A Chart Is a Claim

Correct numbers can still give an incomplete picture

Imagine a quarterly board presentation, at a hypothetical company, showing that adjusted EBITDA increased 20 percent. The underlying figure is correct and ties to the supporting schedule. The chart appears to demonstrate improving performance.

Now suppose the same hypothetical company’s GAAP net loss widened over the same period because of charges excluded from adjusted EBITDA. That does not make either measure inherently false. It means they answer different questions. A presentation showing only the first can leave the audience with an incomplete picture of financial performance.

The problem is not confined to adjusted earnings. A working-capital trend can change if an entity is excluded; an expense chart can look flat or alarming depending on the axis; a forecast comparison can improve simply because the baseline shifted. Tie-outs validate data. They do not, on their own, validate the message created by the measure, comparison and visual design.

What AI changes

In a conventional chart, an analyst chooses a dataset, measure, period, filters, calculations and display. AI tools can compress many of those decisions into a single request: “Show whether profitability improved over the past four quarters.”

That request leaves open important questions. What counts as profitability: GAAP net income, operating income or adjusted EBITDA? Which version of the forecast or actuals is authoritative? Which entities and adjustments are included? Is the comparison quarter-over-quarter or year-over-year?

An AI assistant may make reasonable assumptions, ask for clarification or use context already available in the environment. The risk is not that AI is proven to make more chart errors than people. The risk is that a polished output can be produced before the reviewer has noticed, documented or approved the assumptions that determine its meaning. Microsoft’s help pages for Copilot in Excel say Copilot can create a chart from a description of the chart the user wants and tell users to review, edit and verify anything Copilot creates. The FAQ page also advises avoiding Copilot for decisions in sensitive areas such as finance, legal or medical topics.

Where presentation becomes a disclosure obligation

The U.S. Securities and Exchange Commission (SEC) has explicit presentation requirements for certain non-GAAP financial measures. For SEC filings and earnings releases furnished under Item 2.02 of Form 8-K, the most directly comparable GAAP measure must be presented with equal or greater prominence, alongside required reconciliation and other applicable disclosures. SEC staff guidance specifically lists, among its examples of a non-GAAP measure being more prominent than the comparable GAAP measure, presenting charts, tables or graphs of the non-GAAP measure without charts, tables or graphs of the comparable GAAP measure with equal or greater prominence.

The underlying idea is straightforward: an adjusted result should not visually eclipse the comparable reported result where those rules apply.

In a settled enforcement action, the SEC found that ADT had not given comparable GAAP measures equal or greater prominence in two earnings releases. ADT consented to the December 2018 order without admitting or denying the findings and was ordered to pay a civil money penalty of $100,000. That case involved headlines and highlight bullets, not charts; it supports the importance of emphasis and presentation, not a claim of chart-specific enforcement.

The legal scope matters. Regulation G applies whenever an SEC registrant publicly discloses material information that includes a non-GAAP measure, including disclosures outside filings; it requires the comparable GAAP measure and a reconciliation and prohibits a materially misleading presentation, subject to exceptions. It does not itself impose Item 10(e)’s equal-prominence rule. Whether the latter applies to a particular presentation depends on whether it is filed with the SEC or furnished under Item 2.02 of Form 8-K, and on other relevant facts. Whether visual emphasis alone can make a public non-GAAP presentation misleading under Regulation G’s general prohibition is a question for securities counsel. Internal board and management charts that are never filed, furnished or publicly disclosed fall outside these particular SEC presentation requirements, but a chart reused in an earnings release, filing or public presentation does not.

The SEC requirements are therefore a defined legal obligation in some circumstances, while a broader chart-review standard for internal finance reporting is a control practice CoE | AI Cost & Governance recommends.

The control gap: numbers versus presentation

Finance teams routinely reconcile balances, review journal entries and check reporting schedules. Yet a chart built from controlled data still involves further choices:

A finance reviewer should be able to answer two separate questions: Are the numbers right? and Does this chart fairly communicate what those numbers support?

A practical release standard for material charts

CoE | AI Cost & Governance recommends a proportionate release standard for charts used in board materials, investor communications, significant forecasts and other consequential decisions. It is not presented as a universal legal requirement.

PCAOB Auditing Standard 1215 offers a useful analogy: it requires audit documentation that allows an experienced auditor with no previous connection to the engagement to understand the procedures, evidence and conclusions. AS 1215 binds auditors, not corporate chart preparers; the proposed chart controls borrow its reproducibility principle, not its legal force.

Putting the standard into practice with AI

A useful implementation is an AI Financial Analysis Audit Record attached to the reporting workpaper. It should capture the original request, materially consequential follow-up instructions, the specific data/version used, transformations, validation results, human changes and approval. A prompt-only log is insufficient: the same instruction can yield different results across datasets, context, tool versions and manual edits.

For example, a stronger instruction than “Show profitability improving” is:

Using the approved quarterly reporting dataset identified as [dataset/version], chart GAAP net income and adjusted EBITDA for the last eight completed fiscal quarters. Preserve all entities in the approved consolidation scope. Use the approved adjustment bridge for adjusted EBITDA. Flag periods when the two measures move in different directions. First provide the exact plotting table, with every value that will appear in the chart, its period, entity scope, units and source reference. Then provide a script that renders the chart from that table, with axis settings stated. List the exact filters, calculations and any assumptions that need confirmation. Do not invent missing data, and do not produce a chart without its underlying table.

This is a starting instruction, not a verification control. The reviewer must still check that the selected dataset is authoritative, the EBITDA bridge agrees to approved support, the comparable measure is appropriate, the chart displays both measures fairly, and any disclosure-specific requirements are met.

Enterprise tools may retain parts of the interaction automatically. For example, Microsoft’s Purview documentation, last updated August 26, 2026, says that where an organization has auditing enabled, user interactions with Copilot, including in Excel, are logged automatically. Microsoft says those records show which user interacted, when and where, and include references to the files or other resources Copilot accessed to generate its responses. Availability and retention depend on the organization’s auditing and retention settings. Confirm what is actually captured in your own environment before relying on the logs. Such logs should not be assumed to preserve a complete analytical workpaper or every manual revision.

Keep completed audit records within your organization’s approved information-governance environment, and do not upload sensitive data to an unapproved AI tool.

How to build and check an AI-assisted chart

The most dependable pattern keeps the numbers and the drawing separate, so each can be checked on its own. No chart should be accepted without its underlying table.

1. Produce the plotting table first. Ask the tool for the exact table behind the chart: every value that will appear, with its period, entity scope, units and source reference. Nothing should appear in the chart that is not in the table.
2. Tie the table to approved support before anything is drawn. Agree totals to the reporting package or trial balance, adjusted measures to the approved bridge, and period labels to the fiscal calendar. This is the step that catches a wrong extract, a dropped entity or a filter nobody asked for. Until it ties, an AI-produced table is a draft.
3. Render the chart from the table with something that can be rerun. A short script that reads the table, such as a Python matplotlib script, or a native spreadsheet chart bound to the table’s cells, both draw the chart mechanically from the reviewed values. Both also leave the presentation choices (axis range, sort order, labels, colors) written down where a reviewer can read them. A chart that exists only as a generated image has nothing tying its bars to a number, and should not be used for reporting.
4. Regenerate it independently. The reviewer, or a second preparer, reruns the saved script or workbook against the saved table and should get the same chart. If it cannot be regenerated, it does not meet the reproducibility control above.
5. Check the drawing against the table. Confirm the number of data points, the labels, units and signs, and the scale. A bar chart whose value axis does not start at zero exaggerates differences, and two measures on separate axes can make unrelated movements look connected.
6. Check the message, not only the values. Ask what the chart leaves out, such as the comparable reported measure, a contrary indicator or a change in scope, and whether the title claims more than the table supports.
7. File the package together. Keep the table, the script or workbook, the final image and the audit record as one set, so an approved chart can always be traced back to approved numbers.

The tie-out in step 2 carries most of the weight. A script and a table only prove that the chart matches the table; they do not prove the table is right.

Check the data and the picture separately.

A deck that ties to the model has passed a consistency check

AI can now go from spreadsheet to chart to board slide in one run. That makes the set consistent, and it lets one error travel through all of it. If the model carries an invented assumption, the deck can tie out to it perfectly and still be wrong.

The slide can also add an error of its own.

What finance leaders can do now

1. Identify high-impact charts. Start with external financial communications, board packs, audit committee materials and analysis supporting significant decisions. Apply applicable disclosure rules where relevant.
2. Require an explanation of the chart’s construction. Confirm the source, measure, filters, calculations, comparison and visual design. Ask for the plotting table with every chart, and keep the table and the script or workbook needed to recreate material outputs.
3. Extend the review to AI-generated work. Record material instructions and system context, review transformations and outputs, and approve a clearly identified final version. Apply lighter controls to exploratory work and stronger controls to consequential reporting.

Approve the message as well as the numbers

AI can accelerate financial analysis. That makes it even more important to review the decisions that sit between source data and the finished chart.

Finance already controls the numbers it reports. It should also be able to explain and defend the story those numbers are made to tell.