01What happens.
Internal systems read the organisation's documents as they are. A table read as a stream, a lost heading or a superseded version served as current can each introduce an incorrect fact. The next process retrieves that fact, the following process summarises it, and a later process combines it with correct information in a report. By the time the report reaches a person, it carries the system's authority without the document history that produced it.
02Why it matters.
The systems being built inside organisations depend on the state of the documents they are built on. The state of those documents is the condition AAAnow describes as Digital Estate DEBT™: assets and content that are disconnected, duplicated, outdated or insufficiently governed. MIT research found 95% of enterprise generative AI pilots deliver no measurable return (MIT NANDA, 2025); AAAnow observed a cause that study did not examine: the material the systems were built on.
For an executive relying on an internal system's answer, the question is not limited to whether the system works. It also asks whether the documents were read correctly and whether an underlying fault remained visible on the page.
03The case.
The mechanism is an observed AAAnow finding used in AAAnow's published writing on internal AI failure. The narrative that follows is illustrative and no organisation is identified. The chemicals company in use case 28 met the same mechanism.
The organisation had connected its internal AI and agentic systems to its document stores, so that procedures, reports, specifications and records could be retrieved and summarised on demand. Ingestion was automatic, so a document added to the store was available to the systems in whatever form it had been saved.
Some documents carried faults a machine could not read past. Tables reached the system as streams of figures, headings were lost to fonts that had not been embedded, reading order had never been set, and superseded versions sat beside current ones with no marking. When one of those faults changed how a document was read, the resulting error entered the system looking like any other material.
The systems processed the incorrect fact through several stages. Retrieval pulled it into an answer, summarisation retained it, and consolidation combined it with correct material from other documents.
Each process carried the incorrect figure forward while removing more of its source context. The report that reached the manager contained a figure no document in the store had stated.
The manager acted on the report, and when the figure was questioned, the trail led back through the consolidation, the summary and the retrieval to a document that was correct on screen. The fault could not be seen there, because it did not exist on the page.
The systems had been tested against sample questions before deployment. Everyone had assumed the documents were right, because they were the organisation's own record, and no process had read them as the systems would.
04Exposure.
A decision taken on a figure that no document stated, produced by the organisation's own systems from the organisation's own documents, is the organisation's decision. The exposure runs from the operational cost of the decision to the safety exposure where the documents carry procedures, and to the question the board is asked afterwards: what were the systems reading, and who had checked.
The spend on systems is a visible part of the loss where those systems cannot be trusted. Decisions taken on their answers create further exposure, but the organisation cannot enumerate them because the systems kept no record of which answers rested on a misread document.
05What AI Readiness changes.
Testing covered the systems, but it did not test how their source documents would be read by a machine. Reading the organisation's record as its systems read it, before ingestion and again as the record changes, shows which documents a machine cannot follow, which versions conflict, and which facts a system fed that record will produce. AAAnow provides that reading, evidences it, and places the organisation's position on the AI Readiness Maturity Scale™. On that reading the organisation fixes the documents that fail, marks the versions that are superseded, and builds its systems on a record that reads to a machine as it reads to a person.
The organisation makes the changes, and AAAnow provides the assessment, the evidence, the guidance and the review.
06Related use cases.
The internal AI project failed, was written off, and the documents that sank it were never examined
The chemicals instance of this mechanism.
Funding stopped on submissions that looked fine
The same class of document fault, deciding a public body's outcomes.
The results table AI read had its figures in the wrong columns
The table fault at the source, in a listed company's report.
07References.
- MIT NANDA, The GenAI Divide: State of AI in Business, 2025.
- Digital Estate DEBT™: Lawrence Shaw, AAAnow, 2026.
- 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 mechanism; the narrative is illustrative. No organisation is identified.