01What happens.
An internal AI system answers from the documents it is fed. Conflicting documents produce conflicting answers, while poorly structured documents can create errors that appear to come from the system itself. The project conclusion landed on the AI system even though the cause lay in its source material.
02Why it matters.
The board that funded the project received a familiar result: an AI system that could not be trusted, a budget spent, and a decision not to try again. MIT research found that 95% of enterprise generative AI pilots deliver no measurable return (MIT NANDA, 2025). AAAnow observed a cause that study did not examine: the documents the systems were built on.
Writing off the project without examining its source documents leaves the underlying cause in place after the budget has been lost.
03The case.
The company is anonymised, and the case is an observed AAAnow finding, used in AAAnow's published writing on internal AI failure. The documents are the ones described in use case 5.
The company set out to answer handling and storage questions faster, for its own staff and for its customers, and to do it without adding people. It procured a generic internal AI system and connected it to the document store, where the procedures and safety manuals lived. The store was the company's record, and the project treated it as the source of truth.
The store held several versions of the same instructions, with no version handling, some supplied by a third party, and some with a structure a machine could not follow. The system read the documents exactly as they had been stored. Asked a storage question, it answered from one version; asked again, it answered from another. The answers disagreed with each other, and where a document's reading order had never been set, the answer was wrong in a way that matched no document.
The project team tested the system, found it inconsistent, and reported the result. The system was doing what it had been built to do with the material it had been fed, and the conclusion reached was that the system was not fit for purpose.
The project closed and the spend was written off, and the budget for a second attempt was withdrawn with it. The document store was left as it had been found, still serving the same conflicting versions. The same conflicting versions carried on feeding the company's other systems and the AI tools its customers used.
The review that closed the project examined the system, the vendor and the integration. It did not examine the documents, because the documents were the company's own record and the record was assumed to be right. Nobody tested that assumption before the project was closed.
04Exposure.
The written-off spend is visible, but the wider consequence is the company's conclusion that AI cannot be trusted with its documents. That conclusion stops the next project before it starts. The documents themselves are the safety exposure use case 5 describes, and the failed project was the company's first evidence of it, read as evidence of something else.
For the board, the question is what was examined before the project was closed. A failure attributed to the system when the cause was in the company's own record is a failure the company will meet again, in the next system it builds on the same documents.
05What AI Readiness changes.
The project tested the system against the documents and never tested the documents. The check that was missing reads the company's record as a machine reads it, before a system is built on it. It shows which versions conflict, which documents a machine cannot follow, and what a system fed that record will answer. That check, and the evidence of what it finds, is what AAAnow provides.
Where the foundations fail, the company fixes the record first, so that the next system is built on documents that agree with each other and read correctly to a machine.
The company makes the changes, and AAAnow provides the assessment, the evidence, the guidance and the review.
06Related use cases.
Staff handled and stored chemicals on instructions the company had already replaced
The documents that sank this project.
The company’s AI amplified its own PDF failings
The general mechanism this project met.
The chatbot was replaced twice before anyone looked at what it was reading
The same wrong conclusion drawn in a local authority.
07References.
- MIT NANDA, The GenAI Divide: State of AI in Business, 2025: 95% of enterprise generative AI pilots deliver no measurable return.
- Observed AAAnow finding, anonymised.
Next step
Talk it through, in your terms.
The briefing sets this case in your organisation’s terms, and the discussion follows your questions.
Observed, anonymised. No company is identified.