Earlier this year we ran a three-month pilot of Anthropic’s Claude with 26 colleagues, from a representative sample across Homeprotect. Being an FCA-regulated home and contents insurer, we wanted to balance AI innovation with safe, responsible and governed use of AI and to accrue early experience with this technology across the board.
Having completed a GenAI training course at the end of 2025 focused on context engineering and understanding basic LLM principles, participants were given access to the full Claude toolkit in Q2 2026. This included Claude in the chat and desktop apps for everyday knowledge work, Claude Cowork for agentic desktop tasks, and Claude Code for software delivery.
We went in with a simple question: does this genuinely change how we work? Here are our findings.
The successes
Adoption told its own story. By the end of the pilot the overwhelming majority of participants were using Claude multiple times a day, and every single respondent to our closing survey rated the experience good or excellent. The median participant estimated they had saved around a week of effort over the pilot, and several put the figure at a month or more.
What surprised us most was the breadth of use cases. Code development was the single most common usage, but it was closely followed by research and information gathering, brainstorming, and drafting documents, emails and reports. Approximately half of participants used Claude to surface insights from collections of documents or to create charts and architecture diagrams.
In everyday knowledge work, colleagues used Claude Chat to turn meeting transcripts into minutes and sprint-ready user stories, to summarise long email chains down to the points that mattered, and to transform informational content into polished, board-ready presentations. One participant described finally being able to “bring things to life” visually in a way they never could alone.
In engineering, the results were the most measurable. A brand-new customer renewal journey went from blank page to deployable inside a single two-week sprint. A dependency security remediation, normally a week-plus of careful work, was done in days. And in one fully logged example, a production feature for our claims-handling platform was designed, built, tested and reviewed in about 70 minutes, shipping with 15 new tests. Crucially, the guardrails held: everything went through branches, pull requests and human review, which caught two defects in that same example.
The learnings
Firstly, AI output earns trust through review and not by default. Whether it is code, a summary or a slide, staff members consistently validated before relying on it, and that behaviour proved to be correct.
Second, Claude is strong at the first 80–90% of a task. The final stretch of refining a solution, judging nuance, and working in sensitive legacy areas still belongs firmly to employees with full context of the task at hand.
Third, the biggest gains came from people who treated Claude as a collaborator to be challenged and redirected: planning first, producing second, reviewing in a separate pass.
What we would do differently
Delivered more focused and targeted training. Our structured learning resources leveraging Anthropic’s Skilljar academy arrived a few weeks into the pilot, by which point many participants had already formed their own habits, both good and bad. For our broader Claude rollout, training started before day one.
We are also clear that scaling AI tooling responsibly means scaling the controls around it. Governance, supply-chain assurance and access management all need to grow with adoption, and that work is a clear part of our roadmap.
Next steps
We are now moving from pilot to staged Claude rollout: broadening access beyond the initial pilot group and beyond technology, embedding day-one training into onboarding, building a shared library of reusable plug-ins, prompts and skills so good patterns spread rather than staying with individuals, and hardening the governance framework that sits around all of our AI use cases.