AI for Learning & Assessment

The Problem

When UCL rolled out Microsoft Copilot M365 licences to staff, access alone did not translate into confident or effective use. Staff had the tool available but lacked structured support to engage with it meaningfully - and uncertainty about responsible use, data privacy and what the tool was actually for meant uptake was inconsistent across departments.

The Solution

Two parallel resources were developed: the AI Efficiency & Ethics Workshop (a participatory, practical session delivered six times across UCL - detailed below) and a standalone Copilot Guide infographic, designed to make the offering legible to staff who had never engaged with an AI tool before. The guide explains what Copilot is, where to access it, what it can be used for, and what responsible use looks like - without overwhelming staff with feature lists or technical jargon.

Outcome

The workshop series grew to six institutional deliveries and has been re-invited for further sessions into 2026. The Copilot Guide became a standalone staff resource, distributed alongside workshop delivery and embedded in onboarding materials for new staff. Attendees consistently reported moving from uncertainty about the tool to practical confidence in applying it to their own workflows.

The Problem

As across many large digital learning environments, image descriptions in Moodle course materials were inconsistent and difficult to improve at scale - creating barriers for disabled learners and a compliance risk under the Public Sector Bodies Accessibility Regulations. Manual retrospective auditing was resource-intensive and did not address the problem at the point of authoring.

The Solution

Built using the standard Microsoft 365 Copilot agent experience on a standard licence - approved by the UCL Digital Accessibility team and shared organisation-wide across UCL. The agent improves accessibility at the point of authoring, not retrospectively.

Outcome

Live AI accessibility agent deployed at UCL STEaPP. Staff guidance materials and a responsible AI use framework published. The workflow supports improved accessibility practice by helping staff produce reviewable first-draft alt text at the point of authoring, reducing the friction of starting from nothing while keeping human judgement in the loop.

I also support colleagues and student project teams with the less visible parts of AI agent development: environment strategy, publishing routes, authentication, guest access, licensing, ownership and handover.

A recent example involved a UCL student project team building a Copilot Studio agent for an external partner organisation. The agent had started in the default Power Platform environment, which was fine for early prototyping but not ideal for governance, external testing or eventual ownership. I tested a safer route using a simple agent first: exporting the agent through a Power Platform solution, importing it into a dedicated sandbox environment, enabling Dataverse, and investigating what external guest users could and could not access. The key finding was that migrating an agent into a dedicated environment improves governance, but does not automatically give external guests editing access in Copilot Studio.

For rapid prototype testing, I identified a workable route using a published Demo Website with no authentication. This allows external stakeholders to test the agent and provide feedback, but it has to be treated carefully: the site is unlisted, not private, so anyone with the link can access the prototype. This work helped separate three different needs that are often confused: internal development, external testing and final ownership.