Why AI Exposes Your Company's Data Problems

For years, businesses have operated with imperfect data. Customer information lives in the CRM, orders live in the ERP, and operational data lives somewhere else. And there's probably a spreadsheet somebody maintains because none of those systems quite tells the whole story.

Yet the business keeps running.

Why? Because people are remarkably good at filling in the gaps.

They know that one customer somehow has three records in the CRM. They know which number in the spreadsheet is actually correct. They know that "complete" means one thing to Operations and something slightly different to Finance. And they know who to ask when the system doesn't have the answer.

Then AI enters the picture, and suddenly all those little inconsistencies we've learned to live with aren't so little anymore.

AI didn't create the problem. It just stopped hiding it.

Why Does AI Expose Data Problems?

AI depends on context that people often take for granted. An experienced employee doesn't just look at data. They interpret it.

They've learned which information to trust, which rules have exceptions, which customers require special handling, and which processes don't work quite the way the documentation says they do. AI doesn't automatically have any of that institutional knowledge.

Give it conflicting information and it has conflicting information. Give it outdated information and it has outdated information. Leave an important business rule in someone's head and, as far as the AI is concerned, that rule doesn't exist.

This is one reason an AI proof of concept can look fantastic and then struggle when introduced into the real business.

The model didn't suddenly get worse. The environment got real.

Your Employees May Be Your Best Integration Layer

Here's an exercise: watch what happens when a customer calls with a question that isn't completely straightforward.

Someone opens the CRM, then the ERP, then maybe a spreadsheet. They check an email and message someone in Operations because there's one more thing they need to know. Finally, they give the customer an answer.

We call that "doing the job."

Technically, that employee just performed an integration. They gathered information from multiple systems, reconciled inconsistencies, applied institutional knowledge, and made a decision. Businesses do this thousands of times every day.

That's an important distinction when evaluating AI. The question isn't simply:

Can AI perform this task?

It's:

Can AI access and understand everything a good employee uses to perform this task correctly?

That's a very different question, and usually a much more useful one.

The Spreadsheet Isn't the Problem

Spreadsheets take a lot of abuse. Sometimes they deserve it.

But if a team created a spreadsheet because two systems don't give them the information they need, deleting the spreadsheet doesn't solve anything.

The spreadsheet is evidence.

So are:

• manual exports
• shared documents
• email approvals
• personal notes
• duplicate databases
• Slack or Teams messages that have somehow become part of the process
• "ask Steve, he knows how it works"

These workarounds show you where your formal systems and processes stopped matching how the business actually operates. That's useful information.

Before automating around a workaround, understand why it exists. Otherwise, there's a decent chance you'll spend a lot of money automating the wrong process.

Does AI Need Clean Data?

No. AI needs usable data in context.

There's a difference.

If someone tells you that you need to clean every piece of company data before you can use AI, congratulations. You've just been handed a three-year project that will probably never end.

Perfect data isn't the goal. For a specific AI use case, the information it depends on needs to be accessible, reliable enough for the task, current enough for the decision, and understandable in context.

That's AI data readiness: the degree to which the information required for an AI use case is accessible, reliable, current, and understandable in the context where it will be used.

That definition is intentionally tied to a use case. Don't start with:

How do we clean all our data?

Start with:

What does the AI need to know to do this job well?

Then ask:

• Where does that information come from?
• Which source is authoritative?
• Do our systems agree?
• How current does the information need to be?
• Who owns it?
• What information does an employee know that isn't captured anywhere?

Now you have a problem you can actually solve.

Access to Data Isn't the Same as Understanding It

Let's say you connect an AI assistant to your CRM. Great. Now it can retrieve customer records.

But does it know which record is correct? Does it understand that certain customers have special contract terms? Does it know which exceptions require approval, whether a field is trustworthy enough to make a decision from, or what employees normally do when two systems disagree?

That's context.

And it's why system integration is necessary for operational AI, but isn't sufficient on its own. APIs, integrations, and technologies like Model Context Protocol (MCP) can give AI access to business systems, tools, and information.

That's an important step, but connectivity doesn't magically create understanding.

The business still has to define what information means, which sources should be trusted, what rules apply, what the AI is allowed to do, and when a human needs to make the call.

Giving AI access to everything isn't a strategy. Giving it the right context is.

Why Is Tribal Knowledge a Problem for AI?

Tribal knowledge is operational knowledge employees rely on that isn't formally captured in company systems or documentation.

Every organization has it:

"Don't process those until Tuesday."

"That customer always orders it this way."

"If you see this error, call Sarah."

"The system says X, but we actually do Y."

People learn these things through experience. Eventually, they stop thinking of them as special knowledge. It's just how the business works.

Until you ask a machine to do the same job.

Then you discover how much of your operating model was never actually documented.

AI makes tribal knowledge a technology problem. If a rule matters, capture it. If an exception matters, define it. If employees consistently need information that isn't available in your systems, figure out where that information should live.

That's useful whether AI ever touches the process or not. You're reducing the business's dependence on information living inside individual people's heads.

Start With the Business Decision, Not the Data Cleanup

The worst response to discovering all of this would be:

"We need an enterprise-wide data cleanup initiative."

Maybe you do. But I wouldn't start there.

Start with something the business actually wants to improve. Maybe you want AI to:

• identify orders likely to be delayed
• prepare information before a customer service interaction
• review documents for exceptions
• surface equipment maintenance risks
• route an internal request
• recommend the next action on an account

Pick one and work backward.

What would your best employee need to know to make that decision? Where does that information live? Which parts can they trust? Where are the gaps? What judgment does the employee add? What happens when the answer isn't obvious?

That's your roadmap.

Not for fixing every piece of data your company owns, but for creating enough trusted context to improve one meaningful business process.

Then you learn. Then you expand.

Your Data Problem Might Actually Be a Process Problem

Here's where things get interesting.

Once you start tracing how information moves through a workflow, you may discover that data isn't the real problem. The process is.

Maybe two departments define the same thing differently. Maybe nobody actually owns a critical piece of information. Maybe three approvals exist because of a decision somebody made ten years ago. Maybe employees enter the same information into multiple systems because nobody ever connected them.

Or maybe the spreadsheet is doing exactly what it needs to do because the official workflow is the thing that's broken.

AI has a habit of dragging these problems into the light.

That's a good thing, because fixing them can create value before the AI ever does.

The Opportunity Is Bigger Than AI

Connect two systems and you might eliminate duplicate entry. Clarify data ownership and your reporting gets better. Document a process and onboarding gets easier. Capture institutional knowledge and operational risk goes down. Remove unnecessary steps and customers get answers faster.

None of those benefits require a large language model.

That's why I don't view data readiness as some giant hurdle businesses have to clear before they're allowed to use AI. AI can be the catalyst that finally makes these problems worth solving.

The inefficiencies, disconnected systems, and undocumented rules were already there. Your employees were simply good enough to compensate for them.

Now you have a reason to look more closely.

AI Needs Trusted Context, Not Perfect Data

AI is forcing businesses to answer a question they've been able to avoid for years:

How does information actually move through this company?

Not how the architecture diagram says it moves or how the process document says it moves. How it really moves: through systems, spreadsheets, emails, conversations, and the accumulated experience of the people doing the work.

That's where some of the biggest opportunities for AI are hiding.

But the goal isn't perfect data. It's trusted context.

Understand what your best people know, where they get their information, how they resolve ambiguity, and where your systems fall short.

Do that, and you're not just preparing your data for AI.

You're probably improving the business along the way.

Frequently Asked Questions