Data Verification Checklist for AI-Generated PPTs: A Practical Workflow

    AI-generated presentations can look complete, polished, and professionally analyzed while still containing unreliable data. This article outlines a practical review process for teams that use generative AI to draft slide decks: verify sources, confirm metric definitions, check time validity, recompute calculations, review chart presentation, cross-validate key figures, and label data by confidence level before delivery.

    AI-generated presentations often have a complete structure, consistent formatting, and analysis language that looks professional. This can easily create a false impression: if a slide deck looks enough like a report, the data inside it should be trustworthy.

    In reality, the visual quality of a presentation and the reliability of its data are two completely independent things. A model can quickly generate titles, charts, conclusions, and industry terminology. It can also package an uncertain number into a “professional conclusion” with two decimal places. If the user does not verify the content further, this material may move directly from draft to presentation, and may even affect budgeting, procurement, investment, and business decisions.

    Therefore, reviewing an AI-generated presentation cannot be limited to checking typos and layout. It requires a review process designed specifically for data. This article organizes that process into five core layers: source, definition, time, calculation, and presentation.

    1. Start with a basic fact: online access does not mean correctness

    Using intelligent tools with retrieval capabilities can help supplement research, but “able to search” does not mean “search results are inherently trustworthy.”

    At least the following risks exist in the retrieval process:

    1. Secondary republishing causes error propagation
      One article may cite another article, which in turn cites an earlier source. On the surface, multiple web pages may support the same number, but they may all originate from the same original source.

    2. Context is truncated
      A page may show only that “the market size reached a certain value” without explaining whether this refers to the global market, a regional market, or a specific segment. It may also fail to state whether the metric is revenue, shipments, users, or orders.

    3. Year and publication date are confused
      A report published in the current year does not mean the data represents the current year. Some materials are published in one year but cover the previous year. Some forecast data may also be misread as actual data.

    4. Forecast values are treated as facts
      Expressions such as “expected,” “projected,” “forecast,” and “calculated compound growth rate” cannot be used interchangeably with historical data that has already occurred.

    5. The model fills in missing information
      When source material is incomplete, a model may add a plausible but unverified number based on language patterns. This content often reads smoothly, but its source cannot be traced.

    The correct role of online search is therefore to “help discover candidate sources,” not to “replace human judgment.”

    2. The first gate of data verification: confirm that sources are real and traceable

    Any data intended for a formal presentation should answer the following questions as far as possible:

    • Who published the data?
    • What is the original material called?
    • When was it published or updated?
    • What is the statistical scope of the data?
    • Can the specific chapter, table, or page in the original text be located?
    • Is the material original statistics, or a restatement of other sources?
    • Is there a public methodology description?

    If these questions cannot be answered, the data should not be used directly as a definitive conclusion.

    1. Prioritize original sources

    Source credibility can generally be assessed in the following order:

    The first priority is original material such as government departments, industry associations, public financial disclosures, formal research reports, and company public documents. The second priority is professional media, research institutions, and data platforms with clear citation relationships and editorial standards. The third priority is personal articles, marketing copy, content aggregation pages, and visual or text material without sources.

    This does not mean secondary sources are necessarily wrong. It means secondary sources must be traced further. If a number appears in only one article, and the article does not provide the original report, statistical method, or time range, it can at most be treated as a lead to be verified.

    2. Do not treat search result snippets as evidence

    Search result snippets are usually automatically extracted fragments and may lack qualifying conditions. For formal citation, open the original text, review the full context around the number, and record the following information:

    1. Data name:
    2. Specific value:
    3. Unit:
    4. Statistical object:
    5. Time range:
    6. Geographic scope:
    7. Source name:
    8. Location in original text:
    9. Access or retrieval date:
    10. Verification status:

    If these fields cannot be recorded, the data has not reached the level where it can be cited directly.

    3. The second gate: check statistical definitions, not just the size of the number

    For the same metric, even if the values are correct, different statistical definitions can lead to incorrect conclusions when they are compared side by side.

    For example, “market size” may have multiple definitions:

    • Calculated by sales revenue;
    • Calculated by shipment volume;
    • Calculated by the number of active users;
    • Calculated by transaction value;
    • Calculated separately by software, hardware, or service revenue;
    • Whether related businesses, value-added services, and overseas markets are included.

    When reviewing a presentation, it is advisable to break each key number into five parts: “value + unit + object + scope + time.” Only when all five are clear does the number become comparable.

    A simple definition checklist

    Check item Question to confirm
    Metric name Is it revenue, profit, users, or penetration rate?
    Unit Yuan, ten thousand yuan, hundred million yuan, percentage, or quantity?
    Statistical object Enterprise, product, user, order, or region?
    Geographic scope Global, national, provincial, municipal, or a sample area?
    Time range Monthly, quarterly, annual, or cumulative?
    Data nature Actual, estimated, forecast, or survey result?
    Tax treatment Do revenue metrics include taxes?
    Comparability Are definitions from different sources consistent?

    Pay particular attention to indicators such as “market share.” Market share should at least specify the numerator and denominator:

    1. Market share = relevant business size of the object ÷ total market size under the same definition

    If the numerator is calculated by revenue but the denominator is calculated by shipment volume, the conclusion does not hold. Even if both are calculated by revenue, they cannot be compared directly if one includes service revenue and the other counts only product sales.

    4. The third gate: check time validity and avoid “old data, new conclusions”

    AI-generated content often combines data from different years in the same chart, leading readers to assume that the data belongs to the same period.

    It is advisable to add a clear time label to every number, for example:

    • Actual data for 2022;
    • Preliminary statistics for 2023;
    • Forecast data for 2024;
    • Cumulative data as of a certain month;
    • Future values extrapolated from historical data.

    Also note that the publication date and the statistical year are not the same. For example, a report published in a given year may cover data from the previous year. An annual report may also cover only the period up to a certain quarter.

    For growth rates, the base period and calculation interval must also be confirmed. Common growth rate calculations are:

    1. Growth rate = (current period value - previous period value) ÷ previous period value × 100%

    For multi-year compound growth, the usual formula is:

    1. Compound growth rate = (ending value ÷ beginning value)^(1 ÷ number of years) - 1

    Do not mix single-year growth rates, cumulative growth rates, and compound growth rates in the same chart. More importantly, do not keep only the percentage while removing the start and end years.

    5. The fourth gate: reverse-verify calculations generated by the model

    Numbers generated by a model may come from three sources:

    1. Direct citation from external material;
    2. Calculation based on existing data;
    3. Inference or completion when data is missing.

    The first type requires source checking. The second type requires recalculation. The third type must be explicitly marked as an assumption or pending verification and must not be disguised as fact.

    1. Recheck totals

    If a page shows regional shares, product revenue, or departmental costs, check whether the total is reasonable. Due to rounding, individual items may produce a small deviation when added together. If the deviation is obvious, however, the figures need to be checked again.

    2. Recheck growth relationships

    For example, if a page says a business grew from 100 to 130 and labels the growth rate as 35%, this should be checked immediately. According to the formula, the actual growth rate is 30%. This type of error is not complex, but it is very common in automatically generated pages.

    3. Recheck unit conversions

    Errors in conversions among “ten thousand yuan, million yuan, and hundred million yuan” can directly cause order-of-magnitude deviations. It is advisable to use a base unit consistently in the data table and apply formatting only at presentation time, avoiding manual conversion at multiple stages.

    4. Recheck whether charts and body text are consistent

    Common problems include:

    • The body text says 18%, but the chart label shows 20%;
    • Chart ordering does not match the text description;
    • Bar charts use different vertical-axis scales;
    • A missing year in a line chart is connected by a continuous line;
    • Legend names do not match data table fields.

    A presentation review should not only check the “data table.” It should also check whether the data is expressed correctly after it enters the chart.

    6. The fifth gate: be wary of numbers that are “too precise”

    In fields without a public statistical basis, numbers with two or more decimal places are not necessarily more professional. They may actually be more suspicious.

    For example, if a page from an unknown source gives a “market share of 17.43%” but does not explain the sample size, statistical definition, time period, or calculation method, the digits after the decimal point only increase the sense of certainty. They do not increase credibility.

    To judge whether precision is reasonable, consider three points:

    • Does the original material itself provide the same level of precision?
    • Is the data collection method sufficient to support that precision?
    • Does the page explain the calculation method and margin of error?

    If the original material gives only “about 17%,” it should not be rewritten as “17.43%.” If a number is estimated from multiple public data sources, expressions such as “estimated,” “approximately,” or “calculated from public data” should be used.

    7. Establish a dual-source cross-validation mechanism

    For core numbers that will appear on title slides, conclusion pages, ranking pages, and decision-recommendation pages, it is advisable to find at least two independent sources for comparison.

    The “two sources” here should not simply be two articles that republish each other. They should meet the following conditions as far as possible:

    • Different publishing entities;
    • Different original data or statistical methods;
    • Clear publication time and data scope;
    • Basically consistent metric definitions;
    • Ability to explain the difference between the two.

    If the two sources disagree, do not simply choose the number that better matches expectations. Instead, determine whether the difference comes from:

    • Different statistical periods;
    • Different statistical scopes;
    • Different samples;
    • Different metric definitions;
    • One is actual data and the other is forecast data;
    • Republishing or calculation errors in one source.

    Data with only a single source may be retained, but it must be marked internally as “pending verification,” and the wording on formal pages should be softened. For example, use “public materials indicate,” “one report estimates,” or “preliminary calculation” instead of presenting it as an absolute conclusion.

    8. Add a credibility label to every data point

    To avoid repeated judgment during multi-person collaboration, add a status field to the data table:

    1. Confirmed: Reliable original source exists, definition and time are clear, and cross-validation is complete
    2. Basically credible: Source is clear, but limitations remain in definition or time
    3. Pending verification: Only a single source exists, or the original source cannot be located
    4. Estimated value: Calculated from public data, with the calculation method recorded
    5. Assumption: Used for demonstration or scenario calculation, not representing actual conditions
    6. Prohibited: Source unknown, definitions conflict, or recalculation is impossible

    These labels do not all need to appear in the final presentation, but they should be retained in the draft, data dictionary, or review record.

    9. Complete pre-delivery checklist

    Before submission, a quick check can be performed in the following order:

    Source checks

    • Does every key number have a source?
    • Can the source be opened or located?
    • Is original material available?
    • Has republished content been mistaken for independent evidence?

    Definition checks

    • Is the metric name accurate?
    • Are the numerator and denominator consistent?
    • Are the unit, region, sample, and time clear?
    • Do different pages use the same standard?

    Calculation checks

    • Are totals correct?
    • Have growth rates been recalculated?
    • Is there double counting in percentages?
    • Have unit conversions caused order-of-magnitude errors?

    Presentation checks

    • Are forecast values clearly labeled?
    • Are estimated values described with appropriate wording?
    • Is decimal precision written too high?
    • Are charts, body text, and conclusions consistent?

    Risk checks

    • Is there content with a “single source but written as established fact”?
    • Are there unexplained, unusually precise numbers?
    • Did the model invent an institution, year, or report name?
    • Should a note such as “Data as of a certain date; some content is estimated” be added to the page?

    A more reliable approach is to divide AI-generated presentation work into three stages:

    The first stage is content generation. Let the tool help build the outline, refine arguments, and design the page structure, but do not treat all numbers as facts at this stage.

    The second stage is data review. Extract all numbers from the pages, build a data list, add the source, definition, time, and calculation process for each item, and complete cross-validation.

    The third stage is page backfilling. Put only data that has been confirmed or clearly labeled back into the presentation, and unify footnotes, units, time labels, and explanatory text.

    This can be summarized as:

    1. Generate draft
    2. ↓
    3. Extract all data
    4. ↓
    5. Confirm sources and statistical definitions
    6. ↓
    7. Cross-validate and recalculate
    8. ↓
    9. Label actual, forecast, and estimated values
    10. ↓
    11. Check consistency between charts and body text
    12. ↓
    13. Deliver formally

    AI is well suited to improving the efficiency of information organization and expression, but data responsibility cannot be automatically transferred to the model. The robust approach is not to reject generative tools entirely. It is to place them in the role of “drafting, organizing, and assisting analysis,” while using human review as a gate before final conclusions.

    Whether a presentation is professional does not depend on how many decimal places a number has, nor on how complex a chart looks. It depends on whether the numbers can be traced, explained, recalculated, and verified. As long as teams follow five principles—sources are traceable, definitions are clear, time is explicit, calculations are reproducible, and claims are not exaggerated—AI-generated content can move from a polished-looking draft to reliable material that can genuinely support communication and decision-making.