Most data vendors now sell AI-ready data: records a machine can act on without a human standing by to catch the mistake. Ask one for the pass and fail criteria behind the label and the deck moves to the next slide. I have yet to see a vendor hand over the test.
So here is one. Four checks, all testable against a real CRM or a freshly purchased list, none of them requiring you to buy anything to run.
What is AI-ready data?
The short answer: AI-ready data is data a machine can act on without a human standing by to catch the mistake. For CRM data that means passing four checks: reachability (emails validated past the accept-all layer, phones that get answered), accuracy (firmographics that match the live world), structure (fields that carry a definition and an owner) and freshness (decay measured on your own list). Fail any one and the AI on top inherits the failure at whatever scale you deployed it.
Few pass. A March 2026 Harvard Business Review Analytic Services survey of more than 230 people who make AI data decisions at their organizations, sponsored by Cloudera, found just 7% say their data is completely ready for AI. The feeling of readiness runs well ahead of that. Per the 2026 State of Data Integrity and AI Readiness from Precisely and Drexel University’s LeBow College of Business, a survey of 505 data and analytics leaders, 87 to 88% claim their organization is AI-ready while 43% admit data readiness is their biggest obstacle. Same respondents, same survey.
What makes data AI-ready?
Gartner’s answer is that readiness depends on the job. In its February 2025 guidance on AI-ready data, Gartner says the data “must be representative of the use case,” including the patterns, errors and outliers the AI will meet, and that “if the data has issues, then the data is not ready for AI.” Gartner predicted that through 2026, organizations would abandon 60% of AI projects for lack of AI-ready data.
Most published definitions are written for data platforms and model training. IBM defines AI-ready data as “high-quality, accessible and trusted information” and lists its traits as unified and accessible, governed, secure and supported; Gartner’s own steps run through metadata, data pipelines and observability. All of that holds, and none of it tells a revenue team whether the next record an AI SDR emails is the right person at the right company. A CRM feeding a sales agent is its own use case, so by Gartner’s logic it needs its own test, run one record at a time. If the AI will also write back into the CRM, the bar rises again, which is why we argue AI agents should not get write access yet.
New data fails before it gets old
The decay narrative says data starts clean and rots. Our buying experience says the rot is often pre-installed. When we take a fresh list from a major data provider and run it through a validator that double-checks accept-all addresses, quality typically drops 15 to 20% while the data is still brand new. That is the condition it arrived in.
The accept-all layer is where most of it hides. An accept-all domain answers yes to every address thrown at it, so a standard verification pass marks the record valid and the first real send finds out otherwise. ZeroBounce’s 2026 Email List Decay Report, based on more than 11 billion verified emails, found at least 23% of an email list degrades annually. The same report found catch-all addresses made up more than 9% of everything it checked in 2025, when only 62% of all emails submitted to it were valid.
The counter is a category I rarely see teams run: double verification. Verify every email, then take everything that comes back catch-all and verify those again. That second pass produces significantly more accurate results than most major verification systems manage in one. For one client it changed the entire picture. Across roughly 160,000 contacts, a first read put almost half the list in question because of catch-all domains. The second pass settled it in both directions: over 40,000 contacts were undeliverable or high risk and came off the list before they could bounce and damage deliverability, and the catch-alls that verified clean moved back into the known-deliverable pool. Same list. The difference was knowing which part of the uncertainty was real.
Email is only the measurable edge. The same on-arrival problem shows up in locations, titles, company names, and industry tags, which is the class of field I watched an AI enrichment feature rewrite at scale. And the phone channel is going through its own version: vendors like TitanX are shifting the standard from numbers that exist to numbers that get answered, which is the right question, because a phone field that connects nowhere is an accept-all domain wearing a different costume. Your reps pay for each of these failures in dial time and bounce handling, out of a selling week that was already thin.
Why won't vendors define AI-ready?
Because a definition is a test, and a test can be failed by inventory they have already sold. A data provider that publishes pass and fail criteria invites every customer to run them against last quarter’s delivery. The label works better soft, the same way “premium” works better than a spec sheet.
We can publish the test because we do not sell data. Nothing in our business gets hurt if you run it and pass. Run it and fail, and you have learned something no data vendor could afford to tell you.
The AI-ready data checklist, in four parts
Run each check against your own CRM or a freshly purchased list. A record is AI-ready only if it passes all four, and a list is only as ready as the share of its records that do, measured against the job you plan to hand the AI.
| Check | Pass looks like |
|---|---|
| Reachability | Emails validated past accept-all domains. Phone numbers that get answered, not numbers that merely exist. |
| Accuracy | State, country, title, and company name verified against the live world, not inferred by a model filling blanks. |
| Structure | One field per fact. Every field carries a written definition, a named owner, and no lookalike twins. |
| Freshness | Decay measured on your own list, on a schedule, instead of assumed from industry lore. |
The four checks are not proprietary, and that is deliberate: they are what good practice looks like whether or not we are involved. Our free CRM diagnostic covers the same ground across 40+ read-only checks, and on request we test every contact email in the instance, so you get a measured picture of your data quality instead of an estimate. One boundary worth stating plainly: we are not a data provider. If the results show you need better data, we help you find the right provider for your motion. Our work is finding the problems you did not know you had, which is a different job from selling you replacements.
One design choice is worth stating too: the diagnostic’s score is calculated by deterministic rules, not an LLM, and the connection never writes to your CRM. In an article about AI readiness that may read as ironic. It is the point. You do not grade the student by asking the student.
Two emails worth sending
Before your next data purchase, send the vendor the four checks and ask which ones the file passes. Before your next AI line item, run them against your own CRM.
One of those two emails is uncomfortable to send. The other is uncomfortable to receive. Send both.