The Data Integrity Ladder: Five Rungs From Best-Guess Numbers to Every Figure Cited

Ask a finance leader how trustworthy their numbers are, and almost everyone answers the same way: “pretty good, a few known gaps.” That answer is nearly useless, because it is not describing the data. It is describing how the data feels from the outside. A dashboard built on unreconciled exports and a dashboard built on a ledger-tied model can render the exact same chart, with the exact same confident blue bar. You cannot tell which one is real by looking at it. You can only tell by tracing a number back to where it came from, and most teams have never actually tried.

That is the problem with informal self-assessment. It measures confidence, not integrity. What is missing is a ladder: a small number of distinct, checkable states a finance stack can occupy, ordered from worst to best, with a concrete test for which rung you are actually standing on. Below is that ladder: five rungs, from opaque numbers nobody can defend to every figure carrying a citation back to a source document.

Why this needs to be a ladder, not a checklist

Data quality checklists tend to list dozens of properties (completeness, timeliness, consistency, accuracy) and ask you to rate each one. That produces a scorecard, not a diagnosis. It also lets a team score reasonably well on nine properties while failing the one that matters most: can any given number be traced to its source and proven correct.

A ladder forces ordering. Each rung has to be strictly better than the one below it, and reaching a rung has to require satisfying everything the rungs beneath it require. That structure matters for AI specifically. An AI system sitting on top of your finance data does not average across your data quality properties: it inherits the weakest one. If lineage is broken anywhere, the AI will confidently cite a number it cannot actually prove, and nothing about the interface will tell you that happened. So the ladder is built around one question at every rung: if someone challenged this number right now, what could you show them?

The five rungs

Rung 1: Opaque numbers. Figures exist in reports, decks, and dashboards, but nobody can say with certainty where any individual number came from without manually reopening source files. Totals get typed into slides by hand. If two people pull “revenue” independently, there is a real chance they get different figures and neither can explain why. This is more common than most leadership teams admit, because it is invisible day to day: the reports still get produced on schedule.

Rung 2: Sourced but unreconciled. Every number can be traced to a system it came from: this pulled from the ERP, that from a bank export, that from a spreadsheet someone maintains. But the systems disagree with each other and nobody has reconciled the differences. You have provenance without agreement. This is arguably more dangerous than Rung 1, because it looks more rigorous. Each individual number has a defensible source. The problem only shows up when you compare two “correct” numbers that do not match.

Rung 3: Reconciled at a point in time. Someone has done the reconciliation work (tied the sub-ledgers to the general ledger, matched the bank feed, resolved the variances), but it happened once, for one close, and the reconciled state decays the moment a new transaction lands. This is the state most finance teams are in right at month-end close, for about a week, before it quietly erodes. The reconciliation was real. It just was not built to persist.

Rung 4: Continuously reconciled. The tie-out is not a monthly event; it is a standing property of the model. New transactions get matched and reconciled as they arrive, so the model is always close to current rather than periodically brought current. This is the rung where you can finally trust a number you pull on a random Tuesday, not just a number pulled the week after close.

Rung 5: Cited to the source document. Every figure does not just tie out arithmetically: it carries a pointer back to the specific transaction, statement line, or filed document it was built from, so a person or an AI can click through and see the underlying evidence in seconds. This is the rung that makes AI-generated answers auditable rather than merely plausible. A number without a citation is a claim. A number with one is closer to proof.

Rung	State	The test that reveals it
1	Opaque	Ask where a number in last week’s deck came from. If the answer requires “let me dig,” you’re here.
2	Sourced, unreconciled	Pull the same metric from two systems. If they disagree and nobody has a documented reason why, you’re here.
3	Reconciled at a point in time	Ask if the reconciled state from last close still holds today. If it has already drifted, you’re here.
4	Continuously reconciled	Pull a number today, not at close. If it is current and tied out, you’ve cleared this rung.
5	Cited to source	Click on any figure. If it opens the actual source document or transaction, you’re at the top.

Why teams misjudge their own rung

The honest reason most teams overstate their position is that the failure states are visually silent. A Rung 1 dashboard and a Rung 5 dashboard use the same charting library. Nothing in the UI degrades as you drop rungs: the reports still render, the numbers still look like numbers. The only way integrity failures surface is through consequence: a board question nobody can answer cleanly, an auditor request that takes three weeks instead of three hours, two departments presenting different figures for the same metric in the same meeting.

There is also a specific trap with AI adoption. Teams often believe that adding an AI layer for chat-based Q&A over their financials is itself a maturity upgrade, because it feels more sophisticated than a static report. But an AI wrapped around Rung 1 or Rung 2 data does not raise the rung. It just adds a more persuasive interface on top of the same unreconciled numbers, and a persuasive wrong answer is worse than an obviously rough one, because it invites more trust than it has earned. Climbing the ladder has to happen in the data layer, before the AI layer, or the AI inherits whatever rung the data was already on.

Climbing without a rebuild

The jump from Rung 3 to Rung 4 is usually the hardest, because it requires reconciliation to become a continuous process rather than a close-period ritual: matching transactions as they arrive instead of batching the work into one stressful week. The jump from Rung 4 to Rung 5 is more about instrumentation than process: every figure in the model needs a stored link back to its originating document, not just a correct value. Neither jump requires ripping out your ERP or accounting system. Both require a layer that sits on top of your existing sources, keeps them reconciled continuously, and stores provenance for every number it produces, which is a narrower, more buildable problem than “replace the finance stack.”

The takeaway

Most finance teams cannot name their rung on this ladder, because nothing in daily operations forces the question. The fix is not a bigger audit. It is running the five simple tests above against your own numbers this week: pick three figures from your last board deck and try to trace each one to a source document in under sixty minutes. Wherever that test stalls out is your actual rung, regardless of how the dashboards look.

If you want AI answering questions about your numbers, the ladder is the prerequisite, not an afterthought: Rung 4 and 5 are the floor for letting a model retrieve and reason over your data safely. Book a demo to see what a continuously reconciled, cited data layer looks like against your own numbers.

For the mechanics of getting there, see how verification works inside the platform, and full-year reconciliation for what continuous reconciliation looks like across a real close calendar. On the distinction between one reconciled foundation and the many views built on top of it, see single source of truth vs. sufficient versions of truth.