When two documents disagree about a number

Why a figure that does not tie is usually found late

In most processes the same number appears in several places. It is in the information memorandum, the management accounts, the audited statements, the tax computations, a contract schedule, and the operating model. Each was produced by a different person, for a different purpose, often on a different basis. Revenue may be gross in one file and net of credits in another. Headcount may include contractors in one and not the next. An adjusted earnings figure carries add-backs that a statutory figure does not. Period cut-offs and currency conventions vary.

None of that is unusual and most of it is explainable. The problem is ownership. Checking that a metric agrees across every document in the room is nobody's single job, so it happens in fragments. Whoever builds the model ties out the model. Whoever drafts the disclosure schedules ties out the schedules. A difference neither of them looked at survives until someone on the other side reads two of those documents in the same sitting. That tends to happen in confirmatory diligence, in a purchase price adjustment discussion, or during the negotiation of a representation. The cost at that point is rarely the correction. It is the loss of confidence, the extra round of questions, and the time lost.

Why reconciling figures across documents is hard to automate

There are four separate problems here.

The first is extraction. Numbers live in scanned pages, in spreadsheets with merged cells and hidden tabs, and in footnotes. A system that cannot read the file cannot compare anything in it.

Identity. Two numbers can only disagree if they are supposed to be the same number. Deciding that requires resolving the metric, the entity or consolidation scope, the period, and the basis of preparation. Two figures for revenue that differ by the value of intercompany sales are not in conflict.

Normalization. Thousands against millions, local currency against reporting currency, fiscal year against calendar year, and rounding all have to be handled before a comparison means anything.

The fourth is adjudication. Many real differences are correct and explained somewhere, often in a note on the same page. A system that raises every arithmetic difference produces a queue the deal team stops reading in a day.

What CogniSuite does when two documents disagree

Data room chat answers are grounded in documents retrieved for each question, under the asking user's own folder permissions, from that deal's own database. When the retrieved set contains figures that do not agree, the system prompt instructs the model to say that the sources disagree, to give each figure separately, and to link each one to the document it came from with the date that document carries, rather than silently picking one. Every factual claim in an answer is supposed to carry a titled link to its source document, so the reader can open it and check.

That behaviour lives at the level of a single answer to a single question, and it depends on two things being true. Both documents have to be in the set retrieved for that question, and the model has to comply with the instruction. It is not a sweep of the room.

The one mechanical number check that actually exists

Drafts headed for a counterparty, on the request and Q&A workflow, go through server side verification before anyone sees them. The model returns a structured draft with its citations. The application then checks that each cited document is one the recipient side is permitted to read, and that each quoted passage is genuinely present in the stored text of that document. Citations that fail are dropped. If an affirmative answer survives with no verified citation left, the answer is replaced with a note routing it for manual review rather than being shown.

On top of that, the draft is scanned for figures that appear in the answer but in none of its verified quotes, with the same amount written in different formats treated as the same amount. Those figures are flagged to the human reviewer. The scope matters: this compares an answer against its own cited sources. It catches a number the model produced without support. It does not compare document A against document B.

What is not built, stated plainly

There is no reconciliation engine. Nothing in the platform extracts figures into a structured set, normalises units and periods, compares them across the room, stores a discrepancy as a record with an owner and a status, or raises a flag in the interface. Any surfacing of a conflict happens inside an answer to a question a person asked.

Several other limits bear on numeric work. Retrieval returns a small number of top scoring documents per question, so a conflict between two documents that are not both retrieved will not appear. Each document is represented by a single embedding built from a truncated version of its text, so a figure buried deep inside a long file may not be what drives retrieval. Enrichment runs in the background after upload and is best effort. A file whose processing fails is still stored and viewable but has no embedding, which makes it invisible to chat and search with no indicator on the file itself. The deal team can re-run enrichment across the room to recover those. Scanned PDFs with no extractable text remain the weakest case, since there is no OCR in the pipeline.

If you need assurance that every number in the room ties, that is tie out work and it is still done by people. Nothing here replaces it.

What a person still has to do with a flagged discrepancy

A flag does not decide anything. When an answer shows you two figures that do not agree, the work that follows is judgment. It runs roughly in this order.

Confirm the two figures are meant to be the same figure. Check the metric definition, the entity and consolidation scope, the period, the basis, and the currency. Many apparent conflicts end here.

Decide which source governs for the purpose at hand. An audited statement, a management pack and a signed contract carry different weight depending on what the number is being used for.

Find the explanation. Restatements, reclassifications, add-backs and timing cut-offs are usually documented somewhere in the same room.

Decide the consequence. That may mean updating the model, raising a request to the other side, adjusting a representation or a disclosure schedule, or recording the item as immaterial and moving on.

Record the decision where the rest of the team will find it, because the same figure will come up again in Q&A and in drafting. On the request side the platform can detect that a new request duplicates another request still in flight and, once a member of the deal team approves the match, attach that request's confirmed answer documents to it, so the same item does not get answered twice from scratch. The scan skips requests already marked answered or published, so it works across the open backlog rather than against a closed back catalogue.

How to get useful numeric answers from the room

Ask narrowly. Name the metric, the period and the basis in the question rather than asking for a number in the abstract. Ask for each figure to be listed with its source rather than asking which one is right. Open the cited documents, because the note that explains a difference is often close to the number. Make sure the source files are in the room and have been processed. Where the answer depends on something the other side holds, raise it as a request instead of resolving it internally on an assumption.

Grounded answers with real citations and permission scoped retrieval change how quickly a team can find the two documents that disagree. Deciding what the disagreement means is still your call. More on how retrieval and permissions work is on features and security.