Ask a system a question and it will give you an answer. That much is easy, and it is also the wrong thing to optimize for. The harder question is whether you can tell where the answer came from, and whether anyone with the standing to say so has checked it. Exalt local AI hardware is built around a specific answer to that problem: a four-step chain that runs from the question you type to the person who signs off on the result. Break any one link and the chain stops being worth trusting, no matter how fluent the output sounds.
The question
A good question to this kind of system is narrow and specific: which clause of the travel policy covers a trip extended for personal days, what the return-policy exception is for a damaged item received after 30 days, which revision of a drawing is current. It names a thing, a context, and usually a document type. That specificity matters because it is what lets the system narrow its search instead of reasoning from general knowledge. The system does not try to answer from what a general-purpose model was trained on somewhere else. It takes the question, decides which of the approved sources it should search, and retrieves the passages that are actually relevant to that question - not a summary of the topic, the specific passages.
A vague question - “how does travel reimbursement work” - still gets an answer, but a weaker one, because there is less for the retrieval step to narrow against. The quality of the first link in the chain sets a ceiling on everything downstream. That is normal: a person asking a colleague a fuzzy question gets a fuzzy answer too.
The approved source
“Approved source” means the system only answers from documents someone has deliberately put in front of it - the current manuals, the current schematics, the current policy pages, the current contract templates. It does not mean the documents are infallible. It means the set is known, bounded, and owned. Someone on your team decides what goes into that set, keeps it current when a manual is revised or a policy changes, and removes what should no longer be answered from. That is a curation job, not a one-time upload.
This is also where inspectability starts to matter in practice. Because the source set is fixed and known, when the system produces an answer you can ask a second question: is this actually in the approved set, and is it current? If the answer traces back to a superseded manual that should have been pulled, that is a finding about your document set, not a mystery about the model. A system that instead drew on whatever it had absorbed during training gives you no such handle - there is nothing to point at and nothing to correct.
The qualified output
The output the system hands back is not a clean paragraph of prose presented as fact. It is qualified: the specific passage or document it drew from, and some signal of how confident the match actually is. A well-supported answer, tightly matched to a specific passage, should look different from an answer stitched together from several loosely related fragments. Hiding that distinction behind confident phrasing is the failure mode this design is built to avoid.
- The answer carries a citation to the passage it came from, not just a document title.
- Weak or partial matches are marked as weak, rather than smoothed into fluent-sounding certainty.
- Where the approved sources do not contain an answer, the system says so instead of filling the gap with something plausible.
A confident paragraph with no provenance is not an answer you can act on - it is a guess wearing the clothes of an answer.
The human review
The last link is a person, and the design assumes that person is qualified to make the call - the finance lead who knows the account, the compliance reviewer who knows the policy, the engineer who knows which drawing revision actually shipped. Their job is not to double-check the system’s arithmetic. It is to accept or reject the answer using judgment the system does not have: knowledge of what changed in the business last week, awareness of an exception the document doesn’t mention, plain professional skepticism about a result that looks off.
This only works if the first three links did their job. A reviewer cannot meaningfully evaluate an answer that arrives with no source and no confidence signal - there is nothing there to inspect, only a conclusion to either trust blindly or reject blindly. Inspectability is what makes the review step a real check rather than a rubber stamp. That framing is consistent with how we describe human oversight more broadly on our responsible AI principles: a human in the loop is only doing real work when they have something real to look at.
What breaks when a link is missing
Each of the four steps exists to make the next one possible, which means the chain fails in a specific way depending on where it breaks. Skip the approved-source step and the system answers from wherever it can - including background it was trained on, which no one on your team chose or maintains. Skip the qualification step and every answer looks equally certain, so a well-supported answer and a thin guess are indistinguishable to the person reading them. Skip the review step and the system’s output becomes the decision, made by no one in particular and owned by no one at all.
None of this makes the system correct. It makes the system inspectable, and it makes the review step meaningful instead of theatrical. Those are different claims, and only one of them is one we make. For where the approved set is kept and every correction is recorded, see Exalt Ledger.