Why AI answers need citations, and a one-minute routine to check them

Language models write fluent text. Fluency is not accuracy, and the two are hard to tell apart when you cannot see where a sentence came from. In business documents the typical failures are small and expensive:

• A plausible number that is not in the document. The report says revenue grew 7 percent; the answer says 9 because it blended in another quarter.
• A clause from a different contract. Two supplier agreements have different notice periods, and the answer quotes the wrong one.
• General knowledge dressed as fact. The handbook is silent on carry-over of vacation days, so the answer fills the gap with what is common elsewhere.
• A missing qualifier. The liability cap is stated correctly, but the next sentence excludes confidentiality breaches and the answer leaves that out.

Without a source, each of these looks exactly like a correct answer. The reader either rereads the whole document, which defeats the purpose, or takes the risk.

Not every reference is equally useful. A citation earns its place when it meets four conditions.

1. 1It is specific. Document name and page, not just "from your files". For slide decks, the slide number. For spreadsheets, the sheet and cell range.
2. 2It is per statement. Each sentence of the answer points to its own source. One citation at the bottom of a long answer leaves you guessing which part it supports.
3. 3It opens the passage. Clicking the citation takes you to the page with the supporting text highlighted, so checking takes seconds, not a search.
4. 4It has been checked. The cited passage must actually exist on that page and contain the quoted words. A citation that points to the wrong page is worse than none, because it gives false comfort.

ChatDocument is built around these four conditions. Every statement in an answer carries a citation, the viewer opens at the highlighted passage, and each cited passage is checked in code before the answer is shown: if the passage is not on the cited page or does not contain the quoted text, that sentence is dropped. You can read more on the cited answers from documents page.

A tool that must cite every statement cannot quietly fill gaps. If nothing in the documents supports an answer, there is nothing to cite, and the honest result is "the documents do not answer this". ChatDocument says that plainly and lists the closest passages it checked, with pages, so you can see what was searched.

This matters more than it sounds. In a due diligence review, "the data room contains no change of control clause for this customer" is a finding. In an HR question, "the policy does not cover this" is a reason to update the policy, not to improvise an answer for an employee.

Teams that use AI answers safely tend to follow the same short routine. It fits on an index card.

1. 1Read the answer once, fully. Note which statements you will actually rely on. Not every sentence needs the same scrutiny.
2. 2Open the citation for each statement you rely on. Read the highlighted passage and the sentence before and after it.
3. 3Look for qualifiers. Words like "except", "subject to", "unless", "provided that" and cross references to other sections change the meaning of a clause.
4. 4Match numbers exactly. Dates, amounts, percentages and day counts must match character for character.
5. 5Check the document version. The citation shows the file name. Make sure it is the signed version, the current policy or the final report.
6. 6Keep the page reference in your output. Write "60 days' notice (MSA, p. 14)" in your summary. The next reader can check in seconds.

For a contract question, this routine usually takes under a minute per answer. For a board pack with twenty figures it takes longer, but far less than rereading the source.

• Legal and contract review. A clause summary without a page is a summary nobody can sign off. See how teams use this in AI contract review.
• Finance. Figures from reports and filings must be traceable. A cited answer pointing to the sheet and cell range turns a question into an audit trail.
• HR and policy questions. Employees act on the answer. Citing the policy section protects both the employee and the company.
• Proposals and questionnaires. A security answer copied into an RFP must match the current approved wording, with a source to prove it.
• Customer support. Agents answering from documentation need the article and page they can share with the customer.

When you evaluate any tool that answers from documents, these questions separate a real citation system from a decorative one:

• Does every statement have its own citation, or only the answer as a whole?
• Does a citation open the document at the highlighted passage?
• Is the cited text checked against the page before the answer is shown?
• What happens when the answer is not in the documents? Ask a question you know is not covered and watch.
• Can you restrict answers to a selected set of documents, so nothing comes from outside?
• Are your documents used to train models? (In ChatDocument they are not; the AI model providers are bound by contract not to train on your content.)

If you want the technical background on how retrieval and grounding produce citations, read RAG explained for business teams.

A routine only helps if everyone uses it. A few practical steps for team leads:

• Agree that any AI answer quoted in a client email, contract memo or report carries its page reference.
• Share answers by link rather than by copy and paste, so the reader sees the citations too.
• When someone finds a citation that does not support the statement, report it and keep the example. It sharpens everyone's reading.
• Treat "not in the documents" as a valid result in reviews, not as a reason to ask a general chatbot instead.
• Keep a short list of the questions your team asks most often, with the cited answers. New colleagues learn where the authoritative passages are in their first week.