Book a demo
All notes

Verify3 min read

Data reconciliation across documents: how to find where two figures disagree

Data reconciliation across documents means finding every place a figure appears and explaining why the versions differ. Finance teams reconcile accounts this way. The same method works when the figure is a warranty period in a contract, a sales deck and a chat reply. This note gives a plain procedure you can run in a spreadsheet, the kinds of difference you will meet, and where a tool can help.

By the Corteal team

120 days90 daysabout 4 monthsDecided120 days

What reconciliation is

To reconcile two sources is to compare them on the same fact and account for every difference. A bank reconciliation compares the ledger with the statement. A data reconciliation compares two systems or two files. Across documents, one source is a paragraph in a PDF and another is a message in a chat. Nobody exports them to a table, so the first job is to build the table.

The procedure

  1. Define the claim. Write it as a term, a unit and a scope, for example "warranty period, in months, all customers".
  2. Collect every mention. For each one, note the value, the document, the page, the author and the date.
  3. Normalise. Put all values in one unit. "Two years", "24 months" and "104 weeks" are one value. "120 days" and "four months" are close but not the same.
  4. Compare. Group by value and count how many documents quote each.
  5. Classify each difference. Decide what kind it is, using the list below.
  6. Decide. A named person picks the figure that stands for a scope and a period, and records why.
  7. Check new drafts. Compare each new document against the decided figure.
1Define the claim2Collect every mention3Normalise units4Compare and count5Classify each difference6A named person decides7Check new draftsSteps 1 to 5 fit in a spreadsheet. Step 6 is a person.
Figure 1: The reconciliation procedure, step by step.

Kinds of difference

Not every difference is a conflict. Sorting them saves most of the work.

  • Different unit or rounding. The values agree once converted.
  • Different scope. One document is about enterprise customers and one about all customers. Both can be right.
  • Out of date. A newer document replaced the old one, but the old one is still in circulation.
  • A true conflict. Two current documents for the same scope disagree. A person has to decide.
Unit or roundingAgrees once convertedDifferent scopeBoth can be rightOut of dateRetire the old oneTrue conflictA person decidesSort the first three before anyone spends time on the fourth.
Figure 2: Sorting differences first leaves a short list for a person.

A worked example, with invented numbers. A data retention period appears three times: 120 days in a policy, 90 days in a contract and "about four months" in a chat reply. After normalising, the policy and the chat reply are close, and the contract is lower. The scope of all three is the same. That is a true conflict, and it needs a decision.

SourceSaidNormalisedVerdictPolicy120 days120 daysSame scopeContract90 days90 daysDiffersChat replyabout four monthsabout 120 daysSame as policyA retention period quoted in three places, all for the same scope.Invented sample data.
Figure 3: Three mentions of one claim, normalised. Invented sample data.

Where it goes wrong

Teams that do this by hand run into the same traps. Mentions hide in places nobody searches, such as a meeting recording or a thread. Dates go missing, so you cannot tell which version is newer. The decision lives in someone's head and is lost by the next quarter. And nobody checks the next draft, so the old number comes back.

What our Verify demo does

The Verify demo works on a fictional company, Halden Flowworks. The facts in it are provided, not extracted by a model: the list of claims, each value and the number of documents quoting it was written by hand, and the conflicts are planted. What is real is the part after the conflict is found. A claim has a state. A named person approves one figure with a scope and a review date, every other value becomes a deviation, and a pasted draft is checked against the approved figures. The note Who decides which figure is right goes through the code.

Corteal Verify is built to find the mentions in your own documents, chat and meeting recordings, so that nobody has to type the table in step 2. The finding step is not in the demo. A pilot is where it gets built and tested on your material.

Have two documents that quote different numbers? Start a pilot and send them.

Related notes