ACORD Form Extraction Compared: Sensible vs Indico vs Roots AI

Our team spent three months testing ACORD form extraction across Sensible, Indico Data, and Roots AI. We ran each platform against 400+ ACORD forms (25s, 28s, 125s, 126s, and 130s) collected from real agency submissions, including forms with handwritten entries, low-resolution scans, and multi-page attachments. This article reflects what we found, including where each platform impressed us and where each one fell short.

Why ACORD Extraction Is Harder Than It Looks

ACORD forms are standardized. That should make them easy to parse. In practice, it does not.

The ACORD 25 (Certificate of Liability Insurance) alone has gone through dozens of revisions since the 1990s. Field positions shift between versions. Some agencies still use forms from 2014; others have adopted the 2023 revision. A parser trained on the current version will misread fields on older forms unless it accounts for layout drift.

Beyond versioning, real-world ACORD forms come with problems that clean test data never captures:

• Handwritten entries. Agents fill in policy numbers, effective dates, and additional insured information by hand. OCR accuracy on handwritten insurance data hovers around 70-80% in our testing, depending on handwriting legibility.
• Poor scan quality. Faxed copies of faxed copies are still common. We received ACORD 25s scanned at 72 DPI with visible moiré patterns. All three platforms struggled with these.
• Overstuffed fields. Agents routinely cram additional insured language, endorsement descriptions, and special conditions into the “Description of Operations” box. Parsing this freeform text into structured fields is a different problem than extracting labeled data.
• Multi-page attachments. An ACORD 25 with three pages of additional insured schedules needs the parser to understand table continuation across pages, not just single-page extraction.
• Field naming inconsistencies. The same conceptual field (say, “policy effective date”) appears with different labels and in different positions across ACORD 25, ACORD 28, and ACORD 125. A parser needs to normalize these into a consistent schema.

We tested all three platforms against these real-world conditions, not clean sample forms.

Head-to-Head Comparison Table

Sensible: Developer-First Extraction

Sensible takes a fundamentally different approach than the other two platforms. Instead of training a model on labeled examples, you write JSON-based extraction configurations (called “SenseML”) that describe where fields are on the document and how to extract them. Think of it as a structured, version-controlled way to build extraction rules.

What Works Well

Pre-built ACORD extractors. Sensible ships with pre-built configurations for the most common ACORD forms. We pointed the ACORD 25 extractor at our test set and got usable results within minutes, not weeks. The pre-built configs handled the 2019, 2021, and 2023 ACORD 25 revisions without modification.

Transparent pricing. Sensible charges per document, with published pricing on their website. At the time of our testing, the Business plan runs $499/month for 2,000 documents. That is $0.25 per document at volume, with no hidden enterprise surcharges. For an agency processing 500 certificates per month, the math is clear and predictable.

Developer experience. The SenseML configuration language is well-documented. When the pre-built extractor missed a field, our developer modified the JSON config, tested against sample documents in Sensible’s web editor, and deployed the fix in under an hour. Version control for extraction configs is a real advantage when you need to track why a field extraction changed.

Structured output. Every extraction returns clean JSON with confidence scores per field. Low-confidence fields get flagged for human review automatically if you configure it. In our testing, fields with confidence scores above 0.85 were correct 97% of the time.

Where It Falls Short

Handwritten field accuracy. On our handwritten ACORD field test set, Sensible averaged 68-75% accuracy. Policy numbers with mixed alpha-numeric characters were the worst performers, averaging around 62%. This is not a Sensible-specific problem; it is a fundamental OCR limitation, but Sensible does not offer a way to train a custom handwriting model to improve it.

Enterprise support. Sensible is a developer tool. If your team does not have a developer comfortable with JSON configurations and REST APIs, the onboarding experience will be frustrating. Their support team is responsive but small. We waited 4-6 hours for responses during business hours, longer on weekends.

Complex table extraction. Multi-page additional insured schedules with inconsistent column formatting gave Sensible trouble. When table rows wrapped across pages with slightly different column widths, the extractor sometimes merged or skipped rows. We had to write custom SenseML anchoring logic to handle the worst cases.

Limited document type breadth. Sensible works on any document, but you need to build or customize the config for each new form type. If you process 50+ different ACORD and carrier-specific form types, the configuration maintenance burden adds up.

Indico Data: Enterprise Transfer Learning

Indico Data approaches document extraction as a machine learning problem. You label 50-100 example documents, train a model using Indico’s transfer learning architecture, and deploy the model to process new documents. The platform learns to handle variations in layout, formatting, and even handwriting through training data.

What Works Well

Document type breadth. Indico claims support for 900+ document types across their customer base. For insurance operations that process ACORD forms, loss runs, policy declarations, endorsements, and carrier-specific correspondence, Indico’s ability to handle diverse document types from a single platform is a genuine advantage. We trained models for ACORD 25, 28, and 125 and carrier-specific loss runs in the same Indico instance.

Transfer learning on small sets. Indico’s core pitch is that you can train an accurate extraction model with 50-100 labeled examples. In our testing, this held up. We labeled 75 ACORD 25s, trained a model, and achieved 90% field-level accuracy on typed fields within a week. Adding 25 more labeled examples with handwritten entries improved handwritten field accuracy from 65% to 78%.

Multi-language support. For agencies serving multilingual markets, Indico supports 100+ languages. We tested with Spanish-language ACORD submissions and the extraction quality was notably better than Sensible’s on non-English content.

On-premises deployment option. For carriers with strict data residency requirements, Indico offers on-prem deployment. None of the document data leaves your infrastructure. This matters for carriers subject to state data privacy regulations or handling sensitive claimant information.

Where It Falls Short

No self-serve. You cannot sign up for Indico, upload a document, and get results. The sales process involves demos, scoping calls, and contract negotiation. For an agency that wants to start extracting ACORD 25s next week, this is a dealbreaker.

Pricing opacity. Indico does not publish pricing. Based on our conversations and industry contacts, enterprise contracts start around $100,000 per year and scale with volume and document types. For a 50-person agency processing 2,000 documents per month, Sensible’s transparent pricing is significantly more accessible.

Training data maintenance. Transfer learning is powerful, but it requires ongoing labeled data. When ACORD releases a new form revision, or when a carrier changes their loss run format, you need to label new examples and retrain. Our team estimated 5-10 hours of labeling work per quarter to maintain extraction quality across our document types.

Implementation timeline. From contract signature to production deployment, Indico quoted us 8-12 weeks. Our industry contacts reported actual timelines of 12-20 weeks, including data preparation, model training, integration testing, and staff training. For comparison, Sensible went from API key to working extraction in 3 hours.

Roots AI: End-to-End Insurance Process Automation

Roots AI takes a broader approach than pure extraction. Their InsurGPT platform, trained on over 250 million insurance documents according to their published materials, handles not just data extraction but also downstream tasks like policy checking, quote generation, and submission processing. ACORD extraction is one capability within a larger automation suite.

What Works Well

Insurance-native understanding. Roots AI does not just extract text from fields; it understands insurance context. When extracting an ACORD 125 (commercial insurance application), the platform identifies relationships between named insureds, locations, and coverage lines. In our testing, Roots AI correctly parsed multi-location commercial applications where the other two tools needed per-location extraction configs.

Pre-trained on insurance documents. Unlike Indico, which requires labeled examples for each document type, Roots AI comes pre-trained on common insurance forms. The platform recognized and extracted ACORD 25, 28, 125, and 130 fields without us providing training data. Accuracy on typed fields was 91-94% out of the box.

Process automation beyond extraction. If your goal is not just to extract data from an ACORD form but to use that data to check policy compliance, generate quotes, or populate a management system, Roots AI offers pre-built workflows for common insurance processes. We tested their submission-to-quote flow and found it handled 70-75% of straightforward commercial property submissions without human intervention.

Pre-built insurance system connectors. Roots AI offers connectors to Applied Epic, Vertafore AMS360, and other insurance management systems. Data extracted from ACORD forms flows into the agency management system without custom integration work. In our test environment, the Applied Epic connector worked as advertised.

Where It Falls Short

Newer company, less public validation. Roots Automation was founded in 2018 and rebranded their insurance-specific offering relatively recently. Compared to Sensible’s developer community and Indico’s enterprise reference customers, Roots AI has fewer publicly available case studies and third-party validation. We could not find independent accuracy benchmarks.

Pricing transparency. Like Indico, Roots AI does not publish pricing. Based on our conversations, their platform is priced as an enterprise subscription, positioning it against Indico rather than Sensible. Agencies processing fewer than 5,000 documents per month may find the cost difficult to justify for extraction alone, though the process automation features shift the ROI calculation.

Customization depth. When Roots AI’s pre-trained models got a field wrong, correcting the extraction required working with Roots AI’s support team rather than editing a configuration file ourselves. Sensible’s self-service approach to fixing extraction errors felt faster for our team.

Handwritten field accuracy. Despite the larger training corpus, Roots AI’s accuracy on handwritten ACORD fields (70-76%) was not significantly better than the other platforms. The fundamental challenge of handwriting OCR on insurance forms persists regardless of how much typed training data the model has seen.

The Handwriting Problem: None of Them Solve It

This deserves its own section because every buyer asks about it.

Handwritten entries on ACORD forms are still common. Agents write policy numbers, effective dates, additional insured names, and description-of-operations notes by hand. Every platform we tested struggled with this, and the accuracy numbers tell a consistent story:

These numbers are from our test set of 120 ACORD forms with handwritten entries. Your results will vary based on handwriting legibility, scan quality, and pen color (blue ink scans worse than black ink on most scanners).

The honest recommendation: if your workflow involves significant handwritten ACORD data, plan for human review on those fields regardless of which platform you choose. Use AI extraction for the typed/printed fields and flag handwritten fields for manual verification.

Field Naming Inconsistencies Across ACORD Versions

Another problem that affects all three platforms: the same conceptual field appears differently across ACORD form types. “Policy effective date” on an ACORD 25 is in a different location and with different labeling than on an ACORD 125. If you need to normalize extracted data into a single schema for downstream processing, you need a mapping layer between raw extraction output and your target schema.

Sensible handles this through configurable output field names in SenseML. Indico handles it through labeled training data that maps physical fields to your target schema. Roots AI handles it through pre-built insurance ontology mapping. All three approaches work, but all three require configuration effort proportional to the number of ACORD form types you process.

We mapped five ACORD form types to a unified schema and found that the mapping effort was roughly equivalent across platforms: 4-8 hours of configuration per form type, plus testing.

Which Tool for Which Buyer

Choose Sensible if: You have developers on staff, you process primarily ACORD 25s and 28s, you want transparent per-document pricing, and you want to be extracting data within a week. Sensible is the right choice for agencies and MGAs with technical teams that value speed and control.

Choose Indico Data if: You process diverse document types beyond ACORD forms (loss runs, policy documents, carrier correspondence), you need multi-language support, you have the budget for an enterprise contract, and you can wait 3-4 months for full deployment. Indico is the right choice for mid-to-large carriers or TPAs with high document volumes and varied formats.

Choose Roots AI if: You want more than extraction. If your goal is to automate the entire submission-to-quote or certificate-to-policy-check process, Roots AI’s broader automation capabilities justify the enterprise pricing. Roots AI is the right choice for insurance operations teams that want to reduce process handling time, not just data entry time.

Choose a combination if: Many of the operations teams we spoke with during this evaluation use Sensible for high-volume, well-structured ACORD extraction and a second platform for more complex or diverse document processing. The per-document pricing on Sensible makes it cost-effective to handle the predictable work there and route complex documents elsewhere.

What We Would Do Differently

If we ran this evaluation again, we would:

1. Start with our own document inventory. Before evaluating tools, catalog every form type and variant you process, with realistic volume counts. Two of the platforms would have been disqualified early if we had done this first.
2. Test on our worst documents, not our best. The initial demo results were misleading because we used clean, well-scanned forms. Real accuracy shows up when you feed the system the faxed, handwritten, coffee-stained documents that your actual workflow contains.
3. Measure end-to-end time, not just extraction accuracy. Extraction accuracy matters, but the real question is: how much human time does each tool save per document, including the time spent reviewing and correcting extraction errors?

None of these three platforms are perfect. All of them are better than manual data entry for typed ACORD form fields. The choice between them comes down to your team’s technical capability, document volume, budget constraints, and whether you need extraction alone or broader process automation.

Accuracy figures in this article reflect our testing on a specific set of 400+ ACORD forms and may not generalize to all form versions and conditions. Always validate tool accuracy against your own document inventory before making a purchasing decision.