You connect the new AI tool on a Friday. You expect it to know your customers by Monday.
Then it fires a "we miss you, come back" email at someone who bought two days ago.
The tool was fine. The data under it was a mess.
This is the quiet reason so much AI marketing falls flat, and there is a number for it. In Salesforce's tenth State of Marketing report, a survey of 4,450 marketers across 26 countries, 98% of teams already using AI said they run into at least one data barrier to personalisation. Data silos. Too much data. Poor quality data.
Nearly half, 46%, said they lack the customer preference data to send anything relevant at all.
So the clever part is no longer the problem. The ground it stands on is.
An agent does not read your mind. It reads your data. Show it who bought what, when they bought it, and what they did on your site, and it can work out the next sensible move for each person on its own. Hand it fog and it guesses. That is the real promise behind agentic marketing, and it lives or dies on the foundation you give it.
Hand an agent clean data and it finds the next move. Hand it fog and it guesses.
What data does an AI marketing agent actually need?
Four things, in plain terms: order history, your product feed, customer events, and permission to send. Klaviyo's Shopify data sync shows the shape of it. It builds a profile for every customer with first and last purchase dates, lifetime value, total orders and consent status, pulls in the full catalogue with names, prices, images and variants, and streams live events like Placed Order, Viewed Product and Added to Cart, most of them landing within seconds.
Order history tells the agent the rhythm of a customer. Someone who reorders every 30 days is a different person from someone who bought once in a sale, and they deserve a different message.
The product feed is how the agent recommends the right thing instead of a random thing. No feed, no relevance.
Events are the live signal. A Viewed Product yesterday means more than a purchase last spring.
Permission is the one people skip. If you cannot prove someone opted in, the smartest agent in the world still cannot send. If you want to see how those four sources turn into real decisions, we broke it down for a Shopify and Klaviyo store.
Why does messy data break AI personalisation?
Because the picture is split across too many places. Salesforce found the average marketing organisation has seven separate data sources to join up for agentic marketing, and only around half have full access to the sales and commerce data they need. When those sources disagree, the agent inherits the disagreement.
One system says a customer churned. Another shows them buying last week. The agent picks one, and half the time it picks wrong.
The blast
- One send, one schedule, everyone
- 'We miss you' to a two-day-old buyer
- Confident, and pointed the wrong way
An agent on clean data
- One decision, per customer
- The reorder note, on the day it is due
- Confident, because the orders say so
Bad data does not slow an agent down. It points it in the wrong direction with total confidence. That is worse than doing nothing, because it spends the customer's patience and your sender reputation at the same time. Clean data is not admin. It is the thing that decides whether every future send helps you or quietly costs you.
How do you get your Shopify and Klaviyo data ready for AI?
Start with the joins, not the tool. 61% of marketers told Salesforce that full data integration is still a work in progress, so you are fixing the same thing everyone else is fixing. Five moves get you most of the way there.
- Pick one identity. Email is usually it. Make sure a customer is one profile across Shopify and Klaviyo, not three with different spellings and half a history each.
- Clean the product feed. Every item needs a name, an image, a price and a category that is actually true. This is what the agent recommends from, so a wrong category becomes a wrong email.
- Turn the events on. Confirm Placed Order, Viewed Product, Added to Cart and Checkout Started are all flowing. These are the signals that let an agent act at the right time instead of on a fixed schedule.
- Fix consent. Sync opted in status properly for email and SMS, and keep it current. It protects you, and it makes every send you are allowed to make count for more.
- Keep the history. Do not wipe old orders when you migrate or replatform. A year of purchase history is the difference between a guess and a good decision.
None of this needs a data team. Most of it is an afternoon inside your Shopify and Klaviyo settings, and it is worth more than any feature you could bolt on top. If you want a fast read on where your own foundation leaks, the free revenue leak audit walks your setup and shows the gaps in plain numbers.
How do you know the data foundation is paying off?
You measure it against a control group. Hold back a slice of customers the agent leaves alone, let it work the rest, and compare the two. That is the only honest way to know the data and the decisions are earning their keep, rather than taking credit for sales that would have happened anyway.
This is the part I care about most, because it is where the foundation turns into money. When we model PilotX against a control group a brand sets for itself, a cleaner foundation lifts the modelled return. The lightest mode, where you steer, models around 19% more revenue. Working alongside you and approving each move, around 30%. The confident moves running on your rules, on a brand that has done the data work above, up to 50%, roughly three times what older personalisation tends to top out at.
I want to be straight with you. That is modelled against a control group, not a measured case study, and we have one live customer today. But the model only climbs because the data underneath it does. Better foundation, higher ceiling. That relationship is the whole point.
If you only do one thing this week, open Klaviyo and check whether your events are actually firing. Not the setup screen. The live feed. Watch a real Placed Order land on a real profile. That one look tells you more about whether you are ready for AI than any report will.
When the foundation is there, PilotX finds the biggest gap in your lifecycle and builds the fix on your own products before you connect anything, free, so you can watch it work on your real customers first. That is the offer. The data comes first, though. It always did.