Your AI project is a filing project

Ask an owner what they want AI to do and the answer is usually some version of "I want to ask it a question about my business and get a real answer." Reasonable ask. What stands in the way is almost never the model. It's that the answer lives in four places under three different names, and at least one of them is out of date.

That isn't an AI problem. It's a filing problem, and it's older than any of the tools being sold to solve it.

Two kinds of data, one problem

Structured data is what sits in rows and fields — the AMS, the accounting package, the CRM. It's already machine-readable, which is why people assume it's fine. It usually isn't. The same client exists three times with three spellings. The renewal date in the CRM disagrees with the one in the carrier portal. Nobody can tell you which is right without asking a person.

Unstructured data is everything else: email threads, PDFs, scanned signature pages, contracts, chat history, the shared drive, the folder on somebody's desktop. This is where most of what a business actually knows is kept, and it accumulates the way sediment accumulates — by year, not by design.

Automation needs the structured side, because a workflow has to trigger on a field with a value it can trust. AI needs the unstructured side, because its answers are only as good as the documents you can hand it. Most companies have both in volume and have organized neither.

What "organized" has to mean

Not "we bought SharePoint." Organized means a handful of decisions, written down, that everyone actually follows:

What you hear	What it is	The decision
"Which one's current?"	No version discipline	One live copy, the rest archived
"It's in the drive somewhere"	Files named scan0142.pdf	A naming convention
Two systems, two numbers	No system of record	One owner per fact
Search returns 2019	No retention rule	What expires, and when
"Ask Denise, she knows"	The index is a person	Tags instead of memory

The specifics matter less than the fact that they're decided and written down. A convention as plain as 2026-07-14_acme-corp_renewal.pdf beats an elegant scheme nobody types. A system of record means one system owns the field and every other system displays it without editing it. None of this is a technology purchase — it's governance, and it's unglamorous enough that it usually gets skipped in favor of buying something.

Why this decides whether AI works

An AI assistant pointed at a disorganized business doesn't fail loudly. It fails in four specific ways that are easy to miss until they cost you something:

• It can only reach what's findable. Retrieval works on names, text, and metadata. A drive full of scan0142.pdf is, to a machine, a drive full of nothing.
• Duplicates produce confident wrong answers. Hand a model two versions of a contract with no signal about which one was signed, and it will answer from one of them. It will not mention that it picked.
• It inherits your permissions. This is the one that gets underestimated. An assistant wired into a shared drive is the fastest search tool your company has ever had, applied to whatever access sprawl you've accumulated. If folder permissions have been a "we'll fix that later" for six years, later just arrived.
• You can't tell if it's right. With no system of record, there's nothing to check the answer against — so the answer gets trusted or ignored based on how confident it sounds. Both are bad outcomes.

Automation degrades the same way, just faster and more visibly. A workflow keyed to a field that three people fill in three different ways doesn't produce a wrong answer; it produces an exception queue that somebody has to work by hand, which is the thing you were trying to stop doing.

Priced the same boring way

Time spent looking for files is a process cost like any other, and it prices with the same deliberately boring formula:

people × minutes per day searching ÷ 60 × workdays per year × blended hourly rate

Run it on the modeled 24-person agency we use throughout these articles, at its $35 blended rate, and assume ten minutes per person per day spent hunting for a document or reconciling two systems that disagree:

24 × 10 ÷ 60 × 240 × $35 = $33,600 a year

Two cautions, because this is exactly the kind of number that gets abused. First, the agency is a modeled fixture, not a client. Second — and more important — the ten minutes is our input, not yours. We don't know what your number is, and neither does any vendor quoting you an industry average. Have two or three people log it honestly for a week before believing any version of this figure, including ours. A measured five minutes beats a borrowed thirty.

The order that works

1. Name the system of record for every entity that matters — client, policy, invoice, employee. One system wins per fact.
2. Pick a naming convention and a folder shape that mirrors how work moves, not how the org chart is drawn.
3. Write the retention rule, including who deletes and when.
4. Then automate the process.
5. Then point AI at the result.

Most stalled pilots are an attempt to start at step four or five. The tool works in the demo, on clean sample data, and then meets the drive.

Two opinions worth stating plainly. A shared drive with a convention people follow beats a document management platform nobody follows — the platform is not the discipline. And not everything needs to move: migration is often the most expensive available way to avoid deciding which system is the system of record.

Where this shows up in an assessment

Half of what a process interview captures is really a data map: where each fact lives, what it's called, who owns it, which system wins when two disagree, and how many times a single number gets re-keyed between systems before it reaches a customer. That map is worth having whether or not you ever automate anything, because it's the document that lets you answer "why do these two reports disagree" in minutes instead of an afternoon.

It's also the precondition for everything else. The $999 assessment produces it as a byproduct of documenting how your business actually runs — which is the honest first step of any AI project, no matter what it's being sold as.