01What happens.
A funding body reads submissions by machine before a person sees them. A budget table that looks correct on screen can reach the machine with its headings and figures out of sequence. The machine then decides on figures that differ from the ones the applicant wrote.
02Why it matters.
The same exposure sits in any organisation that decides on documents by machine: grants, tenders, insurance claims, admissions and licence applications. The person who approved the process cannot see the fault, and neither can the person who submitted the document. It is discovered after the decision has been acted on.
The applicant has no reason to doubt an outcome that came through a formal process. When the outcome is wrong, it returns to the organisation as an appeal, a complaint or a claim.
The case figure
64% of the applications carried faults in the way their tables and data were structured.
AAAnow observed figure, belonging to this case only.
03The case.
A government funding body invited applications for project funding and asked applicants to submit their applications as PDF documents. The guidance set out what the documents needed to contain. It did not set out how the documents needed to be built.
Applicants prepared their submissions in the usual way, laying out budgets in tables, spreading multi-year figures across columns, exporting to PDF and checking the result on screen. To a person, each file was complete and correct.
The funding body’s system then read the submissions, and 64% of them carried faults in the way their tables and data were structured. To the machine, a budget table arrived as a row of headings followed by figures from several columns in one sequence. Key figures moved away from the headings they belonged to.
Funding for key projects was stopped on that reading. The decisions were consistent with the figures the system held, and the figures were wrong.
The system was blamed, which is the reasonable first conclusion when a new system produces bad decisions. Considerable staff time went into re-reading applications and checking figures. The budget on the page and the budget in the system disagreed, and the page was right, so the reviewers concluded the system was at fault. It was reset and restarted on the same documents, and it failed again almost immediately.
Nobody involved could see the problem, because it does not appear on screen. A PDF records where each item sits on the page, and it records which figure belongs to which heading only if it has been built to do so. A person works out that relationship from position, and a machine reads the structure inside the file. Where the structure is missing, the machine reads a different document from the one the applicant approved.
Projects that qualified for funding were left unfunded, and applicants who had done nothing wrong had no way of seeing why. Staff time was spent twice on the wrong cause. The funding body was left with a decision process it could no longer stand behind.
04Exposure.
A decision taken through an automated process remains the body’s decision, and the body answers for the outcome and for the review that followed. The same failure at larger scale is on record in Texas, where lawyers identified the state’s automated eligibility system as wrongly denying applicants by the thousands, with claimants unable to reach anyone who could explain the decision (The Contrarian, 2026).
A decision reversed after challenge costs more than a decision made correctly once, because the appeal has to be heard and the review paid for a second time. The larger cost falls on the confidence of applicants and of those who oversee the body, and a second failure on the same documents makes that confidence harder to recover.
A body that can show it tested the documents its system decides on carries that evidence into any challenge and into any discussion of AI-related risk with its insurer. In industry benchmarking, 57% name AI errors and misinformation as the top concern threatening AI adoption (Gallagher 2026, respondents). No outcome is automatic.
05What AI Readiness changes.
What the funding body could not see from the inside was how its system read a submission that looked correct on screen. AAAnow reads documents as AI reads them and evidences where a table, a heading or a figure arrives out of order, before a decision rests on it. With that evidence the body can set out the structure it requires of applicants, test the first submissions against it, and decide whether the system is ready to decide.
Your team makes the changes; AAAnow provides the assessment, the evidence, the guidance and the review.
06Related use cases.
Steps in the wrong order, temperatures transposed
The same fault in a chemicals company, where the consequence is physical safety.
The organisation’s own AI compounds its defective documents
Misread documents feed internal systems and the origin of the error drops out of view.
The documents that sank the internal AI project
A written-off spend that was traced to conflicting source documents.
07References.
- The Contrarian, The algorithm said no, 2026. www.contrariannews.org/p/the-algorithm-said-nocontrariannews.orgWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- Gallagher, AI adoption and risk benchmarking 2026. www.ajg.com/news-and-insights/features/ai-adoption-and-risk-benchmarking-2026ajg.comWe do not enable external links - please copy and paste, you are then certain as to the link being opened
- 64% is an AAAnow observed figure belonging to this case only.
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 organisation is identified.