Is Your Data Actually Ready for AI? (Probably Not — Here's How to Tell)

Every few weeks someone asks me about AI for their business. Usually some version of "can we get something that answers questions about our stuff."

The answer is usually yes. The hard part sits upstream of anything the AI does.

Here's a test you can run in about ten minutes without calling anyone.

The Ten-Minute Test

Pick one question you'd genuinely like answered. A real one:

• Which customers haven't ordered in six months, and did any of them complain first?
• How long does a job actually take, from first call to final payment?
• Which jobs made money last year, and which ones only looked like they did?

Now go answer it.

Count two things while you do. How many programs you had to open, and how many times you had to ask another person.

Three programs and one person is the usual score. That score is the thing standing between you and any AI project, because it means the answer doesn't live anywhere. It gets assembled by hand, each time, by someone who knows where to look.

A machine can't read something that only exists when Dave reconstructs it.

Where It Actually Lives

I've been in enough of these businesses to guess what I'll find before I walk in.

Email is the real system of record. The CRM got bought, filled in for two months, and abandoned, because email already worked. Quotes, approvals, change orders, complaints, the promise somebody made on a Tuesday — all sitting in inboxes, one per person, searchable by exactly one person each.

Everybody is a point of contact, too. A customer calls whoever they spoke to last time, which is fine for the customer and terrible for the record.

Files land in three places: some Dropbox, some Drive, mostly Office 365 by now, plus a share drive nobody has audited since the last IT guy left.

Spreadsheets are load-bearing. Not "we track a few things in Excel." The business runs on four workbooks, and one of them has a macro nobody can explain.

If there's a tech guy, there's an Access database on the network share. It works. Don't touch it.

And there's usually one machine that can't be upgraded. I worked with a print shop whose etcher needed Windows ME and a serial cable. Replacing that etcher was fifty thousand dollars, so the answer was to keep the old machine running as a VM and leave it alone. Any plan that assumed it would go away was a plan for a different business.

All of this accumulates when a company grows by solving problems as they arrive. It leaves you somewhere specific: your information isn't missing — it's stranded. It exists, and it can't get to where the question is being asked.

The Part Nobody Puts in a Brochure

Some of it is stranded on purpose.

In most businesses I've seen, at least one person's standing is built partly on being the only one who knows where something is. That's a reasonable response to feeling replaceable, and it will do more to sink a data project than any technical problem will.

So if you do this, that person has to come out of it more valuable than they went in. Any plan that skips them fails quietly in month three.

Three Doors, and You Only Need One

Once you can see where things live, what people want tends to fall into three shapes.

"I want to ask my data questions." Plain English in, real answers out, drawn from your own records instead of a chatbot's imagination. The most common ask, and the cheapest to do well, provided the records exist somewhere a machine can read.

"I want to see how it connects." Where the time goes, which customers cluster, where jobs stall. A dashboard fed by live connections to the systems you already pay for. This one needs the plumbing to be real rather than reassembled by hand each month, which is where most of the work hides.

"Tell me what to fix." Point it at customer history, payments, complaints and the marketing calendar together, and ask where the gaps are. Most valuable, most groundwork, because it only works when those four things can be looked at side by side.

Notice the order. Each door needs the one before it to be true. Most people ask for the third and need the first, and that's just how the work stacks.

When Not to Do This

Forty invoices a year and you know every customer by name? You don't need any of this. You are the system, and you're a good one. Spend the money elsewhere.

If the actual problem is that the phone isn't ringing, that's a website and local search conversation, and I'd rather have that one with you than sell you the expensive thing.

And if someone quotes a six-figure "AI transformation" before asking how a job moves through your shop, they're selling a platform.

Some of You Can't Send This Anywhere

Handling medical records, legal files, or client financials? The standard advice, pipe everything to a cloud AI service, may not be available to you.

It can run on a box in your office instead. The data never leaves the building, there's no monthly per-seat bill, and nobody's terms of service get a vote. Slower than the big services, and for this kind of work you won't notice.

I mention it because it's usually what turns "we can't do this" into "oh, we can do this."

What I'd Do First: an Office Flow Map

Before any of the three doors, I make one thing. An Office Flow Map: how a job actually moves through your business, and where the information gets stranded.

First contact through quote, work, invoice, payment, and whatever happens afterward, with every place the information lands marked on it, and every place it stops.

It runs two to four days once the sheets are gathered, and it scales with how many processes you have. Ten processes means ten sheets somebody has to actually read. It's worth having even if you never build anything, because it's usually the first time the whole path has been on one page.

Here's the entire Office Flow Map Process. If you'd rather do it in-house, do it:

1. Pick one real job you finished recently. A real one, with a name on it, rather than a typical one.
2. Write down every step from first contact to final payment. Who touched it, in order.
3. At each step, write where the information went. An inbox. A spreadsheet. A whiteboard. Somebody's memory.
4. Mark every place it stopped. Anywhere someone had to re-enter it, re-ask for it, or go find the person who knew.
5. Do it again with a job that went badly. The stops land in different places, and those are the expensive ones.

That's the method. No software required, and a whiteboard is genuinely fine.

What you end up with is a picture of your business nobody has drawn before, and a short list of places where the same fact gets typed in more than once. Fix those and you'll have taken most of the value out of this before AI enters the conversation at all.