In short: only 21% of teams say their CRM data is well prepared for AI. You do not close that gap by buying a tool. You close it by counting three things: how many records have an owner, how many have the right value in the field you segment by, and how many required fields are actually filled by an integration. That takes a week.
The number comes from Validity's State of CRM Data Management 2026, published on August 25, 2026, with 500 marketing professionals in the United States, the United Kingdom, Brazil, Australia and New Zealand. The same report says 62% of organizations lost revenue to poor data quality, that nearly a third of teams spend six hours or more per week reconciling information, and that only 41% have someone accountable for data governance.
The figure that matters most is not the 21%. It is the 78% of C-suite leaders who admit they acted on an AI recommendation they later suspected was wrong because of the data underneath. That is the real invoice for skipping the audit.
It does not mean the database is perfect. It means something more modest and more demanding at the same time: the fields your AI will decide with have a stable meaning. If your industry field was filled automatically, by a data vendor, and nobody checked it, an agent that segments by industry is working on a tidy fiction.
The difference between bad data and ambiguous data matters. An email that bounces is bad data and it surfaces on its own. A contact labeled "technology services" when the company actually sells people's hours is ambiguous data: it passes every validation, looks fine in the report, and sends the wrong message. AI does not correct ambiguity, it amplifies it, because it applies it at scale and without hesitating.
We measured this on our own prospecting database. On November 5, 2026 we enriched 105 B2B contact records from Chile and Argentina with a commercial data vendor and checked each one against the service catalog the company publishes on its own website.
The result: 79 of the 105, or 75%, arrived with the exact same industry value. Looking at their catalogs, 43 of 105, or 41%, describe staffing, people outsourcing, BPO or agency services, which for segmentation purposes is a different conversation. And 7 of 105 carried an email on a domain different from the company website. Counting both problems without double counting the records that had each, 49 of 105 records, or 47%, would have entered the CRM carrying something that breaks a segment.
This is our data, from one specific slice and one week, not a representative sample of anything. But it is the kind of count anyone can repeat on their own database in an afternoon, and it is more useful than a survey.
You do not need a three month CRM data audit to know whether you are close or far. You need three numbers, and all three come from a report you can build today.
| What you count | What good looks like | What it means if it is off |
|---|---|---|
| Records with no owner assigned | Under 5% of active contacts | Nobody answers for that data, so nobody fixes it |
| Records where the field you segment by is empty or still on its default | Under 10% | Your segmentation runs on part of the database and treats the rest as if it were the same |
| Required fields that are actually filled by an integration | Zero, or documented one by one | A rule you think protects you only applies where you put it |
The third one surprises people the most. In HubSpot, required properties set in Record Creation Settings do apply to API writes, but required associations only apply in the interface. If your integration creates deals without associating the company, nobody sees the gap until a report looks odd.
There is a practical reason and a governance one. The practical one: AI bills per action. Every enrichment, every summary and every recommendation that runs on a wrong record is budget leaving the building, and we put numbers on that in our article on HubSpot AI pricing.
The governance one is newer. Since the October 2026 rollup, AI connectors can read properties and engagements flagged as Sensitive Data, in public beta. Only a Super Admin can turn it on and access is read only, but it is exactly the kind of checkbox that gets switched on without a conversation. If your database has no data owner, it also has nobody whose call that is.
When all three counts come back bad at once, the problem is no longer cleanup, it is design: fields that mean different things to two teams, or integrations writing without rules. At that point it pays to map the data model before touching a single record, because cleaning on top of a broken design unravels in three months. That is what we do in a diagnostic inside your portal, and what it costs is published.
If you want to know where your HubSpot stands today, the free self assessment takes 3 minutes (in Spanish).
The three counts above, once a month, take under an hour and can live as a saved report. A deeper review of the data model makes sense once or twice a year, or whenever something structural changes: a new integration, a new product, or a new team starting to write into the CRM.
It is worth it if the order is right. First you define what each field means and who fills it, then you clean, and only then do you connect the vendor with rules about what it may overwrite and what it may not. Cleaning without that definition is work you repeat every quarter.
For mechanical work, yes: deduplicating, normalizing formats, flagging emails that bounce. For deciding whether a contact belongs in a category, not yet, because that judgment is exactly what you are auditing. And every run costs, so measure the cost before handing it the whole database.
Start with the field you segment by today, not with all of them. One well defined, well populated field already changes campaign results and assignment. Trying to fix thirty fields at once is the most common way to finish none of them.