafting machine it is excellent - subject lines, service copy, product descriptions in seconds. As the final authority it is risky, because one plausible-sounding wrong sentence ends up costing more than a slower approval. The gain is in the speed of the draft, not in dropping the review. That gets confused a lot.
And then there is fully autonomous journey steering, the systems that pick strategies on their own. Honestly, I am cautious there. Not because it never works, but because the failure case cannot be cleanly contained. If you want to try it, do it as a deliberate experiment with a hard budget cap. Not as a load-bearing pillar of a plan your revenue hangs on.
What sits above all of it is not glamorous: the value of AI in CRM today comes less from the next model than from the discipline with which you measure what it actually changes. Without that discipline, even the best model ends up disappointing.