Ask a marketing team whether they use AI in CRM and the answer is almost always yes. It tells you nothing. Copy gets written faster, images too, and somewhere a scoring model runs. The question that separates a real change from a tooling update: which decision do you now make differently?
This guide draws the one line that matters in CRM. Automating a process is craft, and most of it has been possible for years. Automating a decision is a different undertaking entirely.
The practical source is a conversation from our own podcast. Laurens Mauquoi, Chief Digital Officer of nexeye, runs marketing and product management for three brands across five countries and more than 800 stores in Europe. On the Reyo podcast 360 Marketing & CRM he described where AI in CRM really works for him and where it gets overrated. His figures are his statements, not benchmarks; quotes below are translated.
Process or decision: the only line worth drawing
Over the past decade marketing has mostly automated processes: welcome flow, abandoned cart, birthday email, win-back. Even that, according to Mauquoi, only a minority of companies have really done. The next step is not about sending. It is about choosing.
"What AI will do in future is not automate marketing processes, but marketing decisions," he says. That sentence is more than a phrase, because it sorts your project list. A flow that fires a template after an event automates a process. A system that decides, per person, which channel and which content the next contact uses automates a decision.
| Process automation | Decision automation | |
|---|---|---|
| What runs without you | The sending | The choice of channel, content and timing |
| Who decides | The team, once, when the flow is built | The system, per person, per contact |
| What you need | Events and templates | History and a forecast per person |
| How you notice | Emails go out that were never on the plan | Two identical profiles get different content |
| Typical failure | Flows nobody has reviewed in two years | Predictions nobody reads |
The two columns are not a ladder but separate building sites, and a decision automated without a working flow underneath gets executed by hand anyway. The left column is covered in the guide to marketing automation in e-commerce.
Without forecasting, AI in CRM adds very little
The most sober line of the conversation is also the most useful: "AI is almost always overrated. In the end the biggest gain in CRM is always this: do you do forecasting or not? Do you do personalisation or not?" In Mauquoi's experience AI then makes a few percentage points of difference, because it is one level more precise. The two switches before it are not gradual. They are on or off.
Good news for teams without a data science function: both switches flip without a model. Forecasting in CRM starts as arithmetic on what you already hold:
- Purchase interval per category. The median number of days between a first and a second order.
- The typical next category. What people buy after a given item.
- Overdue status. How far past their own usual interval someone sits today.
- The moment. Those three numbers give you a date and a content theme per person.
For the question of which promotion someone will buy, Mauquoi says, you need no AI at all: with hundreds of thousands to millions of known customers, the history itself shows when they come back and what they are likely to do. AI sharpens that arithmetic, it does not replace it. Thinner history means a coarser forecast and more review. The level built on this arithmetic is covered in the guide to hyper-personalisation in CRM.
The four decisions worth moving
The useful candidates occur often, cause limited damage when wrong, and show their effect in days, not quarters.
| Decision | Data it needs | What works without AI | What AI adds |
|---|---|---|---|
| Channel per person | Response behaviour, last few months | Rule: no open in 90 days, switch route | A ranked order of routes per person |
| Content per person | Purchase history, catalogue, stock | Rule: next category, exclude what was bought | The order inside that selection |
| Timing | Purchase interval per person | Arithmetic on median intervals | Correction from recent behaviour |
| Discount depth | Margin, response to earlier offers | Tiers by segment | A suggestion per person, never unapproved |
The first two rows are the ones Mauquoi names explicitly: channel choice per person, up to a letter or a phone call for people who never open email, and content per person. Timing and discount depth are the obvious continuation, and the last row is the one you keep your hand on longest, because it hits margin and the mistake only shows up in the accounts. How deep the German market goes is in the guide to discount strategy in e-commerce.
Someone has to monitor the predictions
Machine learning predictions have sat inside marketing platforms for years. What is new is that they became usable, which is why a role appears here rather than disappears.
"You still need someone who understands the whole picture and monitors the AI: these predictions make sense, and these are too wild because the data underneath them is not enough," Mauquoi says. Routine analytical work in CRM goes away. The judgement does not.
In practice this is a recurring half hour, not a new headcount. Four questions carry it.
- Is there enough behind each prediction? A forecast built on eleven orders in a category is a guess with decimal places.
- Where is the output suspiciously uniform? When a flow recommends nearly everyone the same thing, a field has usually lost its contents.
- What happens at the edges? New accounts with no history, returns, gift purchases. Each of those needs a defined fallback.
- What would have happened without it? The only question that proves value, and the only one needing a control group.
Access rights and the review path are covered in the AI policy template for marketing teams, the border between a fixed workflow and an autonomous agent in the guide to AI marketing agents.
Segmentation fades, the lifecycle stays
Mauquoi's forecast for the craft is uncomfortable: old-school segmentation "will fade away", because personalisation per person is far stronger than any bucket. Five invented segments with five claims describe a state but never a path, and the same person lands in a different one depending on how you cut.
That does not mean segments leave the toolbox tomorrow. They stay useful for reporting, for finding where something is stuck, and as a fallback when a forecast does not hold. What fades is the idea that they are the destination. The lifecycle stays, because it describes the path, and the forecast sits on top of it. How to build the layer underneath is in the guide to CRM marketing in e-commerce.
CRM becomes a conversation
The trend almost nobody discusses, in Mauquoi's view: CRM becomes conversational. Until now it has been a one-way street, campaign out, response measured. With language models, genuine two-way communication becomes possible, with higher engagement and a different kind of loyalty.
The reading here is a channel question. Mauquoi observes a competitor running a messaging channel exactly like a newsletter channel, and calls that "completely wrong", because a messaging channel works in both directions, closer to social media. nexeye is not there yet either, by his own account.
Before you add any channel, two things follow. Replies need a named owner and a response time, or you have built a return channel with nobody at the other end. And every reply is data: it belongs in the customer record, not only in a support inbox.
Market data as input for those decisions
Your own data tells you what your customers did. It does not tell you which offer sat next to yours in the inbox. For decisions about timing and discount depth that is the missing half, and no forecast on your own history replaces it.
Mauquoi puts the principle plainly: "Every retailer should know who their competitors are and what their promotions are. Not in order to copy them, you should not do that. A strong brand draws its power from its own positioning. Nevertheless we have to understand what offers the competition is running." What matters is not how often the others send, but their headline promotion.
That outside view is what Reyo supplies. Competitor monitoring records which promotions and mechanics run in your market, market predictions place them in the shape of the year, marketing recommendations turn them into suggestions, and the campaign planner holds what your team decides. Two limits apply to any automation sitting on this data: we see what was sent, not how it performed, because nobody holds other shops' opens, clicks or revenue. And we never name the shops, because the point is the pattern, not the sender.
If the decision itself should happen inside your AI client, Reyo MCP is the interface. Your client, for example Claude, ChatGPT or Claude Code, reads your competitors' campaigns and emails, creates entries in the campaign planner and writes comments, always only inside your own account. Access is included from the Professional plan, and the guide covers the steps per client.
How that view enters your planning is in the guide to competitor monitoring, the underlying terms in what is marketing intelligence, the direction of travel in the marketing intelligence trends for 2026.
The order to start in
| Step | What you do | How you know it holds |
|---|---|---|
| 1 | Calculate purchase interval and next category per category | You can name a date per person |
| 2 | Put one flow on that interval instead of the calendar | Emails go out that were never on the plan |
| 3 | Move one decision, starting with the channel | Non-openers automatically get a different route |
| 4 | Book the weekly half hour of review | Odd predictions surface before they send |
The order is not arbitrary. Steps one and two cost days and hold for good, step three costs weeks and delivers the first real gain, step four keeps the earlier ones from breaking quietly. Start at three and you have moved a decision with no basis yet.
The multi-team rollout is in the guide to how to use AI in marketing.
Five mistakes that are easy to avoid
- Buying AI before forecasting exists. Without a forecast on your own history, the model sharpens a decision nobody makes.
- Measuring the campaign by its own revenue. Every discount produces some. Only the incremental part counts, which means the gap to a control group.
- Letting predictions run without a reviewer. A wild prediction on thin data sends just as reliably as a good one.
- Treating the return channel like a broadcast. Two-way communication needs ownership and a response time, not a template.
- Leaving out the outside view. Timing and depth are decided against a market your own data cannot see.
Conclusion
AI in CRM marketing pays off where it is invisible. Not in the copy and not in the image, but in the choice of channel, content and timing per person. Two switches sit in front of that, and no model flips them for you: forecasting yes or no, personalisation yes or no. Both work without AI, and both deliver more than any model layered on top.
If you do one thing this week, calculate the median purchase interval for your largest category and count how many people are well past it. That is your first forecast, it needs no model, and it tells you today who to write to tomorrow.
Two episodes go deeper: episode 3 with Laurens Mauquoi on bundles, omnichannel and hyper-personalisation, and episode 1 with Dr. Markus Wübben on CRM and AI. Both sit on the podcast page; the show is in German.
As of 8 November 2026. The statements and quotes from Laurens Mauquoi come from episode 3 of the Reyo podcast 360 Marketing & CRM, published on 14 October 2026, translated from German; his figures are his own, not Reyo data.