nib develops Cortex-powered AI to streamline data migration

nib Group has built a set of AI-powered workflows to support a complex data migration program following its acquisition of NDIS provider All Disability Plan Management.

The ASX-listed health insurer used large language model (LLM) functions available in Snowflake Cortex to automatically extract structured insights from unstructured text such as emails and call notes, and map them into the new system fields.

The project saw nib Group create prompts using Snowflake Cortex’s integration with Meta’s LLaMA 70B.

Speaking at Snowflake Summit in San Francisco, Pip Sidaway, a data product manager at nib Group, said her team effectively “wanted to see how much automation [they] could use to get all these different types of data” into a new consolidated system.

Six months’ work cut

nib Group first started using Snowflake in 2018, but recently migrated everything off its legacy data platform and moved fully onto Snowflake’s data warehouse platform.

Facing a complex migration, the group identified two key use cases for AI that could be developed entirely within the Snowflake ecosystem.

One was to streamline and classify up to 30,000 daily emails into single threads; another was to update over 16,000 free-text fields from legacy records.

To tackle email workflows, nib had multiple inboxes set up to handle incoming messages, but customers would often search the company name and send emails to whichever address appeared, including “no-reply” addresses.

“We wanted to see if we could identify the high-priority emails as well as classify based on this content,” Sidaway said.

“We wanted to collapse that into a single thread as well, so that we could reduce the handling time and duplication of work.

“We wanted to tag emails to categorise them, so [they were] going to the best people to help them.”

To create an AI prototype, Sidaway said her team “ended up trying out different models to see the sort of responses we got. So, we did find that the smaller models didn’t really provide a consistent output.”

As such, the team opted for a 70-billion parameter model that fell within nib Group’s budget for the project.

“We gave it all the categories so that it knew based on the context as to what we wanted to classify, and then we told [it] how we wanted the output as well, she added.

Initial testing with 100 records was successful. However, when scaled to 10,000 records, the model frequently failed due to output formatting errors.

According to Sidaway, this was attributed to the LLM not returning a valid JSON and instead returning empty strings or malformed output that failed parsing due to syntax errors.

This was corrected using fallback logic and more defensive prompts to handle cases where the LLM returned empty or malformed output.

The testing phase also reinforced the need for human oversight.

“Don’t just let AI go crazy without checking it,” Sidaway told the audience. “In our emails, when you look at the actual code behind the emails, it’s a lot of HTML tags in there. So, we set up a Streamlit app.

“That enabled us to actually look at what the email was and the other to see if the tag that was applied to it was correct.”

The results from the program enabled nib Group to collapse 20 percent of its emails into unified threads, reducing duplication and improving customer traceability.

Sidaway added that the team also successfully tagged approximately 60 percent of high-priority emails, making them easier to identify and respond to.

Most notably, however, the team used AI to update 10,000 fields to help identify records where third parties are authorised to act on behalf of a customer.

“If we didn’t use AI, someone would have manually had to go in and cut and paste it between systems, and so we estimate that was going to be about six months’ work,” she said.

Looking ahead, the group is now exploring using Cortex Analyst for streamlined data discovery and semantic models for better AI responses, using data lineage from Select Star to inform semantic models and improve output accuracy.

Eleanor Dickinson travelled to the Snowflake Summit in San Francisco as a guest of Snowflake.