Every CRM has shipped its AI. HubSpot calls it Breeze, Salesforce calls it Agentforce, Zoho has had Zia for years and Pipedrive has its AI features scattered across the pipeline. The demos are good: the assistant summarizes the thread, drafts the follow-up email, scores the lead and tells you who to call today. If you're evaluating which one to turn on, the question they're making you answer is «which one has the better AI». It's the wrong question.
The thesis in one line: the AI your CRM is selling you is designed to solve your CRM's problem —getting you to use it more and fill it up more—, not your operation's. And your operation's problem is almost never writing emails: it's that the data is dirty, incomplete and out of date, and that nobody knows the next step on half the open deals.
What AI for CRM promises in HubSpot, Salesforce, Pipedrive and Zoho
The four platforms converge on the same thing under different names: a conversational assistant inside the record, email drafting, conversation summaries, lead scoring and —new this cycle— agents that run whole sequences without anyone pressing a button. The real difference between them isn't model capability: it's how much of your data they already hold, and how much it costs you to get out what they produce.
| CRM | What it calls its AI | Where it genuinely helps | Where it leaves you on your own |
|---|---|---|---|
| HubSpot | Breeze (assistant + agents) | Drafting and summaries inside the record; works well if your team already lives in HubSpot | Everything that happens outside HubSpot: the email nobody logged, the ERP where the billing history sits |
| Salesforce | Agentforce | Agents over complex, highly configurable processes, with serious permissions behind them | The real cost of configuration: without someone to model it, a powerful agent on badly modeled data amplifies the mess |
| Zoho | Zia | Predictive scoring, anomaly detection and best-time-to-contact suggestions at a price well below the other two | Depth once the process leaves the standard path; the ecosystem pushes you to stay inside Zoho |
| Pipedrive | AI features across the pipeline | Simplicity: scoring and suggestions without anyone having to administer anything | Anything that demands your own judgment or outside data; it's a visual pipeline, not a platform |
None of that is wrong. The thing is, all of it is designed on the same assumption: that the data inside the CRM is correct and complete. At most companies that assumption is false, and that's where the promise breaks.
The vendor's bias: your CRM solves CRM problems
Think about it from the other side of the table. Your CRM vendor measures success in adoption: records created, fields filled, active users, licenses renewed. Its AI, logically, attacks the friction that gets in the way of that —writing is work, filling in is work, looking at the dashboard is work—. It's an AI that makes using the CRM more comfortable.
Your operation measures something else: deals that move, forecasts that hold, customers that don't go cold. And the friction blocking that is almost always outside the CRM or underneath it: in the email that never got logged, in the contact duplicated three times, in the «industry» field every rep filled in their own way for four years, in the history that lives in the ERP.
The practical consequence is ugly. An assistant that writes well on an incomplete record produces a well-written, wrong email. An agent that moves stages on dirty data hands you a forecast with more decimal places and the same falsehood. AI doesn't fix the data: it inherits it and spreads it faster. It's the same reason so many AI automation projects fail: you start with the visible layer instead of the one holding it up.
The boring stuff is what moves the number
There are three things that do change the number, and none of them shows up in the demo because none of them is flashy. They look more like plumbing than artificial intelligence, and that's exactly why they work.
1. Data hygiene
Duplicates that merge on their own instead of piling up. Company names normalized so «Acme S.L.», «ACME SL» and «Acme» are one. Categorical fields filled from a closed taxonomy and not from whatever each person felt like that day. Contacts marked inactive when the email bounces three times. This is what makes the lead scoring you already paid for start getting it right: not a better model, less dirty data.
2. Enrichment
Half the records in your CRM have no company size, no real industry, no tech stack and no signal on whether they're growing. Without that, any prioritization is a hunch with an interface. Enriching means filling those fields from outside sources and from your own systems —what they already bought, what they already returned, what tickets they already opened— and keeping them fresh. It's the least glamorous part of integrating AI with your systems and the one that returns the most per dollar spent.
3. The written next step
The question leadership asks every Monday isn't «how many emails did we send?». It's «what's missing to close this and who has to move it?». That lives in the conversations, not in the fields: a commitment with a date, a specific objection, a person who decides. Pulling it out of the threads and writing it into deal properties turns the CRM into a state system instead of a filing cabinet. It's exactly what we do when we connect HubSpot with Gmail: the value isn't moving the email across, it's extracting the state out of it.
What you shouldn't automate yet
Three things the demo makes look harmless and that you pay for in production:
- Sending customer emails with nobody reading them first. Not out of fear of a typo: because of the implicit commitment. An assistant that promises a date or hints at a discount creates an expectation your operation will have to hold up.
- Moving pipeline stages automatically from day one. The stage is the data your forecast comes out of. If a model moves it with no auditable criteria, you lose the ability to defend that forecast to anyone, yourself included.
- Scoring leads with a model trained on your dirty history. It will learn your past biases and hand them back looking like science. Clean first; score after.
The pattern is always the same, and it doesn't depend on the CRM: automate first what's reversible and cheap to correct; leave for later whatever touches the customer or the forecast. It's the same logic behind which processes to automate with AI, applied to sales.
Three questions before you turn anything on
Before you buy your CRM's AI module —HubSpot's, Salesforce's, Pipedrive's or Zoho's, it hardly matters— answer these three. If you fail one, that module isn't your next purchase.
- What percentage of my records is complete on the five fields I use to decide? If you don't know, that's this week's task, not turning on an agent.
- What happens when the important data lives outside the CRM? If your billing history is in the ERP and your conversations are in email, your CRM's AI is going to weigh in having seen half the movie.
- Can I see why the system did what it did? If the answer to «why did it move this deal to negotiation» is a shrug, you're not ready to let it run loose yet.
The uncomfortable conclusion
If your data is clean, complete and connected, the native AI in any of the four will give you a good result and you don't need anything else: turn it on and save yourself the project. If it isn't, none of the four is going to save you, and buying the most expensive one just puts the mess on prettier screens.
That's why the useful question isn't «which CRM has the better AI?» but «what is my data missing so the AI I've already paid for starts getting it right?». Almost always the answer is the boring stuff: duplicates, empty fields and conversations that never made it into the record. It's less exciting than an autonomous agent. It's also the only thing that changes the number.