Two addresses on an insurance form: can AI tell them apart?
An employee enters information from an insurance form into the system. The tool reads the address without a typo, but also puts the company’s registered office address in the warehouse location field. Correcting it requires returning to the document. Useful assistance shows each address alongside the relevant part of the form so the employee can check the data before saving it.
Syntalith
Keep the address with the right part of the form
Suppose a business property insurance submission contains two different addresses. The first appears in the company details section and identifies its registered office. The second appears in a separately labeled section for the warehouse being submitted for insurance. After extraction, the employee sees the registered office address in both places.
The correct record takes the warehouse location from the warehouse section. Each value points to the passage it came from. The employee can check the address and correct its placement without searching every page. Reading the characters is only part of the task; the tool also needs to recognize what the entry refers to.
If the warehouse section is blank, the registered office address does not fill that gap. The employee needs to see that the location requires clarification. Preparing information from the form does not determine acceptance for insurance or the scope of coverage.
Multiple locations need to stay together
Another form may describe two warehouses in separate sections. The employee then needs two groups of information, each containing the address and description for the relevant location. A list of extracted addresses and a separate list of descriptions leaves them with the work of reconnecting the details.
The employee should be able to review and correct each location separately, including when the same section repeats on later pages.
Check whether the receiving system can store that information. If the form contains several locations but the planned import allows only one, improving the model alone will not resolve the problem. The project also needs to address how those groups are saved in the destination system.
What to check before training a model
For a consistent form layout, an existing extraction template may be sufficient. If the company collects responses through a web form, separate fields for the registered office and property locations can reduce the need to extract information again. Start with the capabilities already available in the intake system.
For scans, distinguish incorrectly read text from correct text placed in the wrong field. OCR converts a page image into text. If it misses a section heading, the subsequent address assignment may rely on incomplete material. Check the document reading stage before assuming that understanding the address is the problem.
Compare available extraction tools on forms the team regularly corrects. Google describes Custom Extractor using a foundation model and adaptation options, including fine-tuning, that depend on the processor version. Compare its output with the corrections your team makes before deciding whether to adapt a model.
When adaptation may help
Fine-tuning means training an existing model further using examples with the correct information identified. Here, it is worth considering when readable forms repeatedly lead to confusion between the registered office and property location, or to details from several properties being mixed together. The problem is a recurring failure to associate fields with their meaning across different layouts.
The intake team contributes its understanding of what belongs in each field. Employees show the corrections they make today and explain how they separate locations. Without that understanding, a model may reproduce ambiguities in the current field descriptions. Also establish which information is simply absent from a document and should remain for follow-up.
Assess the trial on separate forms and versions that were not used for adaptation. Employees check misplaced addresses, mixed location details, and omitted values, along with the work needed to correct the whole record. Comparing this with the existing extraction helps establish whether adaptation warrants additional maintenance.
Reviewed information in the existing system
Syntalith proposes an application for extracting insurance form data that shows employees the values and their source locations while keeping properties separate. Once reviewed, the result moves to the agreed stage of intake in the existing system. This addresses the copying and reconnecting of information the team currently performs by hand.
We start with forms employees regularly correct. Your team explains what belongs in each field and agrees on how materials can be used. We assess the current tool and the import into the system, proposing model adaptation for errors that continue to recur.
Describe one form where correctly read information ended up in the wrong place. Explain how the employee repaired the record and where the result should go. That gives us a starting point for discussing the project; engagement information is on Syntalith’s pricing page.
Match a model to the task you need it to perform
Describe where your current AI falls short. We will compare model customization options, data requirements and the cost of running the resulting system.
Private LLMs and fine-tuning