Check a Scanned Document Before AI Summarises It

Inspect the fictional note

The comparison contains a complete four-page original and a deliberately altered three-page extraction. The original is represented as structured text in the exercise pack at the end of this article. It is not an image of a real scan, and the altered text is not a measured result from an OCR product. All names, quantities and procedures are invented for this exercise.

Your requested summary covers equipment quantities, issue and return steps, and the collection condition on pages 1–3. It excludes the paint-colour appendix. This boundary matters: an unrelated missing appendix does not automatically invalidate every possible summary.

For each inspection, write the issue, its location, its possible consequence and the next step. Compare the equipment-and-procedure scope with a whole-note summary: the missing-page decision changes. The full original and altered extraction are reproduced below so the exercise does not depend on a widget, upload or visual puzzle.

Completing the observations records your reasoning. It does not repair the extraction. Keep a proposed correction separate from a corrected copy that has actually been checked. If the inspection activity is available on this page, choose the summary scope, record the four observations and reveal their explanations. The complete original and altered text below also support a paper comparison; selections do not modify the source or approve its use.

First, count identities, not just pages

A file with four pages might contain one duplicate and omit a different page. Match each file position to its printed label, contents entry and cross-references. Keep both identifiers: a cover sheet can make file position 3 correspond to printed page 2.

In this exercise, page 1 explicitly points to Appendix A on page 4, while the extraction stops after page 3. Inspect the complete original to establish what is missing. Here it is paint-colour information, explicitly outside the requested scope. A whole-note summary still needs it.

For a real document, retain the received file unchanged. Record its filename, version and intended use before making a separate corrected copy. Check for cropped edges, unexpected blanks, duplicates, rotation and illegible regions. If the source page itself is missing or unclear, obtain it from the document owner. Do not fill the gap from what a model says is likely.

Then check words and sequence

OCR converts images of text into machine-readable text. W3C's PDF7 technique describes checking each converted page for complete text and correct reading order. PDF3 explains how multi-column and other complex layouts can produce an incorrect reading sequence.

Apply those checks to the actual representation available to the summariser. In the exercise, every step on page 2 survives, but the extraction alternates between the issue and return columns. Restoring only missing words would leave this problem untouched. Reconstruct each sequence under its own heading, then compare all six steps with the original.

Also inspect negations, decimal points, units and qualification words such as “only”, “unless” and “after”. On page 3, the supervisor-check wording survives as a loose note, while the collection sentence appears unconditional. Restore the connection, not merely the words.

These W3C sources are PDF accessibility techniques. Applying their completeness and structure checks to AI-summary intake is this article's editorial inference. The techniques do not certify an AI summary's accuracy, and this exercise is not a PDF accessibility audit.

Check what every number belongs to

For any table value that the summary needs, identify its row label, column heading, unit and attached note. W3C's PDF6 technique explains the need to preserve relationships among table headers and data when presentation changes.

The fictional table's total of 16 display boards survives the altered extraction. Its meaning does not: swapping the two quantity headings turns 12 available and four awaiting repair into 12 awaiting repair and four available. Checking the total alone would miss the mismatch. Both equipment rows need to be checked against the original after correcting the headings.

Figures need their own review. A surviving caption does not establish the meaning of a missing legend, axis or plotted value. If the summary needs a figure, supply a checked description or hold that claim until its evidence can be inspected. The fictional exercise contains no chart and does not test chart extraction.

Choose a next step for a named scope

There are three useful responses to a mismatch:

• Repair against a readable original, preserve the correction log and recheck the corrected copy.
• Use a clearly bounded part of the document with an explicit limitation, when the unresolved material is genuinely outside that scope.
• Hold the affected use when required content remains missing, ambiguous or changed.

A missing paint appendix can be excluded from this equipment-and-procedure summary. The mixed steps, reversed headings and detached condition all affect that summary and still need repair. A scope limitation cannot make those changed facts dependable.

Holding is a defensible response to uncertainty. Repairing the appendix is also valid when the complete original is available, even if this particular summary does not need it. The important distinction is between a planned action and a checked result. Selecting “repair” in a form does not make a corrected document exist.

Keep an intake record someone else can follow

Record the source identity, intended summary, file position and printed page label, observed mismatch, likely effect, correction reference, reviewer and next decision. Keep unresolved points explicit. Where the tool does not expose what it extracted, record that limitation; a clean export from a different application does not establish what the summariser received.

You can copy the record fields into an approved document or print them. This manuscript contains no upload, automated extraction or AI request. Keep practice records fictional; use the authorised document-handling process for actual material.

For real work, use an approved environment and document-handling process. Preserve the original, corrected representation and review record together. Once the required input has been checked, compare the generated summary's substantive claims with that source, especially its numbers, conditions and exceptions. Intake is one checkpoint; output review remains necessary.

Complete fictional exercise pack

Original document

North Quay Workshop: equipment and collection note
Version 1.0 | FICTIONAL TRAINING DOCUMENT
Fictional snapshot, 1 October 2026

FILE POSITION 1 | PRINTED PAGE 1 OF 4
Purpose and contents
This fictional note records a single equipment snapshot. It is a practice document, not current stock information.
Requested summary: equipment quantities, issue and return steps, and the collection condition on pages 1–3. Paint colours are outside this summary.
Contents: page 1, purpose and contents; page 2, issue and return steps; page 3, equipment table and collection condition; page 4, Appendix A: paint colours.
For paint colours only, see Appendix A on page 4. The appendix contains no equipment quantities, operating steps or collection conditions.

FILE POSITION 2 | PRINTED PAGE 2 OF 4
Issue and return steps
Issue steps
1. Record the equipment requested.
2. Confirm the required items are available.
3. Hand over the recorded items, subject to the collection condition on page 3.
Return steps
1. Count the items returned.
2. Put damaged items in the repair area.
3. Record the returned quantities.

FILE POSITION 3 | PRINTED PAGE 3 OF 4
Equipment snapshot and collection condition
T1. Equipment counts at this snapshot. Unit: individual items.
Equipment | Available (items) | Awaiting repair (items)
Display boards | 12 | 4
Trolleys | 3 | 1
Collection of available display boards is permitted only after a supervisor records a check.
Counts describe this snapshot; they do not confirm a future booking.

FILE POSITION 4 | PRINTED PAGE 4 OF 4
Appendix A: paint colours
Display-board frame colour: deep blue.
Trolley frame colour: slate grey.
These are paint-colour notes only. No additional quantities, steps or collection conditions appear here.

Deliberately altered extraction

North Quay Workshop: equipment and collection note
Version 1.0 | FICTIONAL TRAINING DOCUMENT
Fictional snapshot, 1 October 2026

FILE POSITION 1 | PRINTED PAGE 1 OF 4
Purpose and contents
This fictional note records a single equipment snapshot. It is a practice document, not current stock information.
Requested summary: equipment quantities, issue and return steps, and the collection condition on pages 1–3. Paint colours are outside this summary.
Contents: page 1, purpose and contents; page 2, issue and return steps; page 3, equipment table and collection condition; page 4, Appendix A: paint colours.
For paint colours only, see Appendix A on page 4. The appendix contains no equipment quantities, operating steps or collection conditions.

FILE POSITION 2 | PRINTED PAGE 2 OF 4
Issue and return steps
Issue steps
Return steps
1. Record the equipment requested.
1. Count the items returned.
2. Confirm the required items are available.
2. Put damaged items in the repair area.
3. Hand over the recorded items, subject to the collection condition on page 3.
3. Record the returned quantities.

FILE POSITION 3 | PRINTED PAGE 3 OF 4
Equipment snapshot and collection condition
T1. Equipment counts at this snapshot. Unit: individual items.
Equipment | Awaiting repair (items) | Available (items)
Display boards | 12 | 4
Trolleys | 3 | 1
Collection of available display boards is permitted.
Counts describe this snapshot; they do not confirm a future booking.
Note: only after a supervisor records a check.

Compare your findings

The extraction omits page 4, interleaves the two procedures on page 2, reverses the quantity headers on page 3 and separates the supervisor condition from collection permission. The appendix can be outside the stated equipment-and-procedure summary because the complete original establishes its contents. The other three errors affect that scope and require correction and rechecking. For a whole-note summary, the missing appendix also needs restoration. None of these observations establishes that a corrected extraction already exists.

One practical next step

Bring a fictional document and your intake observations to an AI workshop to practise distinguishing a source problem, an extraction problem and a summary problem.

Relevant reading

Practical AI Literacy: Four Skills Beyond Prompt Writing places this source-checking exercise within a wider approach to working with AI.

Source note

Primary sources rechecked on 4 October 2026: W3C PDF7, PDF3 and PDF6, linked at the claims they support above. Fictional examples, inspection decisions and the extension to AI-summary intake are original editorial guidance. No product-specific extraction success rate, accessibility certification or accuracy score is claimed.