What changes when the system is given documents, and what does not

The most common upgrade a manager will be offered is to point the system at the firm's own material: past cost plans, the commission's drawings, the tender pack, the firm's standard clauses. It is usually a genuine improvement and it is routinely oversold. You need to be able to say precisely what it fixes.

Two of those deserve expansion because they are where the expensive failures live.

Retrieval matches wording, not authority. A search across your documents returns the passages that most resemble the question. Resemblance is not the same as governing. The clause that actually controls a payment term may be phrased in language that shares no vocabulary with the question, sitting in an amendment schedule, while a superficially perfect match sits in a superseded revision or in the wrong contract entirely. Deciding which document governs is judgement work, it remains yours, and no amount of document access changes that.

A citation suppresses checking. This is the counter-intuitive one. Attaching a reference to a sentence makes the sentence look verified, so readers verify it less. The failure mode that costs the most is therefore not a fabricated source; it is a real, correctly cited source that does not say what the sentence claims. That error passes every check a busy reviewer performs, because the check they perform is "is there a reference", not "does the reference say this". If your firm introduces cited output, the review standard has to change with it: spot-check what the citation says, not that it exists, and sample enough of them to have a rate.

Worked example 7: the superseded revision

A tender pack for an GBP 8m scheme contains preliminaries at revision C and revision E. A surveyor asks the system what the payment terms are. The system returns a clause from revision C, because its wording matched the question more closely than the reworded clause in revision E, and it cites the document and page correctly. The answer is accurate to a real source, properly referenced, and wrong.

The control that costs 0.2 hours is not a technical control. It is a rule that any output referencing a document must be checked against the revision register before use, and the register is a document the firm already maintains for entirely unrelated reasons. That is the pattern to look for: the cheapest control for a novel failure is usually an existing professional discipline applied at a new point, not a new system. Across 340 live commissions, a 0.2-hour check applied to, say, two document queries a commission a year is 136 hours, GBP 5,576 at cost. One instance caught before the third row pays for it many times over.