Citing a source is not the same as checking it

Artificial intelligence (AI) models can now do several things worth understanding before you buy a tool built on one. They can show where an answer came from, which is citing sources. They can read a scanned document or a photograph of a page, not just typed text. And a well-built system can reuse earlier work instead of asking the AI model to redo it, which keeps the running cost down. Each sounds like a small feature. Each changes how much you can trust, and pay for, what the tool gives back.

Citing sources and reading documents

When an AI model cites a source, it is pointing to the document, page or record it used to build its answer. This matters because an AI model can write confident sentences that are wrong. A citation lets you check the claim against the original, instead of taking the answer on faith.

The same models can also read documents that are not plain text: a scanned invoice, a photographed receipt, a contract kept only as a scanned file. This is useful for a business that still works on paper. It is also a new way in, if the document comes from outside the organisation or if the AI model is given more access than the task needs.

A report on a defence related breach shows how that access can go wrong. A contractor connected an AI coding agent to a secure government network, and the flaw it introduced ran undetected for 2 months before anyone found and fixed it (WIRED). The third-party tool behind the agent had already been flagged as high risk, because it could be tricked into leaking credentials, before it was used to take data out of the network (WIRED). Government contractors must report a cybersecurity incident within 72 hours of finding it, yet one firm reportedly did not report this breach at all (WIRED).

None of this means avoid the capability. It means ask who built the connection, what it can reach and who is told when something goes wrong.

Keeping the running cost down

Every time an AI model reads a long document or answers a question, it costs something to run. Caching is a way of storing the result of earlier work, a read document or a repeated question, so the system does not pay to redo it every time. A well-built assistant caches the parts of a conversation or a document that do not change, and only pays for the new parts.

Without caching, a tool that reads the same handbook or price list for every question becomes expensive fast. With it, the same tool can answer many questions for a fraction of the cost. Ask any supplier building you an assistant how it handles repeated reads, and what happens to your running cost as use grows.

What to ask a supplier: where will the connection reach, what happens within 72 hours of a problem being found (WIRED) and how is the cost of repeated reading kept down.

Sources

1. A Defense Contractor Gave Their Personal AI Agent Access to a Secure Pentagon System, WIRED, 30 October 2026