The core technology updates Instafill.ai shipped between August 2025 and February 2026
Instafill.ai's core technology is improved all the time. Some changes are announced one by one and many are deployed quietly. This post of February 25, 2026 collects the most significant engine-level changes of the previous six months, the ones that affect how accurately and reliably forms are processed. All of them were live and applied automatically.
1. A newer AI engine
The GPT-5 model family. The previous-generation model was replaced with OpenAI's GPT-5 family across the whole platform, for both fine-tuning and form filling. Several variants were evaluated (GPT-5, GPT-5 mini, GPT-5.1, GPT-5.2 and GPT-5 nano), and each type of operation, such as field analysis, dependency detection or data mapping, uses the variant that gave the best results for it. The platform has since moved on again: see the GPT-6 Sol upgrade.
Automatic choice of fine-tuning algorithm. All forms used to go through the same page-by-page fine-tuning, which could not follow content across a page break. A section such as "Employment History" that starts halfway down page 4 and runs to page 6 was split into fragments. Now each page's layout is analyzed during fine-tuning. A single-column page is grouped into logical sections that can span pages, and a multi-column page keeps the page-by-page approach. One form can therefore use both strategies, chosen automatically. Column detection runs on each page with GPT-5 mini, and a full scan takes 5 to 10 seconds even for a large form.
2. Better accuracy on complex forms
Text overflow handling. After filling, every field is checked against its real boundaries on the page, with font size, field size and single-line or multiline in mind. When the text does not fit, a separate AI step shortens or reformats it without losing the meaning. A 65-character value in a 40-character field might turn "123 North Washington Boulevard, Suite 400" into "123 N Washington Blvd, Ste 400".
Cross-page field dependencies. Fine-tuning now scans the whole form for fields that depend on another page, such as a "Same as mailing address" checkbox on page 5 that refers to an address entered on page 1, or a total on page 8 that sums values from page 6. They are stored and used when filling so that dependent fields stay consistent.
Section-based fine-tuning. The form is first parsed into logical sections, such as "Applicant Information" or "Financial Details", and fine-tuned at the section level instead of page by page. A section longer than 5 pages is split into smaller chunks. If clear section boundaries cannot be found, the form falls back to page-by-page processing, so existing forms are not affected.
Tables and lists. Tables are filled row by row, with values placed in the right columns. The AI reads the column headers, so wide tables with many columns no longer mix up which value goes where. Lists are a separate data type: each item maps to its line, and only rows with data are filled, with no phantom entries in empty rows. Instafill.ai detects whether a group of fields is a table or a list, and you can override it. For a credentialing form with a 5-column, 10-row employment history table, the AI fills one row per source record and stops when the data runs out. Table and list processing describes it today.
Repeatable sections. Many forms repeat a block of fields, such as Officer #1, #2 and #3 on separate pages, or a "Dependent Information" block per family member. Fine-tuning now detects these repeating structures and labels each instance as part of the same section, and filling maps different source data to each repetition. This ends the problem of every officer block being filled with the first officer's details. See intelligent field grouping.
Flat-to-fillable PDF conversion. Many government applications, court filings, insurance claims and credentialing packets are flat PDFs with no fillable fields. Instafill.ai now converts them automatically instead of having each text box, checkbox and signature field placed by hand. A primary detector reads the PDF's vector data to find table borders, underlines and field boundaries. If it finds too few fields, a fallback detector finds blank areas from whitespace patterns in the text layer, and checkboxes are detected separately as small squares near label text. The result is saved as a reusable template, so the conversion happens once per form type. See flat-to-fillable conversion.
3. Improving quality over time
Style replication from filled samples. Upload forms you have already filled as examples, and Instafill.ai learns how you abbreviate, how you format dates and numbers, and how you phrase narrative answers, then applies that style to new sessions of the same form. If your office writes "N/A" rather than leaving a field blank, or formats phone numbers as "(555) 123-4567", a few examples are enough for the AI to pick it up. More examples give a closer match. See style replication.
Regenerate fine-tuning. Improvements to the algorithms used to require re-uploading a form. Now a single click re-runs the latest algorithms on an existing form and keeps your manual edits to field names, types and groupings.
Fine-tuning for new or selected fields. When you add fields in the field editor, only those fields are fine-tuned, so they are ready within seconds without touching the rest. You can also select specific fields and fine-tune just those. See AI training and field management.
4. Bigger source documents
Page-by-page source reading. Text is now extracted from PDF and Word source files page by page, with a dedicated AI operation per page, instead of in bulk. Some pages of large documents could be skipped or read partly before, especially with multi-column layouts, mixed text and tables, or scans. It takes slightly more time, and the gain is largest for source files over 10 pages.
Large source files. Before filling, the AI now maps which pages of your source documents are relevant to which parts of the form, and then each form section is filled from the relevant content only. With a 70-page credentialing packet behind an 18-page application, the "Education" section might use pages 12 to 15 and "Employment History" pages 22 to 38. If the source is longer than the model's context allows, the mapping is split into several calls automatically.
5. Infrastructure
A rebuilt processing pipeline. The backend now uses a dedicated task queue, so CPU-heavy work such as PDF parsing, screenshot generation and fine-tuning no longer blocks other work. Up to 30 tasks run in parallel.
Date, time and number formats. Date, time and number are separate field types now. You set a preferred date and time format at the workspace level, such as MM/DD/YYYY, DD/MM/YYYY or YYYY-MM-DD, and it is used across all forms in that workspace. Date and time fields are detected during fine-tuning.
These improvements applied to every account without any action. Questions go to [email protected].
