AI is getting smarter. Your business data might be what holds it back.

The AI available to businesses is improving incredibly quickly. But when you ask it to do real work inside your company, another problem appears: does your business actually know where anything is?

Imagine hiring someone brilliant. They can write. Research.

Analyse. Code. Spot patterns.

Work incredibly quickly. And learn how your business operates. Then, on their first morning, you give them access to:

three different versions of your price list, a CRM nobody has updated properly since March, hundreds of documents with names like `FINAL-v4-USE-THIS-ONE.pdf`,

two spreadsheets containing different customer information, a shared drive organised according to a system nobody remembers, and an employee who says:

Ignore what's in there. Ask Dave. He knows how we actually do it. How useful is your brilliant new employee now? This is roughly the problem businesses are beginning to encounter with AI.

The models are getting better. The information we're giving them often isn't.

ChatGPT knows a lot. It doesn't know your business.

This distinction gets missed surprisingly often. A frontier AI model may know an extraordinary amount about marketing, accounting, sales, software, management and almost any other subject you can think of. What it doesn't automatically know is:

your current pricing, which products you actually sell, how you deal with refunds,

what you promised a particular customer, which supplier you currently use, your internal terminology,

who approves what, where a project has got to, or the slightly strange process your company developed six years ago that everybody now treats as completely normal.

That's business context. And if you want AI to move from being a clever general-purpose assistant to something genuinely useful inside your company, it needs access to the right context.

This wasn't such a big problem when we were just asking AI to write things

If you're asking: Give me ten ideas for a LinkedIn post. the AI doesn't need to understand your entire organisation.

You can provide a little context in the prompt. But now imagine asking: Which customers should we follow up with this week?

That's different. The AI needs to know who your customers are. When they were last contacted.

What they bought. Whether there's already an open conversation. Whether somebody else on the team is dealing with them.

What counts as needing a follow-up. Possibly how valuable the account is. And perhaps whether there are reasons not to contact them.

The quality of the AI isn't enough. The quality of the information surrounding it starts determining the quality of the answer.

AI makes your information problem visible

Most businesses already have a data problem. They just don't necessarily call it that. You see it when somebody asks:

Where's the latest proposal template? And three people send three different files. You see it when the website says one price and the sales spreadsheet says another.

You see it when somebody leaves and half the knowledge about a process leaves with them. You see it when two systems contain different addresses for the same customer. You see it when somebody says:

Don't use that field in the CRM. We stopped updating it ages ago. Humans are remarkably good at working around this. We learn which spreadsheet to trust.

We know Jenny's version is normally the latest one. We remember that a particular customer always needs something slightly different. We ask someone.

We notice something looks wrong. AI doesn't automatically inherit that unofficial knowledge.

Your business probably has more than one version of the truth

This is one of the first things I'd look at before giving AI more access to a company. Where is the authoritative information? Take something simple:

What does this product cost? Is the answer in: the website?

the ecommerce platform? the accounting system? the CRM?

a spreadsheet? the sales team's proposal template? If those all say £500, brilliant.

If three say £500, one says £450 and another says £550, what should the AI believe? Humans may know that the spreadsheet is old. The AI needs some way of knowing that too.

This is sometimes called having a single source of truth. It doesn't necessarily mean putting absolutely everything into one enormous system. It means knowing which system or source should be trusted for a particular kind of information.

That's much more achievable.

AI doesn't need all your data

This is another misconception I'd avoid. The answer isn't: Put everything the company has ever created into the AI.

More isn't automatically better. If you give an AI: old policies,

duplicate files, superseded price lists, irrelevant emails,

unfinished drafts, outdated product information, and contradictory instructions,

you haven't necessarily made it more knowledgeable. You've made it harder to know which information matters. The useful question is:

What does the AI need to know to do this particular job? That's a very different exercise.

Start with the work, then find the information

Suppose you want AI to help answer customer enquiries. Don't begin by uploading the entire company drive. Map what a good employee needs in order to answer an enquiry properly.

Perhaps they need: current product information, pricing,

delivery information, returns policy, previous customer correspondence,

stock availability, and rules about when something should be escalated. Now we have something useful.

We know what information the AI needs. We can ask where each piece lives. We can check whether it's current.

We can decide which source wins when two disagree. And we can decide what the AI shouldn't see at all. That's a much better starting point.

The problem isn't always bad data. Sometimes it's hidden knowledge.

Some of the most valuable information in a business isn't stored anywhere. It's in people's heads. Ask how something works and somebody says:

It depends. Then they explain five exceptions that don't appear in the official process. Or:

We normally do it this way, unless it's one of these customers. Or: Technically that's the policy, but here's what actually happens.

That knowledge matters. An experienced employee has accumulated it over years. If you're asking AI to perform part of their work, you need to work out which bits of that judgement the system needs.

This is why introducing AI can uncover process problems businesses didn't realise they had. You start trying to explain the work to the AI and discover: We don't actually have an agreed process.

That's useful information in itself.

"Ask Dave" isn't a scalable knowledge system

Most companies have a Dave. Maybe yours is called Lisa. There is somebody who knows:

where things are, why something is done that way, which supplier to call,

what happened last time, which spreadsheet is correct, how to fix the strange problem,

and who actually needs to approve something. This works surprisingly well. Until Dave goes on holiday.

Or leaves. Or gets tired of answering the same question six times a day. AI creates an interesting opportunity here.

Not to replace Dave. To stop requiring Dave to be the company's search engine. If useful organisational knowledge can be captured and made accessible, AI can potentially help everybody find it.

That's good for the AI. It's also good for the business.

Your documents become part of the system

Businesses have traditionally thought about documents as things people read. Policies. Proposals.

Contracts. Guides. Specifications.

Meeting notes. Product sheets. Process documents.

With AI, those documents increasingly become machine-readable business knowledge. That means how they're maintained starts to matter more. Is this current?

Who owns it? When was it updated? Is there another version somewhere else?

Does it contradict another document? Would somebody know this was superseded? The PDF sitting forgotten in a folder may suddenly be information an AI uses to answer a customer's question.

That changes the importance of document housekeeping quite considerably.

The CRM problem gets harder to ignore

CRMs are a perfect example. Almost every company wants better information in its CRM. Almost every company also has people who don't update the CRM consistently.

Now imagine connecting an AI agent to it. The agent asks: When did we last contact this customer?

The CRM says February. The actual answer is yesterday, but the salesperson didn't log the call. The AI isn't necessarily wrong.

It's correctly using incorrect information. This distinction matters. A business can spend a lot of time trying to improve the AI when the actual problem is further upstream.

Garbage in, garbage out isn't quite enough anymore

We've used that phrase in computing forever. Bad data goes in. Bad result comes out.

With agentic AI, there's an extra step. Bad information can produce a bad decision. And the bad decision can produce an action.

An AI might not merely tell you the customer hasn't been contacted. It might send them another message. Update their record.

Trigger another process. Or tell somebody internally that they need chasing. As AI becomes more capable of acting, information quality becomes more consequential.

Then there's access

Suppose the information is perfect. There's still another question: Should the AI be allowed to see it?

A business might hold: employee information, financial records,

customer data, contracts, commercially sensitive information,

passwords or credentials, health information, legal correspondence,

payment details, and internal conversations. An AI agent doesn't need access to all of that simply because it works for the company.

Access should follow the job. If an AI is helping with product enquiries, perhaps it needs product and customer-service information. That doesn't mean it needs payroll.

The same principle we use for employees and software permissions applies here too. Give the system what it needs, not everything you have.

Clean data doesn't mean perfect data

This is important because "get your data ready for AI" can sound like an enormous transformation project. It doesn't have to be. You don't necessarily need to spend a year rebuilding every database before doing anything useful.

Start with one process. Identify the information that process depends on. Then ask:

Where does it live? Is it current? Is it complete enough?

Which source should be trusted? Who owns it? Who can access it?

How will it stay current? That's manageable. And it gives you something far more useful than vaguely deciding the company needs to "sort out its data".

Sometimes the right answer is not AI

This is another useful outcome. You may investigate a process and discover the problem isn't intelligence at all. Two systems contain conflicting information.

People aren't following the process. A field isn't being updated. Nobody owns the document.

The same information is being manually entered three times. In that situation, adding AI may simply add another layer to the mess. Fix the underlying problem first.

Then decide whether AI adds anything.

AI readiness isn't really about AI

This is perhaps the part businesses should pay most attention to. A company that has: clear processes,

well-maintained information, sensible access controls, useful integrations,

known sources of truth, and an understanding of who owns what is in a much better position to use AI effectively.

But those are also characteristics of a well-run digital business generally. AI hasn't suddenly invented the need for good information. It has made the cost of bad information more obvious.

The model may stop being the differentiator

There is another consequence of all this. Imagine two competing companies have access to exactly the same AI model. Company A connects it to organised, current, relevant business information.

Company B gives it a folder containing seven years of duplicated documents and a CRM nobody trusts. They technically have the same AI. They don't have the same capability.

As frontier models become more widely available, part of the competitive advantage may move away from simply which AI do you have? Towards: What can your AI reliably know about your business?

And then: What can it safely do with that knowledge?

Before you buy another AI tool, look at what you already have

I'd start here. Pick one process you think AI could improve. Write down the information a good person needs to perform it.

Find where that information currently lives. Work out which version is authoritative. Check whether it is actually maintained.

Identify the information the AI should not have access to. Then look at the process again. You may discover an obvious AI opportunity.

You may discover a simple automation would solve it. You may discover the process needs fixing first. All three are useful answers.

Because the objective isn't to get more AI into the business. It's to make the business work better.

AI is getting smarter. Your business still has to make sense.

The capability of AI will continue improving. Models will reason better. Agents will work for longer.

They'll use more tools. They'll connect to more business systems. And they'll become capable of taking more actions for us.

But every step in that direction increases their dependence on something much less exciting: the information underneath the business. The latest AI model can't tell which of your three price lists is correct unless there is some way for it to know.

It can't retrieve knowledge nobody recorded. It can't reliably follow a process nobody has agreed. And it shouldn't have access to information it doesn't need.

So before worrying about whether your business is "AI ready", there may be a more useful question: If a brilliant new employee joined tomorrow, could they actually find the information they needed to do the job properly? If the answer is no, I'd start there.

Where to go next

• AI agents aren't digital employees. Here's what they actually are. Once AI can access your business information, the next question is what work you allow it to perform.
• AI can now use the computer. That changes what businesses can automate. AI is increasingly able to work with the software businesses already use, even when there isn't a neat integration available.
• Don't start with AI. Start with the work. A practical way to find worthwhile AI opportunities inside an existing business.

Related: AI agents aren't digital employees. Here's what they actually are..