The Question Was Simple.
It Still Got It Wrong.

The last piece's conclusion was: write down the judgment that only lives in your senior staff's heads, and AI finally has raw material to work with.

01One simple question, three different answers

You ask AI a question a grade-schooler could answer: who was our biggest customer last year?

It thinks for a moment and gives you a name. Something feels off — that company isn't that big. You rephrase and ask again. It gives you a different name. Still not satisfied, you ask for the top three by revenue too — and the third name on that list turns out to be the same company as the first.

You start wondering if the AI is just dumb. But this question genuinely isn't hard — the hard part is invisible to you: in the data you gave it, it didn't find one company. It found three.

In the sales system, this customer is "Taiwan Semiconductor Manufacturing Company, Ltd." In finance, it's "TSMC." On the shipping records, because the goods went to the Southern Taiwan Science Park, it's listed as "TSMC Fab 14." Three records, three names, and not a single field tells the AI: this is the same company.

So, faithfully, it split one company into three, each carrying a third of the actual revenue, and then earnestly calculated a wrong answer from that wrong foundation.

The problem was never whether its math was right. It's that what it was handed was already fragmented from the start.

02The same thing, five different names

Zoom out from that one customer example, and you'll find the same thing in every corner of your company.

The same product: the ERP uses a model number, the e-commerce backend uses a product name, the warehouse uses a shorthand code only the picking staff can decode. The same employee: HR records a national ID number, the time clock records an employee number, the project software records an English nickname. The same payment: accounting sees a ledger line, sales sees a deal, and the owner just sees "that thing so-and-so still owes us."

Every individual system is internally correct. Put them together, and none of it lines up. Because none of them were ever designed to be viewed together — each went live on its own, solved its own problem, and grew its own naming conventions. No one, at the time, expected that one day you'd want to pour all of it into one place and ask a question that spans every system.

A human brain doesn't care about this. Your veteran employee sees "TSMC Fab 14" and automatically knows it's TSMC, because there's an unwritten lookup table in their head. AI has no such table. Without a shared key, it can only take things literally — different text means a different thing.

This isn't a flaw in AI. It's the price of its honesty. It won't "guess what you meant" the way a person would — it just answers faithfully based on the data you gave it. Fragmented data in, fragmented answers out.

03Fix the road before you run the car

At this point, most people's instinct is to buy a stronger tool, or switch to a pricier model. Wrong direction.

No matter how smart the model is, it can't conjure a key that doesn't exist in your data. This has to be handled one level back — not the AI itself, but the road underneath it: making the same customer, the same product, the same payment recognizable to each other across every system. This layer of work has a distinctly unglamorous name: data governance. But you can just think of it as plumbing and wiring — nobody gets excited about pipes, but without them, nothing downstream runs at all.

Fix the road before you drive the car. Most companies get stuck because they reversed the order — they already bought the car (the AI tool) before the road (clean, connected data) was ever paved, so the car just bounces along your potholed data, and you blame the car.

Here's what that looks like in practice. A company bought an automated reporting tool, expecting a weekly sales report to generate itself every Monday morning. The first one was wrong: the same product had two model numbers across two systems, so the tool counted it as two separate items being sold, and every number was skewed. After two weeks of this, everyone quietly went back to pulling numbers manually in Excel, and the tool has sat unopened in the backend ever since. Money spent, road unfixed, car going nowhere.

Why does the road always get fixed last? Because it looks the least like an accomplishment. Buy an AI customer-service tool, and you can show it off the next day. Spend three weeks unifying customer names, and when you're done, nothing on the screen looks any different — only you know the foundation moved. One is visible. One isn't. But the visible car only runs because of that invisible road.

04You don't need a full cleanup — work backward from one question

The good news: you don't need to stop everything and scrub the entire company's data clean first — that would turn into the same kind of endless, eventually-abandoned megaproject as the company-wide knowledge audit from the last piece.

Flip the direction. Don't sweep forward from the data. Work backward from the question.

Pick the one question you most want to ask, and ask most often — say, "who's our biggest customer." Then trace backward only along that question: which systems' data does it touch? How many different ways does the same customer get written across those systems? Build a lookup table that strings them into one shared key. Fix only that one road. Leave everything else alone for now.

The table doesn't need to be pretty. This is enough:

One row per customer, listing every scattered name side by side. From then on, whichever column AI reads, it knows they all point to the same row. A table with a few dozen rows like this, and the question that got answered wrong three times finally has a clean answer.

Once it's fixed, you'll get, for the first time, a real answer to a question you've been asking for a long time but never actually got right. More importantly, you'll know exactly how long this particular road is, and just how fragmented your data really is — which is more honest than any "digital transformation assessment meeting" you'll ever sit through.

Then move on to the next question. One road at a time. Fix the next one once this one's running.

Once this is done, the judgment you painstakingly wrote down in the last piece finally becomes usable — because AI can finally recognize that the customer you're talking about and those three names in the system are the same entity.

Fragmented data, fragmented AI

Back to that question that got answered wrong three times.

It was never that AI is dumb. It's that "who's our biggest customer" is a question your company has never had a single place capable of answering fully. The pieces of the answer are scattered across three systems under three names, waiting for someone to string them together. That job used to run on a veteran employee's mental math. If you want to hand it to AI now, you have to actually build that key — for real, inside the data.

Want to know exactly how fragmented your data is, and which road to fix first? Run a free assessment: answer a few questions about your existing systems, and we'll send back a prioritized list of your data gaps. No contract required, and you don't need to touch a single system first.

You can buy the tool anytime. What you can't buy is a road that actually lets AI run.

Rooted in craft. Built for the new wild.