We make improvements to Instafill.ai’s core technology on a regular basis. Some we announce individually. Many are deployed quietly. This article compiles the most significant engine-level changes from the past six months – the updates that affect how accurately and reliably Instafill.ai processes your forms.
All of these are live and apply automatically. No action needed on your side.
6 months of core technology updates
1. Smarter AI Engine
Upgraded to GPT-5 model family
We replaced the previous-generation model with OpenAI’s GPT-5 family – the latest generation of reasoning models – across the entire platform, for both fine-tuning and form filling. GPT-5 handles complex form layouts more accurately and processes information faster.
We evaluated multiple variants (GPT-5, GPT-5 mini, GPT-5.1, GPT-5.2, and GPT-5 nano) and selected the best one for each type of operation. Field analysis, dependency detection, data mapping – each uses whichever variant delivers the best results for that task.
Automatic fine-tune algorithm selection
Previously, all forms went through the same page-by-page fine-tuning process. That approach had a fundamental limitation: it couldn’t handle content that crosses page boundaries. A logical section like “Employment History” that starts halfway down page 4 and continues through page 6 was split into disconnected fragments, losing context and reducing accuracy.
Instafill.ai now analyzes the layout of each page during fine-tuning. The system detects whether a page has a single-column layout (content flows top-to-bottom) or a multi-column layout (content arranged side-by-side). For single-column pages, the AI groups content into logical sections that can span across page breaks, so related fields stay together. For multi-column pages – where section detection is less predictable – the system keeps the standard page-by-page approach.
This means a single form can use different processing strategies for different pages, depending on their layout. The selection happens automatically during fine-tuning. No manual configuration needed.
| Technical detail: The column detection runs on each page using GPT-5 mini with reasoning disabled for speed. The full scan takes 5-10 seconds even for large forms. Pages identified as single-column are grouped and sent to the section detection pipeline. Pages identified as multi-column are processed individually. The results are then merged, so every page in the form is covered regardless of its layout type. |
2. Better Accuracy for Complex Forms
Smart text overflow handling
A common problem with form filling, both manual and automated, is text that doesn’t fit inside a field. A company name that’s too long for the box. An address that overflows a single-line input. A narrative answer that exceeds a text area.
Instafill.ai now automatically detects overflow after filling. The system checks every filled field against its actual pixel boundaries on the PDF page, accounting for font size, field dimensions, and whether the field is single-line or multiline. When text doesn’t fit, a separate AI operation reformats or shortens the content to stay within bounds without losing the meaning.
| Example: If a field allows 40 characters but the AI fills out 65, the system automatically trims or abbreviates the value so it fits. “123 North Washington Boulevard, Suite 400” becomes “123 N Washington Blvd, Ste 400”. This works for both single-line and multiline text areas. |
Cross-page field dependencies
Many multi-page forms contain fields that reference information from other pages. A checkbox on page 5 that says “Same as mailing address” (entered on page 1). A “Total” row on page 8 that should sum values from page 6. An employer field on page 3 that should match a name from page 2.
During fine-tuning, the AI now scans the full form structure, identifies these cross-page relationships, and stores them. When filling, it uses this information to make sure dependent fields are populated consistently. This runs in the background during fine-tuning, so it doesn’t slow down the process.
Section-based fine-tuning
We changed how Instafill.ai understands form structure at a fundamental level. Previously, the AI processed forms strictly page by page. Now, it first parses each form into logical sections – groups of related fields like “Applicant Information,” “Employment History,” or “Financial Details” – and fine-tunes at the section level.
This matters because form sections don’t always align with page breaks. An employment history block might start halfway down page 4 and continue through page 6. By understanding sections rather than pages, the AI maintains better context and produces more accurate results.
| Technical detail: Sections that span more than 5 pages are automatically split into smaller chunks. If the system can’t determine clear section boundaries for a particular form, it falls back to the standard page-by-page processing, so no existing forms are affected negatively. |
Extended support for tables and lists
Tables are now filled row by row, with values placed in the correct columns. Previously, the AI could mix up which value goes where in wide tables with many columns. Now it reads the column headers, understands the table structure, and maps each value to the right cell.
Lists are now treated as a distinct data type, separate from tables. Each item maps to the correct line, and only rows with actual data get filled – no more phantom entries in empty rows. The system auto-detects whether a group of fields is a table or a list, with optional manual override if needed.
| Example: A credentialing form with an employment history table – 5 columns (Employer, Position, Start Date, End Date, Reason for Leaving), 10 rows. The AI reads each source record, fills one row at a time, and stops when the data runs out instead of repeating or guessing. |
Automatic detection of repeatable sections
Many forms repeat the same block of fields across multiple pages. Think of a corporate filing form where page 2 asks for Officer #1’s name, title, and address, page 3 has the same fields for Officer #2, and page 4 for Officer #3. Or an insurance application that repeats a “Dependent Information” block for each family member.
The AI now detects these repeating structures during fine-tuning and labels each instance as part of the same section. During filling, it maps distinct source data to each repetition instead of copying the same values everywhere. This eliminates the problem of every officer block being filled with the first officer’s information.
Flat-to-fillable PDF conversion
Many commonly used forms – government applications, court filings, insurance claims, healthcare credentialing packets – are distributed as flat PDFs with no fillable fields. Before Instafill.ai can fill a form, it needs to be fillable. Previously, that meant manually placing every text box, checkbox, and signature field by hand.
Instafill.ai now converts flat PDFs into fillable forms automatically. The conversion pipeline uses multiple detection engines in sequence. The primary detector analyzes the PDF’s vector data to locate table borders, underlines, and field boundaries. When that detector finds insufficient fields, a fallback detector identifies blank areas by analyzing whitespace patterns in the text layer. Checkboxes are detected separately by locating small square regions near label text.
| Example: An immigration attorney uploads a flat USCIS I-485 form. Instafill.ai detects the field locations, creates an interactive fillable version, and saves it as a reusable template. The conversion only needs to happen once per form type. Every future fill reuses the same converted template. |
3. Improving Quality Over Time
Fine-tuning sets the baseline for how well the AI understands a form. But accuracy doesn’t have to stop there. We’ve added several ways for you to keep improving quality after the initial fine-tune, without starting over.
Style replication from filled samples
You can upload previously filled forms as examples, and Instafill.ai will learn your filling patterns from them. The AI analyzes how you abbreviate, how you format dates and numbers, how you phrase narrative answers, and replicates that style when filling new sessions of the same form.
The system reads the full text of each example file, maps its pages to the corresponding form pages, and extracts field-level patterns. The more examples you provide, the more accurately the AI matches your style.
| Use case: If your office always writes “N/A” instead of leaving fields blank, or formats phone numbers as “(555) 123-4567” rather than “555-123-4567”, upload a few filled examples and the AI will pick up on these patterns and apply them consistently. |
Regenerate fine-tuning
We improve Instafill.ai’s AI and processing algorithms on a regular basis. Previously, the only way to benefit from these improvements was to re-upload your forms and start from scratch. Now, you can regenerate the fine-tuning for any existing form with a single click. The system re-runs the latest algorithms on the same form, preserving any manual edits you’ve made to field names, types, or groupings.
This is especially useful for forms you uploaded months ago. As our AI improves, regenerating the fine-tune lets those forms benefit from the latest updates without any rework on your side.
Auto fine-tune for newly added fields
When you add new fields to a form using the field editor, the system now automatically fine-tunes only those new fields without re-processing the entire form. This means you can add missing fields at any time and have them ready for filling within seconds, with no impact on the fields that were already fine-tuned.
You can also select specific fields on the form and trigger a targeted fine-tune for just those fields – useful when you want to refresh the AI’s understanding of a particular section.
| Together, these three tools – filled examples, regenerate fine-tuning, and partial fine-tuning – mean that accuracy is not static. The initial fine-tune is a starting point. Every form can be refined further based on your feedback, your examples, and our ongoing AI improvements. |
4. Handling Bigger Documents
Page-by-page source reading
Instafill.ai now extracts text from PDF and Word source files page by page, using a dedicated AI operation for each page. Previously, text extraction was done in bulk, which meant some pages in large documents could be skipped or read incompletely, especially when the layout was complex (multi-column, mixed text and tables, scanned content).
The page-by-page approach makes sure every piece of information in your source documents is actually available to the AI when it fills the form. It takes slightly more processing time, but the accuracy improvement is significant, especially for source files over 10 pages.
Large source file processing
When you upload source data for form filling (resumes, contracts, reports, credentialing packets), that data is used to populate the form fields. Previously, the full text of your source files was passed to the AI for every form page, which could slow down processing and reduce accuracy on very large documents.
Now, the system adds an intermediate step before filling begins. The AI first maps which pages from your source documents are relevant to which parts of the form. Then, when filling each form section, it works only with the relevant source content, not the entire document.
| Example: If you upload a 70-page credentialing packet as the source for an 18-page application, the AI first determines which of those 70 pages are relevant to each part of the form. For the “Education” section, it might pull pages 12 to 15. For “Employment History,” pages 22 to 38. This focused approach reduces errors from irrelevant context and speeds up processing. |
For source files that exceed the AI model’s context limits, the system automatically splits the mapping into multiple calls. You don’t need to manually trim or pre-process your documents.
5. Infrastructure
Rebuilt processing pipeline
We rebuilt the backend processing pipeline to use a dedicated task queue. Previously, CPU-heavy operations like PDF parsing, screenshot generation, and fine-tuning could block each other when running at the same time. Now, these operations run independently and concurrently, with up to 30 tasks executing in parallel. The system handles multiple requests more reliably and doesn’t slow down under load.
Custom date, time, and number formats
Date, time, and number are now distinct field types. You can set your preferred date format (MM/DD/YYYY, DD/MM/YYYY, YYYY-MM-DD, and others) and time format at the workspace level, and the AI will use them consistently across all forms in that workspace. The system auto-detects date and time fields during fine-tuning. This is useful for international teams or when working with forms from different countries.
Had forms that didn’t work well before?
These improvements apply automatically to all accounts. Upload your forms and see the difference:
These updates are part of our ongoing work to make Instafill.ai more accurate, faster, and capable of handling more complex documents.
If you have questions about any of these updates or want to discuss how they apply to your forms, reply to this email or reach out at [email protected].