Scott Phillips created Pension Lab to solve the problem of tracking all pension schemes people paid into throughout their careers. His product lets users view all their balances on one dashboard. 

As the customer base grew, registering personal or business account information took 15 minutes per document — and there were thousands of users to process.

Our first Pension Lab project helped cut client onboarding time after we:

• introduced new APIs that communicate with external services,
• developed templates that admins could use to start work faster.

Pension Lab returned in 2024 to ask if AI could help with even faster data processing.

Since the AI MVP’s scope was limited, only 2 AI specialists were needed

This two-man team included a Python programmer, who developed the core application, and a software architect who safeguarded the project’s requirements – a loose relationship with the existing system and a big potential for future scalability.

AI discovery

We’d been working with the Pension Lab for a few years by now and already had ideas about how to best use AI to help the client achieve business goals faster.

We’ve brainstormed these ideas during a series of meetings consistent with our battle-proven GenAI Rapid Prototyping Sprint™, which involves a sprint that accomplishes a number of tasks:

• choosing the best use for AI in your application,
• analysing business-, data-, and architecture-related aspects of the implementation, including possible roadblocks and quick wins,
• delivering the first Proof of Concept within one week,
• coming up with a plan for a full-fledged solution,
• developing a roadmap for further development,
• fitting it all into an actionable GenAI report, complete with process flows as well as time and cost estimates.

A workshop like this is the quickest way to put all of the business considerations we’ve talked about up until now in the context of your organization.

Proof of concept

The Software House reviewed different OpenAI models to find one that was the most cost-efficient and fast enough. GPT-4o mini was the top choice.

We delivered the first Proof of Concept within 1 week to demonstrate the document processing functionality on a small sample.

The OCR solution extracted 15-20 document fields in 40 seconds, while the OpenAI-driven module handled data categorization quickly.

The first build

Pension Lab soon received a plan for a full-fledged solution based on iterative development with weekly demos.

The company chose to develop the OCR and AI-based features and automate document processing workflows.

We added a function that sends scanned documents forward and delivers their summary automatically.

Refinement

To improve the solution’s document processing power, we fine-tuned the OpenAI model for each dedicated document and Letter of Authorization (LoA) provider.

Next, the team added a feature that anonymizes sensitive data before OpenAI processes it and then matches the result with the original data set when the model returns it.

Before the Python-based application’s release, we created a workflow that would allow Pension Lab to improve data extraction capabilities in the future.

The AI-powered Python application we developed to automate document processing met all requirements for the solution:

• it could scale with user growth,
• it minimized the parsing time for 1 document from 15 minutes to 40 seconds,
• it’s easy to decouple and its text data analysis capabilities have great potential for future refinement.

Thanks to data anonymization, Pension Lab can also assure users that their data is never exposed to third parties, and no LLM models can ever use it for training.

Stakeholders approved of the solution.

Technology choice

Python, OpenAI, AWS, Serverless, AWS Lambda, AWS Step Functions