Copilot cited the file, not the line. Here is how to find it

You asked Copilot about a 40-page contract, a board deck, or a policy document. The answer is confident and reads well. You want to put one line of it in an email, so you need to see where it came from. The citation names the document. You open it and start scrolling.

That is the whole friction, and it is not a small one. Verification only survives contact with a working week if it takes seconds. A citation that names a file turns a ten-second check into a five-minute hunt, and a five-minute hunt does not happen at scale. So people stop checking, not because they stopped caring, but because the interface priced it too high.

This is worth stating with evidence rather than as a vendor jab, so here is a report from Microsoft's own Q&A site (February 2026, build bizchat.20260210.47.1). A user documented that Copilot's citations "both inline and in the final Sources section will be incorrect", in the sense that "Copilot will make a specific statement with an inline citation at the end, but that citation will not be relevant to the statement." Their example cited an unrelated GitHub file. They add that when challenged, Copilot acknowledges the error and then reissues a corrected answer containing the same citation, and that in their experience "the majority of response contain at least one incorrect citation."

One detail in that report matters more than the rest. They describe a rarer failure where "the citation looks correct… but the actual URL in the link will go to an unrelated page." That is the failure mode worth internalising, because it survives the check most people actually perform. Seeing a citation is not verification. Seeing that the citation supports the claim is.

One report is one report, and your experience may be better. Treat it as a reason to check rather than as a measured error rate.

You do not need a different tool to close this gap, and the habit is worth having regardless of what you use.

Ask for the quote, not the source. Instead of accepting the summary, ask: "quote the exact sentence from the document that supports this." Then copy that string and search it in the file with Ctrl+F or Cmd+F. Three outcomes, all useful. The sentence is there and says what the answer said, so you are done in seconds. The sentence is there but means something narrower in context, which is the most common real problem and the one a file-level citation hides completely. Or the sentence is not in the document at all, in which case you have caught a fabrication before it reached a client.

Scope it. Do not do this for everything, or you will stop doing it. Do it for the claim your decision rests on, any number you are going to repeat, and anything that surprised you.

An honest comparison has to include this, and Copilot's advantages are real ones we do not match.

It is already there. Copilot lives inside Word, Excel, Outlook and Teams, which means no new tool, no migration, and no asking colleagues to adopt anything. For most organisations that is the single biggest determinant of whether an AI tool gets used at all. It reaches across the whole tenant, so it can answer from mail, chat, and files together rather than from a library you curated. And it is permissions-aware against your existing Microsoft groups, so access control you already maintain is respected without a second system.

If your question is "where is anything across everything we own", that breadth is the right instrument. The trade is provenance: the wider the index, the harder it is to show you precisely which sentence in which artefact produced a claim, and the more likely a passage came from a half-finished draft rather than the signed version.

Tatsulok is narrower on purpose. Answers come only from your own documents and curated collections, and every claim links to the exact passage behind it, so checking is a click that lands on the sentence rather than a scroll through the file. When the sources do not answer, it says so instead of reaching for something plausible. That is the same three-outcome check described above, except it costs one click instead of five minutes, which is the difference between a habit and an intention.

We are not the right answer to "search my whole tenant." We are the right answer to "I have to sign this, so I need to see what it rests on."

This week's drill, no signup needed. Next time Copilot answers from a document, ask it to quote the supporting sentence, then search that string in the file. Note which of the three outcomes you get. Do it three times and you will know, from your own data rather than anyone's marketing, how much the citation was doing.