- The Unglamorous Prerequisite
Getting Your Data Ready for AI
AI amplifies whatever is underneath it. Five conditions determine whether that amplification helps you or produces confident nonsense faster than anyone can check it.
WHY THIS COMES FIRST
The Failure Is Quiet, Which Is What Makes It Dangerous
When a traditional system meets bad data, it usually breaks visibly: an import fails, a field rejects, a report comes back empty. When an AI system meets bad data, it produces a fluent, plausible, well-formatted answer that happens to be wrong. There is no error message, and the output looks more authoritative than what a person would have written.
That asymmetry is the whole argument for doing this work first. You are not preparing data so the technology functions; you are preparing it so that when the technology functions, you can trust what comes out.
The good news is that this work has value regardless of what happens with AI. Clean records, sane permissions, and reachable systems make everything else in the business easier, which is why we are comfortable recommending it even to clients who decide against an AI project entirely.
Data preparation is where AI projects at Dallas small businesses quietly stall, usually because the records live in three systems that were never meant to talk. We see the same three-system tangle across practices and firms in this market, which is why we look at the data before anyone signs off on a tool.
FIVE CONDITIONS
What "Ready" Actually Means
1
Reachable
A tool can query it without a person exporting a file first. Manual export is the single most common reason a promising workflow never becomes a habit, because the friction always exceeds the benefit. The four ways we connect systems sets out what “reachable” looks like in practice.
2
Consistent
One client, one record. Three spellings of the same company, duplicate patient entries, and free-text fields holding structured information all produce output that is wrong in ways nobody catches.
3
Permissioned
Access reflects who should see what, today. AI tools that surface documents inherit your permissions exactly, including the folder shared with the whole company in 2019 that everyone forgot about.
4
Current
Superseded versions are archived rather than sitting alongside the live one. A tool that cannot tell which price list is current will cite the wrong one with complete confidence.
5
Classified
You know which data is regulated, confidential, or public. Without this, every AI decision becomes a case-by-case argument rather than a rule anyone can follow.
Scope this narrowly
You do not need company-wide data perfection. You need these five conditions to hold for the specific data the first workflow touches. Businesses that attempt a full data cleanup before starting anything typically never start anything.
How long does data preparation take?
Scoped to one workflow, usually days to a few weeks. Permissions review is fast; deduplication depends on how many records and how bad the drift is. It is almost always shorter than businesses fear and longer than vendors imply.
Can AI clean up our data for us?
For some tasks, genuinely well, matching duplicates and extracting structure from free text are things it does better than a rule. What it cannot do is decide which of two conflicting records is correct. That still needs someone who knows the business.
Our data lives in an old system with no integration options.
Common, especially in practices and firms running specialist software. Sometimes there is an export or reporting layer we can work with; sometimes the honest answer is that this workflow is not the one to start with, and a different process is. That is a finding, not a dead end.
BK
Last reviewed August 2026 · Questions we have not answered here? Call 214-444-3583. Kehr Technologies is based in Plano, Texas, and works with businesses throughout the Dallas area.