Extract NAR1 with ChatGPT: check the results before using them in Excel

Try it yourself: turn two NAR1 pages into a table

Use the public two-page excerpt of ARCH MARKETING LIMITED’s historical NAR1. The task is to capture its name, original company number, return date, address and share capital, with source locations and unresolved items. It is not a complete company file and excludes director and member details. A shared public input makes differences easier to locate before working with client material.

Open the sample page and keep the PDF alongside a blank worksheet with five columns: field, extracted value, source text, PDF page and review status. Keeping the result beside the source makes it easier to check each item. Once checked, you can arrange the data in the Excel, JSON or other format your colleague needs.

ClariRecord: two-page NAR1 source and extraction sample ↗

Check readability before requesting extraction

File handling varies by ChatGPT client, account and processing method. An uploaded attachment or a statement that it has been read does not prove access to every page, scanned word or table. Ask for the pages it can actually read and a transcription of the company-name area on page 1 and share-capital row on page 2. Compare those with the PDF yourself.

If it cannot read an area, note the page and location. An unreadable field is different from one left blank in the original. If your interface supports images, try a clearer image of that page, keeping its page number. You can also supply OCR text and check it against the image. Ask it to read the same passage again; leave anything still unclear marked for review rather than using it as checked data.

OpenAI: using ChatGPT with files ↗OpenAI: image input guidance ↗

Tell ChatGPT exactly what to extract, then save its first answer

After the precheck, use: “Using only the two attached NAR1 pages, extract company names, company number with its source label, return date, registered-office address and each share-capital row. Give source text and a PDF page for each. Identify pages received. Distinguish N/A, blank, page not supplied and unreadable. Do not search to fill gaps or infer current company status. End with a review list.”

The downloadable prompt contains three language versions; use one. Save an untouched first response before making corrections so that model output and human changes remain distinguishable. A prompt defines the task but does not ensure a correct response.

Start with four details you can check in the PDF

Use the displayed sample as a reference, then return to the PDF. These four details help you confirm that you have the right company and filing, and that you have read the share-capital figures correctly. They do not validate every field: the address, blank values and other requested items still need their own checks.

Scroll sideways to see all columns; focus the table and use arrow keys on a keyboard.

ClariRecord: two-page NAR1 source and extraction sample ↗

Try spotting four mistakes in this example

This deliberately faulty output is hypothetical, not a measured ChatGPT response: “BRN = 1063959; document date = today; directors = 0; issued shares = 20,000.” Instead of merely requesting a rerun, identify each interpretation problem and its source location or coverage limitation.

The number’s label is wrong, not necessarily its digits. The return date comes from page 1. Director details are outside the excerpt, not zero. The share count needs checking on page 2 for accidental addition of the detail and printed total. Keep each original value and correction reason so the revision is evidence-based.

Follow up with: “Correct only these four items. Show the previous value, revised value, evidence and remaining uncertainty. Do not add information from other sources.” Verify the second response yourself. An explanation is not proof; a renewed claim to have read director pages outside the excerpt remains unresolved.

Check the data again after moving it into Excel

In your spreadsheet, give company details, share-capital rows and unanswered questions their own space. Keep the share class and currency with the figures, and note who will check anything unclear. If your ChatGPT interface cannot create an Excel file, you can copy its table into a worksheet and carry out the same checks.

After import, recheck identifiers, dates and numbers for formatting changes. Keep the company number as text, label the meaning of dates, and separate share quantity from monetary amounts and currency. This sample has 10,000 ordinary shares, a total amount of HKD 10,000 and paid-up amount of HKD 10,000. Retain the printed total separately rather than summing it as a second detail row.

ClariRecord: two-page NAR1 source and extraction sample ↗

Before you share the finished spreadsheet

Ask the person who will use the spreadsheet to find the company name, return date and share-capital row, then check those details in the PDF. Can they tell which values you have checked and which questions remain unanswered? A short note beside an empty cell can save them from assuming that missing information means there is nothing to report.

Retain the PDF, first output, correction record and accepted version, labelled as results from a two-page excerpt. Valid JSON or a passing schema checks structure, not fidelity to the source. Reassess coverage for other documents; this exercise does not validate all years, continuation sheets or a whole batch.

For a service request, provide the actual files, fields, target format and review requirements. Standard extraction and manual review are distinct services; Excel preparation and custom formatting also need agreed scope. Use the public sample to discuss presentation, then confirm the work against your own documents.

ClariRecord: two-page NAR1 source and extraction sample ↗ClariRecord: scope and terms ↗