Not the model. Everyone uses much the same frontier models, and the good ones are largely interchangeable for this work. The difference is what the model is standing next to.

If the specification lives in one system, the batch records in another, the certificates in a shared drive and the deviations in a third system, then a model attached to any one of them can only see a fraction of the batch. It can summarise. It can't check anything, because the thing to check against is in a system it can't reach.

That's the whole argument, and it's an architectural one. It also explains why bolted-on AI consistently produces summarising and searching features: those are the only jobs available to it.

Ask for a failed result, then ask what the system did about it: which batch, which specification, which open production runs, which customers, which documents now need reissuing. A connected system answers immediately because it's a query. A bolted-on system produces a thoughtful paragraph about what you should probably check.

Second test: ask where an answer came from. A grounded system points at the record. A model with no record to point at will describe its reasoning instead, which is a different thing wearing similar clothes.

No, and it's worth being fair about it. Search over your own quality documents is genuinely useful, and a summarising assistant saves real time in a business drowning in PDFs. If your systems work and the complaint is that nobody can find anything, that may be all you need.

It just isn't the thing that removes the reconciliation work. That requires the underlying record to be shared, which is an integration project rather than a feature. The full version of this argument, and what the alternative looks like, is set out in why an AI operating system.