Cross-document figure reconciliation

What happens to a number before it reaches the other side

Your team drafts an answer to a due diligence request and asks the AI to write the first version. What comes back is not what the model produced. The model returns a structured draft in which every claim is attributed to a quoted passage from a named document. The application then checks, on the server, that each cited document is one the receiving side is permitted to read, and that each quoted passage is genuinely present in the stored text of that document. Citations that fail either check are dropped. If an affirmative answer is left with no verified citation at all, the answer text is discarded and replaced with a note routing the item for manual review. The unsupported version is never shown.

The draft is then scanned for money figures that appear in the answer but in none of its verified quotes. The same amount written in different formats counts as the same amount, so an amount written in the answer with a currency symbol and an abbreviated unit is not treated as different from the same amount spelled out in the quote. Unsupported figures are flagged to the person reviewing the draft. Flags do not block sending. They exist so your reviewer looks at the number before the counterparty does. Quotes that pass verification become links that open the source document at the quoted passage, so checking one is a click rather than a search.

Be exact about the scope. This compares an answer against its own cited sources. It catches a number the model asserted without support in the documents it quoted. It does not compare document A against document B.

Which documents the AI may quote when it answers the other side

Every AI feature that reads deal documents goes through one retrieval function, and that function applies a folder read check before a document is admitted to the candidate set. There is no separate, laxer path for drafting than for chat.

For an answer that will be shown to a counterparty the rule is stricter. The document set the model may quote is scoped to what that side is permitted to read, not to what the drafting team can see. Where more than one counterparty organisation can see the same request list, a document is admitted only if every one of them may read it. If the platform cannot positively resolve who owns the list, it denies rather than falling back to the wider view.

Retrieval runs against that deal's own database file. Each deal is a physically separate store served from its own subdomain, so a question asked in one room cannot reach another room's material. That is a property of where the files sit, not a rule the model is asked to follow.

Why a figure that does not tie is usually found late

The same number appears in several places in most processes: the information memorandum, the management accounts, the audited statements, the tax computations 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.

Most of that 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, and 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 in the negotiation of a representation. The cost then is rarely the correction. It is the loss of confidence and the extra round of questions.

What data room chat does when two retrieved documents disagree

Chat answers are grounded in documents retrieved for each question, under the asking user's own folder permissions. 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 is supposed to carry a titled link to its source.

In chat the citation rule is a prompt instruction rather than a verified one, because the server side quote verification described above runs on drafts headed for a counterparty, not on chat replies. Both documents also have to be in the set retrieved for that question. Open the cited documents.

What is not built, stated plainly

There is no reconciliation engine. Nothing 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 in 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. Your team can re-run enrichment across the room to recover those. Scanned PDFs with no extractable text are the weakest case, because there is no OCR in the pipeline. Retrieval admits any folder the reader is not blocked from, including view only and watermarked tiers, so an answer can quote text from a document the reader may read but not download.

The guardrails against prompt injection are prompt level: a real mitigation, not a security boundary. The boundaries that hold are the per deal database, the folder check before retrieval and the server side quote verification.

If you need assurance that every number in the room ties, that is tie out work and people still do it.

What a person still has to do with a difference

When an answer puts two figures in front of you 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 for.

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

Decide the consequence: update the model, raise a request to the other side, adjust a representation or a disclosure schedule, or record the item as immaterial and move on.

Record the decision where 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 one already in flight and, once a member of the deal team approves the match, attach the earlier request's confirmed answer documents to the new one. Buyers can dispute a merge. The scan covers the open backlog rather than a closed back catalogue.

How to get useful numeric answers out of the room

Ask narrowly. Name the metric, the period and the basis 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. Where the answer depends on something the other side holds, raise it as a request instead of resolving it internally on an assumption.

Verified citations, permission scoped retrieval and counterparty scoped drafting change how fast a team finds the two documents that disagree, and they stop an unsupported figure reaching the other side unnoticed. What a disagreement means is still your call. More on retrieval and permissions is on features and security.