1. Documentation gap: Which requirements haven't you documented?
You can check your documents against the standard you follow, say ISO 9001. Limit the analysis to the documents that relate to it, and ask AI to point out the parts of the standard you haven't addressed. The answer can be as specific as: You have no documentation for clause 4.2.
The limitation here is that the standard usually comes as a locked PDF, which the AI may not be able to read. If it can't, you can flip the analysis around: Read a section yourself, give the AI the points, and ask how you meet those exact points.
2. Review gap: Which documents should you review first?
Documents in a management system usually have a review date, and that information is already in the system. Imagine 1,000 documents, 450 of which are past their deadline. Instead of starting from the top, you can ask AI to sort them by how critical they are, so you start where an outdated document can do the most harm.
Fire safety instructions that haven't been updated in four years are critical. If you've added a new warehouse or swapped a door for a loading gate, they may no longer match the building, and that can be serious if a fire breaks out. The procedure for ordering food for a customer meeting can wait.
How does AI know what's critical? Again, it's statistics. A language model is trained on huge amounts of text and has learned that documents about danger to people and buildings carry a lot of weight. It also knows that many companies have fire safety instructions, which tells it something about how important they are.
The same principle works for risk assessments. If the system logs when an assessment was last reviewed, AI can find assessments whose review keeps getting pushed back year after year, and which may no longer reflect reality.
3. Process gap: Are procedures followed in practice?
A document can be approved without being followed in practice. AI can't see that from the document alone. But if you give it something to compare against, such as your own observations or the form employees use, it can help you in three ways.
First, you can ask it to compare the procedure with the form. If the procedure describes four steps, for example, does the form have the same four? Second, you can ask it to find patterns in your nonconformities. If the same nonconformity is logged again and again in one place, the cause may lie earlier in the process, with whoever handles the step before and may not be following the updated procedure. Then you know where to look. Third, if you have a suspicion, you can ask it to suggest how to collect data on it, and what you can adjust once you know the cause.
4. Reporting gap: What do your registrations tell you?
You can't fix errors you don't know about. That's why a reporting culture is a prerequisite. But the reporting itself can also tell you something. If all nonconformities come from 5 of 15 stations, you have 10 stations that never report. Another pattern could be that all registrations and updates pile up in the weeks leading up to an audit. AI can make patterns like these visible, because it can go through more data than you have time for.
The foundation: AI needs data it can work with
AI can't pull answers out of thin air. All four analyses above depend on structured data.
First and foremost, that means the same data goes in the same field. A date shouldn't read 04/03/2026 in one row and "the first Thursday in March" in the next, and an item number shouldn't be something you can write five different ways. Second, it means the data is linked. When a complaint is tied to the supplier in the system, instead of the supplier's name just being typed into a text field, you can see the same data from several angles. One employee can follow the individual case, while another can see that the supplier has 14 complaints in total. Those are the relationships AI can analyze.
That's exactly the structure AI needs to run these gap analyses. And this is where I see the strength of a system like ours: If you've built your management system in IPW, you get that structure as standard.
Not all data needs to go into the system, though, because not all of it tells you anything useful. Some production equipment takes 60 measurements a second, and the vast majority of them just show that everything is running as it should. Here, an agent can monitor the measurements and pass on only the ones that fall outside a set limit, say, 10 percent. Those then become nonconformities in the system, and you get registrations that are actually worth looking at.
