AI Readiness: What Your Customer Data Needs Before AI Can Do Anything Useful
Someone in the leadership meeting has asked when marketing will “use AI.” The vendor has features. The team has a prompt box on every screen. What nobody has is an answer to the question underneath: is the data the AI would read in a state where the answer would be right?
Gartner found in February 2025 that 63% of organizations lack AI-ready data or can’t confirm they have it. It predicted that through 2026, 60% of AI projects without AI-ready data would be abandoned. The finding is usually quoted as a warning. It is more useful as a test, because “AI-ready” for customer data means four specific things, each of which a CRM team can check in an afternoon.
What ready means
Every AI feature in a marketing platform reads the same inputs a new junior colleague would read on day one: the event list, the profile, the consent record, and the catalog. That’s true whether it writes copy, builds a segment, times a send, or assembles a whole campaign. It reads them faster and with more confidence, and it doesn’t ask the person next to it whether purchase and order_completed are the same event.
So readiness isn’t about models or vendors. It is whether a stranger, human or machine, could read your customer data and get it right the first time. Four gates.
Fig 1 · Four gates before AI · Event names, identity, consent, and catalog IDs
Gate one: one name per business moment
Open the event list. Count how many events mean “a customer bought something.” If there’s more than one, an AI segment built on purchases uses whichever event the model picked. It won’t tell you which.
The same applies to properties. product_id, productId, and sku on three different events are three fields to a model. A naming convention with duplicates merged or retired is the first gate, and it is the one most tenants fail.
Check: events with the same business meaning, counted. Properties with the same meaning under different names, counted. Both should be zero.
Gate two: one customer per person
A model that scores churn risk per profile scores a person with three profiles three times, and a shared family tablet once. Every prediction, every audience, every “customers like this one” inherits the identity rules, and inherits their gaps.
Check: profiles sharing an email or customer number, counted. The key hierarchy, written. The shared-device rule, decided. If the duplicate count isn’t zero on a strong key, the model is scoring the wrong population.
Gate three: consent on the profile, per purpose, with a date
An agent that builds a campaign checks consent at the node. It can only check what is on the profile. If consent lives in the banner tool, in the old newsletter platform, or in a spreadsheet, the agent either sends to people who withdrew or suppresses everyone.
Check: every capture surface writes to the profile; every channel maps to a purpose; every record has a timestamp and an expiry. The consent model is the gate the DPO will ask about first when AI enters the conversation.
Gate four: the catalog joins to the events
Recommendations, price-drop triggers, “customers who viewed this also,” next best offer: every one of them joins a product in the catalog to a product in an event by ID. If the shop sends a variant SKU and the feed carries the parent ID, the model has nothing to join and recommends from the void.
Check: pick a product, capture its view event, find it in the catalog, compare the ID character by character.
What passing looks like
Not perfection. An account passes when all four counts are known, small, and owned. Someone keeps duplicates at zero. The identity rule is written in the tracking plan. Consent has one source of truth. The catalog ID was checked this quarter. The features can then be switched on in the order that costs least when wrong: analytics first, content tools second, agents last.
A tenant that fails a gate and switches the feature on anyway gets the same campaigns as before, faster, on the same broken data. That is where the abandoned projects in Gartner’s prediction come from. Nobody abandoned the model. They abandoned the confidence.
The afternoon
Four checks. Each is a count or a comparison. Write the four results on one page with a name next to each. If two or more are red, spend the next month on data first. The honest answer to leadership is “after those four.” If all four are green, the answer is “next week,” and it will be true.
The takeaway
AI readiness for customer data is four gates: named events, one identity, consent on the profile, a catalog that joins. They were the requirements before AI arrived. AI raised the cost of failing them, because a mistake that used to reach one campaign now reaches every campaign the model builds. Check the four, fix the red ones, then switch the features on. In that order, the features do what the vendor promised.
Read next
WRITTEN BY
Jan Sacha
Co-Founder · Bloomreach Consultant
Bloomreach Marketing consultant, in his third year at Asteroad running daily CDP and marketing automation work for European clients. Former Exponea enterprise consultant and CEO of Digiline.
Put these ideas to work.
If an article names a problem you are dealing with, tell us. We will listen first, then tell you honestly what we would do. No pitch, no commitment.