How Novus’s AI PDF Field Scan Works (and When to Fix It by Hand)
See how Novus’s local field scan combines AcroForm widgets, drawn geometry, and text labels, when to trust the boxes, how to rescan and clean leftovers, and when manual placement is the better path.
Written against the Jul 27, 2026 release·what has changed since
Novus’s ‘AI scan’ is easy to misunderstand. It is not a cloud vision model reading your PDF on a remote GPU. It is a local, three-source detector that runs inside the browser tab: real AcroForm widgets when they exist, drawn geometry such as stroked boxes and underlines, and text-label heuristics that synthesize fields beside cues like currency symbols or vocabulary labels. Those suggestions are merged, filtered, and shown as editable overlays you can keep, move, or delete.
Understanding that pipeline helps you trust the scan when it is right—and fix it quickly when a leftover box appears on a header or beside a row label. This guide explains what each source contributes, how duplicates are resolved, what recent cleanup rules remove, and when you should ignore the scan and place fields by hand.
Three sources, one editor canvas
- Widgets (AcroForm) — The highest-confidence path. If the PDF already contains form fields, Novus imports their rectangles, types, and—when present—filled or locked values.
- Geometry — Stroked rectangles and underline-like lines sized like inputs become candidate text, number, or checkbox fields, labelled from nearby text when possible.
- Text synthesis — Labels ending in a colon,
$amount cues, blanks, and known vocabulary can invent a field when a drawn affordance is missing. These suggestions score lower so real widgets and boxes win conflicts.
How merge and dedupe decide what you see
After each page is scanned, Novus filters non-user zones (for example office-use bands) and runs dedupe. Widgets suppress overlapping geometry and text. Geometry usually beats text when they target the same spot. Same-row rules drop a beside-label text box when a drawn amount cell already sits further right—the pattern that used to leave extras next to labels like ‘telephone’ or ‘service income’ on blank expense tables.
Column titles such as Amount or Total Amount are treated as headers, not fill targets, and unlabeled geometry scraps that sit on a header row or just under a section title are discarded. Protected pairs like Month/Year and Yes/No stay separate so option groups are not collapsed into one box.
What the scan does well
- Digital government and benefits forms with real widgets or clear
$amount columns. - Tables where geometry boxes align with currency glyphs and row labels.
- Checkbox rows for Yes/No or Month/Year when boxes and labels are nearby.
- Already-filled or read-only widgets, importing values you can still review locally.
When boxes look wrong—and what to do
Leftovers usually mean the text pass guessed a field where only a header or gap exists, or geometry picked up a decorative cell. Delete the stray overlay, nudge misaligned boxes, or add a missing field with the manual tools. If you change the document (for example after rotating pages outside the editor), run scan again rather than trusting stale overlays.
- 1
Scan once on a clean digital PDF
Open the form in the editor and let the first pass populate fields.
- 2
Spot-check amount columns and headers
Confirm boxes sit in cells, not on ‘Amount’ titles or section banners.
- 3
Delete or move leftovers
Use layers and selection to clear noise without wiping good detections.
- 4
Fill gaps manually
Add text, number, check, or signature tools where the scan stayed silent.
- 5
Rescan only when the page structure changed
A second scan can help after major cleanup; it is not a substitute for deleting one bad box.
Try the field scan on your form
Open a digital PDF in the editor, review the detected fields, tidy leftovers, and export locally.
Scans and photographs are a different job
Image-only pages lack widgets, drawn cells, and selectable labels. The detector has almost nothing reliable to latch onto, so sparse results are expected—not a failure of ‘AI.’ For those files, follow the scanned PDF guide: fix orientation, place fields by hand, and sign locally. Prefer a native fillable PDF when the issuer provides one.
A practical trust checklist
- Does the PDF select as text, or is every page a flat image?
- Are amount boxes aligned to
$cells rather than row labels? - Did Month/Year or Yes/No survive as separate checks?
- Did filled widget values import without becoming new empty fields?
- Have you deleted header or between-section scraps before exporting?
Key takeaways
- Novus’s field scan is a local merge of AcroForm widgets, drawn geometry, and text-label cues—not a cloud OCR model.
- Widgets and geometry outrank text synthesis; same-row and header filters remove common leftover boxes.
- Trust detections on digital forms with clear cells; verify headers and option groups before you fill.
- Delete, move, or manually add fields when the scan misses or invents something; rescan after structural changes.
- Image-only scans need hand placement—see the companion guide rather than expecting automatic boxes.
Was this article helpful?
Your answer stays in this browser. There is no backend to send it to — see how it works.
Keep reading
Reviewed and maintained by
Novus Stream Solutions Editorial Team
The Novus Stream Solutions Editorial Team maintains Novus PDF Studio's product documentation, tutorials and PDF explainers. The team checks product claims against the current browser-local implementation and tests, prefers primary specifications and vendor documentation, and corrects material errors openly. The byline identifies the responsible organization; it does not imply a named expert or professional adviser.
Privacy note: every tool mentioned in this article runs entirely in your browser. Nothing is uploaded or queued on a server. A PDF stays in the tab unless you explicitly use Save on this device, which stores that session in this browser without storing passwords. More on the how it works page.