The mixing steps answered out of order, from a document that looked correct.

01What happens.

Think about what a PDF actually holds: the position of each line on the page, and little else. Unless the file has been built to say so, nothing in it says which line you are meant to read first. A person works that out from the layout without a second thought. A machine has only the order the file carries, and where nobody has set that order, the sequence it hands back is whatever it happens to recover.

02Why it matters.

A procedure's safety depends upon its steps remaining in sequence. Steps that are correct in themselves and wrong in order make a different procedure, and a technician following an AI answer on the floor has no way of seeing that the order has changed. The document that produced it is the company's own, approved, and correct on screen.

The same fault sits in any instruction where order matters: a start-up sequence, a shutdown, a decontamination, an emergency response, each of them approved on the page and read from the file.

03The case.

Both companies are anonymised, and both examples are based on observed AAAnow findings. The chemicals example describes one of the failures within use case 5.

The chemicals company published its mixing procedure, produced from a document that had been written, reviewed and approved in the usual way. On the page the nine steps ran down in order, numbered, each with its temperature and its timing. Whoever generated the file had not set the reading order, and no one producing safety documents had been asked to check, because a check of that kind had never been part of producing them.

The company's internal system, and the AI tools its customers used, read the file rather than the page. Each recovered the steps in the order it found them in the file's structure, which was not the order on the page. The returned sequence included step 1, then step 3, then step 9, then step 6 and then step 2, with the temperatures of 2 steps transposed along the way.

A worker asking for the procedure received an answer backed by the company's own document. The source offered the worker no reason to expect its sequence to differ from the approved page.

The manufacturer's operating instructions had passed human review and were published as a PDF. Read by machine, the instructions were reconstructed in the different order, and the review that had approved the page had no way of catching a fault that does not appear on the page.

In both companies, the people responsible had completed the established human review. The fault was in how the file had been built, and neither company's process included a step that read a safety document as a machine reads it.

04Exposure.

A procedure served in the wrong order under the company's name is a direct safety exposure, and the company's directors are accountable for the instruction in circulation. That the fault was invisible on the page changes nothing: the document was the company's, it was published as approved, and the machine read what it held.

The consequence is carried first by any person who follows the answer, and then by the company through its product and public liability, its licence position and the regulator's questions. Where the company's own internal system served the wrong order, the exposure is inside the company before it is outside it.

05What AI Readiness changes.

A review that reads the page cannot find a fault that lives in the file. The check the company lacked was of the file as a machine reads it. Does the order a machine recovers match the order on the page, do the temperatures stay with their steps, and what does AI return when asked for the procedure? We make alert to that check being required and evidences the result, document by document, before a procedure is served.

On that evidence the company rebuilds the files that fail, sets the order in the ones it publishes from now on, and checks what its own systems return for each safety-bearing procedure.

06Related use cases.