We keep hearing that enterprises must get their data ready for AI. I wonder whether we are asking the wrong question. Perhaps we should first ask why we have made it so difficult for our own employees to find, understand and trust information.
For years, every new application promised better control. Every acquisition brought another set of systems. When information became difficult to find, we created another repository. Now we expect a large language model to make sense of it all.
But making a complicated system searchable is not the same as making it useful. We risk spending millions helping people navigate problems we should have fixed years ago.
The danger of making old information sound convincing
Consider a supplier approval. The contract sits in one system, the commercial exception in an email, and the supporting documents on a shared drive. A revised policy is published elsewhere. Someone asks why the supplier was approved.
An LLM might retrieve all four records in seconds. But which policy applied on the approval date? Was the exception formally authorised? Has the contract since been amended? The documents alone may not answer those questions.
A document can be accurate and still be misleading. It may have been correct when written but no longer reflect the business today. Unless we establish which information carries authority, AI can give outdated information new credibility. A confident answer based on the wrong version is potentially worse than a slow search.
Not every mess needs a massive cleanup
The usual response is to propose a large programme: classify everything, remove duplicates, standardise metadata, migrate repositories and build a common knowledge layer. Some of that work is necessary.
But it can become an expensive project with no clear end.
I would start somewhere smaller. Which business decision is being delayed because people cannot find or trust the information they need? Is supplier onboarding stalled by repeated checks? Are service teams working from conflicting contractual commitments? Are engineers unable to locate the history of a recurring failure?
Pick one important question. Trace the information required to answer it. Fix the ownership, gaps and handoffs around that question. Only then decide where an LLM adds value. We do not need to make every document AI-ready before solving a real business problem.
Preserve the decision, not just the document
Enterprises are often better at keeping records than preserving the reasoning behind them. A contract captures what was agreed, but not always why a clause was accepted. A maintenance report records an incident, but may omit the judgement that resolved it. An approval trail may show who clicked a button without explaining the exception.
This missing context matters. Useful enterprise AI needs connections between information, the business process, the decision taken and the person or role authorised to take it. Sometimes that means better metadata. Sometimes it means linking existing systems. Sometimes it means asking employees to record a short explanation at the point of decision.
None of these changes is particularly glamorous. All are more valuable than another impressive demonstration that cannot be trusted in daily operations.
Stop. Trust. Connect.
Before funding a new architecture, I would ask three simple questions.
What can we stop storing? Not everything should be retained forever. Identify duplication and information with no continuing purpose, while respecting legal, regulatory and business retention requirements.
What must we trust? Name the authoritative source for critical information. Make ownership, approval status, access rights and validity clear. If two records disagree, people and systems should know which one governs.
What should we connect? Link the few sources that matter to a business decision, rather than connecting every repository simply because the technology allows it.
This is not an argument for removing every layer of complexity. Multinational businesses have genuine regulatory, commercial and operational differences. Necessary controls should remain. The goal is to remove the needless effort employees face when working within those controls.
There is also a practical question of accountability. If an AI-generated answer influences a purchase, a customer commitment or a compliance decision, someone must be able to trace it back to the relevant records and challenge it. An answer without a reliable source, a date and a clear owner should not quietly become a business instruction. This is where simpler processes and sound governance reinforce each other: employees know what to check, and the system knows when it should not pretend to know.
Measure the work, not the demonstration
Don’t judge a successful pilot by how many documents an LLM can read or how fluent its answers sound. Did a supplier get onboarded faster without compromising checks? Did a service team resolve a dispute using the correct agreement? Did employees spend less time confirming which version to believe?
If those outcomes do not improve, we may have built an elegant interface to an unchanged problem.
We have spent years asking technology to accommodate the way our enterprises work, however complicated that may be. Perhaps AI gives us a reason to reverse that thinking.
Before asking what an LLM can do with our information, we should ask why our information became so difficult to use in the first place.
Otherwise, we risk spending millions making yesterday’s complexity work faster, instead of building a simpler business for tomorrow.