Sam Taylor, AI Implementation Manager at Unilink, and David Purkess, the company's Innovation and Technical Director, used their session at this year's Modernising Criminal Justice conference to share how their own internal AI rollout shaped a set of prison-based AI tools now being deployed with UK prison and probation customers.
Sam opened with Unilink's AI adoption journey. Before rolling anything out, the company surveyed its entire workforce to understand attitudes toward AI, finding around 24% of staff concerned about its impact on their role, alongside genuine curiosity mixed with uncertainty about governance and trust. That insight shaped Unilink's approach: treat this as a people-first initiative, not a technology rollout. Rather than issuing policies or mandating training courses, the company gave staff explicit permission to experiment safely, running a hackathon to let people try tools like Microsoft Copilot, ChatGPT and Cursor on both work and personal projects, and building a curated internal learning platform with role-specific examples. Sam said confidence grew through doing and through sharing - AI forums and team channels gave staff a safe space to ask questions and learn from colleagues who'd already hit the same problems.
She was candid that even Unilink's own senior developers were initially sceptical, but now see AI tools as a partner rather than a threat - trust built through the ability to verify and validate outputs themselves, rather than being asked to accept them blindly. Sam was firm that human accountability and expertise remain non-negotiable throughout, whether in software development or the operational tools built for prisons: the goal is augmenting practical, low-risk, high-value use cases with people firmly in control, not automating high-risk decisions.
David then walked through specific prison-based applications. An AI-assisted notice creation tool lets staff generate consistent, well-formatted communications - covering everything from job adverts to regime changes - through a conversational interface, applying built-in templates and guardrails so the result is always structured and on-brand, regardless of how roughly the request was typed or spoken. The same underlying translation infrastructure automatically converts FAQs, notices, surveys and menus into other languages, useful both for prisoners with limited English and for facilities like immigration removal centres with rapidly shifting populations.
One of the more striking demonstrations was AI-generated food imagery for menu selection, addressing a genuine legal obligation to accommodate dietary, cultural and nutritional needs where text descriptions alone often aren't enough. Rather than relying on stock photography or difficult-to-manage in-cell photography, catering staff can generate consistent, repeatable images - down to specific plate types, lighting and halal or kosher labelling overlays - within tightly defined guardrails that prevent inappropriate or off-brief content.
David also demonstrated a system for turning long, dense policy documents into interactive, question-and-answer format content for prisoners - addressing what staff had previously done manually, one FAQ at a time. The tool ingests something like a 30-page HMPPS policy document, breaks it into categories, generates plain-language summaries with an AI confidence score, and flags anything potentially sensitive for human review before publication. Staff retain full control through a review workflow, and the same content becomes available to prisoners as a self-service voice-and-text assistant, with configurable tone, reading level and vocabulary to support neurodiverse users and those with lower literacy.
Sam closed by returning to the human dimension: the risk isn't AI adoption itself, but organisations standing still while others move forward, given wider estimates that AI could unlock some £140 billion in UK economic output. Her central message was that the organisations that succeed won't be those that simply buy AI tools, but those that invest in helping their people learn to use them well.
Jessica Kimbell, GovNet
