AI CRM: what it really does and how to choose one for your SME

Blai Lloveras Blai Lloveras
Updated:
un equipo comercial de pyme elige su CRM con IA

Almost every CRM on the market now puts the word AI on its homepage, and almost none explains which part of that AI will take work off your sales team next week and which part needs years of data to work. For an SME where the selling is done by the managing director and one salesperson, that difference decides whether the expensive plan pays for itself or is just one more subscription.

Which CRM AI features save work and which are marketing

A salesperson coming out of a meeting has three things to do before moving on to the next one: write down what was discussed, log it in the customer record and send the follow-up email. It’s exactly the work nobody does well at the end of the day, and it’s where a CRM’s AI makes a difference first.

The other half of the catalog, the part that promises to guess which customer will buy and how much you’ll bill next quarter, runs on statistics over your own history. With little data, it guesses wrong.

The ones that save time from day one

These features work on text that already exists (an email thread, a call transcript or the notes in a record) and turn it into a shorter text or a draft. They need no history, no volume and no months of ramp-up. The day you switch them on, they’re already summarizing.

In practice, they look like this. HubSpot lets you ask its assistant, Breeze, for a summary of a contact, company or deal record from the record itself, and according to its documentation this is available on all plans as long as assistant access is enabled in the AI settings. Pipedrive summarizes email threads and includes in the summary the tone of the conversation, a score from 1 to 10 on how ready the contact is to buy, and suggested next steps. Zoho CRM transcribes call recordings and produces summaries from those transcripts.

The follow-up draft is the one that gives back the most hours, because the post-meeting email is the one that gets delayed the most. What you should check before paying is whether the draft comes out with what was actually discussed in that meeting or with a generic template with the customer’s name swapped in, which any email template already does.

Automatic logging, where the call or email gets recorded in the customer record without anyone copying it over, has a limit vendors rarely mention. It only logs what goes through channels connected to the CRM. If half your sales are closed over the salesperson’s personal WhatsApp, that half doesn’t exist for the AI.

The ones that need a lot of sales history to get it right

Predictive scoring gives each contact a probability of buying by comparing them with those who bought before. Zoho calls it Zia scores and also offers to predict, from a field in the record, whether a deal will be won or lost. Sales forecasting does the same at the scale of the whole pipeline.

The problem is arithmetic. A model that learns from your past sales needs plenty of past sales, both won and lost, properly logged and with fields filled in consistently. An SME that closes a few dozen deals a year and until now kept part of its sales in a spreadsheet doesn’t give it enough material, and the number that comes out looks precise without being so.

That doesn’t mean they’re never useful. They are when there’s volume, when the sales cycle repeats in a similar way and when the team has been logging everything in the CRM for a while. If that’s not you, don’t pay for the plan upgrade because of them, and come back to them once you’ve built up the history.

the follow-up work that CRM AI takes off the salesperson

How to choose an AI CRM for an SME

Comparison tables put a green tick next to every feature, and all the big CRMs end up with a full column. What the table doesn’t show is whether those features will get any data at your company, in your language and through your channels. That’s where they part ways.

Make sure the AI works in Spanish and on your channels, WhatsApp included

A summary that comes out in English from a call in Spanish, or a draft in a tone none of your customers would recognize, gets corrected by hand, and then the savings disappear. The test is simple and you can run it in the demo or the free trial, with a real customer email, a real call and a request for a summary and a draft reply. If you have to rewrite almost the whole draft, it’s not worth it.

The channel matters even more. Many SMEs in Spain sell over WhatsApp and by phone as much as by email, and a good share of CRM AI features work on email and meetings. Ask which channels feed into the customer record, whether WhatsApp comes in through the official WhatsApp Business API or a third-party extension, and whether the AI reads those conversations or just stores them. A CRM with better AI that can’t see your main channel summarizes the part of the sale that matters least.

Native integrations versus connections through Zapier or Make

A CRM’s AI is only as good as what it receives. If email and calendar are connected natively, every meeting and every customer reply land in the record on their own, and the summaries and drafts have something to work from.

When that connection doesn’t exist and you have to build it with Zapier or Make, it works, but the division of labor changes. Someone has to set up each flow, make sure it doesn’t break when the vendor changes something, and pay for the runs separately. For a simple connection, like sending website forms to the CRM, that’s reasonable. For the whole team’s email, which is the flow all of the CRM’s AI depends on, our rule is that it must be native, and if it isn’t, we’d rule out that CRM even if it wins on the table.

A practical test for the demo is to ask them to show you a Google Calendar or Outlook meeting showing up in the record, with the follow-up email next to it and the follow-up draft generated from both. If they need to configure something in front of you to show it, you know what you’re in for.

Top AI CRMs for SMEs and what each one offers

The four names that show up in almost every comparison are HubSpot, Zoho CRM, Pipedrive and Salesforce. What follows was checked against each vendor’s documentation in November 2026, and plans and prices change often, so check the pricing page on the day you decide.

HubSpot has a free plan with contacts, companies, deals and pipeline, and its Breeze assistant can summarize records on all plans. The more ambitious features are sold separately. Its prospecting agent prepares follow-up drafts after each meeting and suggests record updates for the salesperson to approve, and it’s part of Agent Hub, which HubSpot includes in the Starter, Professional and Enterprise editions. Agents you build yourself with its agent builder consume HubSpot credits, a variable cost worth estimating up front.

Zoho CRM brands its AI as Zia and is the most prediction-oriented of the four: contact scoring by likelihood to convert, prediction of whether a deal will be won, anomaly detection when something breaks from the trend, and a suggestion of when and through which channel to contact each customer. It also summarizes call transcripts. The Zia page doesn’t say which editions include it, so ask before assuming it comes with yours.

Pipedrive focuses on email. It has a sales email generator that works from instructions and an email thread summary that, according to its help center, is in beta and only on Premium plans and above. Its sales assistant analyzes your data, flags the probability of winning each deal and recommends the next step.

Salesforce includes in its small business packages, Starter Suite and Pro Suite, record summaries, email writing with the Draft with AI option and an assistant called Employee Agent, at no extra cost according to the vendor itself. Anything not on that list, such as building your own agents, needs to be checked for the specific edition before you count on it.

Our rule for choosing between them is to look at the plan first, not the feature. If what you want is to save time on logging and follow-up from the first month, all four cover the basics and the difference lies in which plan they do it on. If you’re interested in prediction and you have history, Zoho is the one that develops it furthest. And if your team lives in email, take a careful look at Pipedrive, knowing that the AI you want is on the higher plan.

How to build your own AI feature on top of your CRM with n8n or Make

There’s a third option the comparisons don’t mention. If you already have a CRM that works and all you’re missing is the summary and the follow-up draft, you can build it on top with an automation tool and a language model, without switching CRMs or upgrading your plan.

It’s the route we prefer when the vendor’s AI sits on a plan that isn’t worth it for you, when it doesn’t work well in Spanish, or when the conversation happens on a channel the CRM doesn’t read. And we advise against it if nobody at the company is going to look after the flow when it fails, because at some point it will.

There are three pieces: an automation tool, n8n or Make, which has nodes for the usual CRMs such as HubSpot, Pipedrive, Zoho CRM or Salesforce; a language model with an API, Claude, ChatGPT or Gemini; and a place in the CRM to put the result, usually a note in the record and a task for the salesperson. Building a first working version takes someone who already knows the tool roughly one or two days, and quite a bit longer the first time.

The trigger and the model call

The flow starts with a CRM event: a meeting marked as finished, a logged call or an incoming email linked to a contact. In n8n that’s a CRM trigger node or a webhook; in Make, a watch module or a webhook.

With the event in hand, the next step gathers the context: the transcript or meeting notes, the latest emails with that customer, the deal stage and whatever is in the record. All of it goes to the model in a single call.

The prompt is where you win or lose the team’s trust. It has to tell the model to use only what’s in the context, not to include dates, prices or commitments that don’t appear in the notes, and, if a piece of information is missing, to mark it as pending instead of filling it in. A draft that promises the customer a discount nobody offered is worse than no draft at all, and the salesperson stops using it by the second time.

the salesperson reviews the follow-up draft before sending it

An output the salesperson reviews before sending

The flow doesn’t send anything. It writes two things to the CRM: a note with the summary of the conversation, what was agreed and the next steps, and a draft follow-up email, which can be left as a task assigned to the salesperson or as a draft in their inbox.

The salesperson opens it, fixes whatever needs fixing and sends it themselves. That review is what makes the system usable with real customers, and also what tells you whether it works, because if every draft has to be rewritten from scratch, the problem is in the prompt or in the context it receives, and that’s where you fix it.

It’s a good idea to store in the note itself which data the model used, for example the source emails and meeting, so that whoever reads it three months from now knows where everything comes from. With that, you have the feature CRMs sell on their higher plans, in your language, on your channels and with the salesperson deciding what goes out to the customer.

If you’d like to see how this flow would look with your CRM and your channels, get in touch and we’ll look at it with you.

About the author

Blai Lloveras

Articles by Blai

Let's talk

Do you have a project in mind?

Tell us about it and let's grow your business together.

Analyze my sales opportunities