The conversation about enterprise AI has been about models for three years — which one, how large, hosted where, at what cost per token. Meanwhile the projects that stall almost never stall on the model. They stall weeks in, when somebody tries to assemble the inputs and discovers what the organisation actually knows about itself.
The finding is consistent and unflattering. The records exist. They are in several systems that disagree, they use fields whose meaning changed when a team reorganised, and the definition of the central entity — customer, order, site, employee — differs between the finance version and the operations version, for reasons everybody has forgotten.
The technology exposed a filing problem
None of this is new and none of it was created by AI. It is the accumulated residue of every migration, acquisition and reorganisation the company has been through, and it has been survivable because humans compensate. An analyst pulling a report knows that the field means one thing before 2021 and another after. A salesperson knows which of the three records for an account is the live one.
A model does not know any of that and cannot be told implicitly. Ask it to reason across those records and it produces something confident and wrong, and the wrongness is inherited from the inputs rather than invented — which makes it much harder to detect, because nothing in the output looks like a hallucination.
Lineage is the second layer and it becomes acute the moment anything is regulated or contested. If a system produced a decision, somebody will eventually ask what it was based on, and the answer needs to be more specific than the name of a warehouse. Organisations that cannot trace an input to a source cannot audit an output, which is why provenance became the feature buyers were paying for rather than a compliance afterthought.
The uncomfortable part is that fixing this is not an AI project. It is data governance — canonical definitions, an owner for each entity, retirement of duplicate systems, documented lineage — which is unglamorous, slow, has no demo, and was deprioritised for a decade precisely because it never produced anything visible. Companies now funding it are funding it under an AI budget line, which is arguably dishonest accounting and is the only way it was ever going to get paid for.
It also explains a puzzle that has bothered a lot of executives: why capable models produce disappointing results in one company and good ones in another with the same tools. The difference is rarely technical sophistication. It is whether the organisation had already done the boring work, and companies that invested in a clean warehouse years ago are getting a return they did not know they had bought.
Underneath all of it sits the oldest version of the problem, which is that the company does not have a data problem so much as a definition problem. Two departments reporting different revenue are not miscounting. They are counting different things, correctly, and no model resolves that — it only makes the disagreement faster and more fluent. Which is also why so many organisations still cannot tell whether their AI works: measuring it requires agreeing what right would look like.
Topics aidataenterprise
