ECC provides a specialist HR platform used by universities and colleges across the UK to score job roles and determine pay grades. The process relies on trained role analysts answering 50 structured questions for each role, with scores weighted through ECC's own proprietary algorithms to produce a final result. The multi-stage process, involving first scoring, second scoring, and an approval step, is thorough but rightly demands time and resource, and the stakes are high: the outcomes directly influence staff pay grades and are subject to scrutiny from academic unions.

With AI adoption accelerating across sectors, ECC recognised both the opportunity and the pressure to act. A research project at one of their partner universities had already demonstrated the theoretical possibility of AI-assisted scoring, but the academic prototype was never production-ready. ECC needed a trusted development partner to take the concept further, building something commercially viable, technically robust, and appropriate for a regulated, unionised environment.

Propel Tech worked with ECC to design and build an AI assistant embedded directly within the ECC platform. Rather than replacing the role analyst, the solution supports them: as a user works through the 50 scoring questions, the AI analyses the uploaded job description and offers a suggested response alongside its reasoning for each question - helping the user to make faster, better-informed decisions.

The project began with a rapid proof of concept, with Propel Tech demonstrating a working AI integration within days. This early prototype - connecting to an AI model via AWS Bedrock and returning a response to a single question - gave ECC the confidence to move forward.

The build phase involved careful alignment with ECC's existing data structures and proprietary scoring algorithms. Propel Tech ensured the AI's output mirrored the same formats used for human-scored records, allowing it to sit seamlessly alongside existing workflows rather than introducing a parallel system. Significant effort also went into prompt engineering - iterating and refining the instructions given to the AI model to achieve the accuracy required for this sensitive use case.

Initial development used Amazon Nova via AWS Bedrock, but after rigorous evaluation across a wide sample of scored records, Propel Tech switched to the Claude model within the same AWS Bedrock infrastructure. The results were markedly more accurate across all 50 questions, and the improvement justified the additional cost, which Propel Tech supported ECC in evaluating through detailed billing projections and evidencing.

The solution was designed with human oversight at its core. The AI never creates or submits a score record independently; it is always the role analyst who makes the final call. This approach was a deliberate response to the sensitivities of the sector, preserving trust while supporting a more streamlined and consistent evaluation process.