CSCS – AI Certificate Scanning on AWS

About CSCS

The Construction Skills Certification Scheme operates the UK’s construction skills card scheme. A CSCS card proves that the person holding it has the training and qualifications for the work they do on site, and for most UK construction sites it is a condition of entry. More than two million workers hold one.

StableLogic has worked with CSCS for many years, delivering the My CSCS mobile app, card verification APIs and the technology behind the scheme’s digital skills passport programme.

The Challenge

CSCS offers a large number of card types. Eligibility for each is defined by hundreds of rules, and an applicant needs to satisfy only one of them to qualify. That flexibility is good for workers, but it makes the scheme genuinely hard to navigate from the outside.

In practice, applicants could rarely work out which card they were entitled to. They had to interpret their own qualification certificates — awarding body, qualification title, certificate number — and key the details in by hand. The consequences were predictable:

• Many applied for a lower card than their qualifications actually supported
• Many applied incorrectly and needed the application corrected
• Some abandoned the journey altogether
• Applicants with lower reading confidence found the process particularly difficult

For CSCS this meant long completion times, avoidable support contacts and rework for scheme staff. For workers it meant being held on lower-value cards than they had already earned.

The AWS Solution

StableLogic designed and built an AI agent that removes the interpretation step entirely. The applicant photographs their certificate; the agent does the rest.

How it works

In the native iOS and Android CSCS app, the applicant takes a photograph of their qualification certificate. The image is passed over an authenticated API to an agent orchestration service running on Amazon ECS, which drives an agent loop using Claude on Amazon Bedrock with native tool calling through the Bedrock Converse API.

The model is given a fixed set of tools — find awarding body, find qualification under awarding body, and match against card rules. These are backed by Amazon DynamoDB for high-read lookups of awarding bodies and qualifications, and Amazon RDS for the relational rule data. The agent extracts the identifiers from the image, calls those tools iteratively to resolve the qualification, reasons over the rule matches, and returns the highest card in the scheme’s predefined hierarchy that the applicant qualifies for, together with a pre-filled application.

Why Amazon Bedrock

The task needed one model that could do three things at once: read certificates of varying layout, condition and image quality; extract structured data reliably; and carry out multi-step tool use with reasoning over the rule results. Claude on Amazon Bedrock met all three.

Amazon Bedrock was chosen over self-hosted alternatives so that CSCS would not be managing model infrastructure, so that data stays within AWS, and so that tool calling is handled by a managed service rather than by fragile text parsing. Processing runs in eu-west-2 so applicant data remains in the UK.

The human stays in control

The agent does not submit applications. The pre-filled form is returned to the app for the applicant to review and submit through the existing journey, so nothing is lodged without human confirmation. Where a certificate cannot be matched with confidence, the application falls back to the existing manual path rather than guessing.

The agent inherits the applicant’s existing CSCS session token, so no new credential or login flow was introduced and it only ever operates in the authenticated applicant’s own context. Uploaded certificate images are stored encrypted at rest in Amazon S3.

Deterministic where it matters

Card rules are held as a predefined hierarchy rather than left to the model’s judgement. Once the qualification is resolved, the card outcome is deterministic — the AI does the reading and the matching, the rules do the deciding.

The Results

• Application completion time reduced by approximately 70%, measured end to end on the mobile journey before and after the AI-assisted workflow was introduced
• Incorrect applications down by an estimated 30%, reducing corrections, support contacts and manual intervention
• A 20% change in the card applied for — applicants moving onto higher-value cards their qualifications already supported
• Over 80% first-pass card match accuracy, validated before launch against sample certificates with known expected outcomes, with unmatched cases routed to the existing manual path
• No new login or credential surface, with all applicant data processed and stored in the UK