Your Guide on How to Prepare Your Company's Data for AI

1. The boundary problem: a chunk can split meaning right down the middle

Chunking decides where one passage ends and the next begins. Most default settings do it by size, roughly every 300 to 500 tokens, with no awareness of what’s inside.

What breaks:

• Tables get cut mid-row.
• Numbered exceptions get separated from the sentence that qualifies them.
• Clauses that only make sense next to the line above them get split apart.

Steps to check this, this week:

1. Pick your single highest-value document type, a loan file, a clinical note, a permit form.
2. Run one through your actual chunking process.
3. Pull four or five resulting chunks.
4. Read each one against the source, focused on tables, numbered lists, and dependent clauses.
5. The file loaded tells you nothing. Only reading the actual chunks does.

2. The ownership problem: clean isn’t the same as owned and dated

Most companies can’t name who’s responsible for a document staying accurate, or when it’s next due for a check.

Fix it like this:

1. List the documents feeding your top three to five AI use cases only, not the archive.
2. Assign one named owner per document type.
3. Tie the recheck to something already on a calendar, a renewal, an annual review, not a new task nobody will run voluntarily.

This is precisely what we do, before an agent goes anywhere near production data: get every document owned, dated, and tagged so nothing stale slips through unflagged. It’s not a feature we bolt on afterward, it’s the first thing we build. We’ve written before about what happens when an assistant has no memory of how your company actually works, The Amnesiac Genius, and this is the other half of that same problem.

3. The confidence problem: the costliest failure never looks broken

There’s a name for this in production AI: context poisoning. A fluent, confident answer built on a source that’s dead or superseded, delivered in the exact same tone as a correct one.

Why it’s worse than a missing answer:

• A missing answer gets questioned.
• A confident one usually doesn’t.

Here’s the fix:

1. Require citation grounding on anything decision-relevant, every answer names the specific chunk it drew from.
2. Sample your high-stakes answers monthly.
3. Pull the citation on each. Check its last-verified date against today.
4. Without grounding switched on by default, this failure only surfaces after it’s already cost something.

4. The threshold problem: ready is a setting, not a property of data

The diagram below shows one customer record feeding two jobs. A marketing recommendation can run broad and loose, a miss costs nothing. A credit decision needs the identical data tuned for precision, with a mandatory human sign-off. Same data. Different bar.