What Skipton says it has done
Skipton is a member-owned mutual with 82 branches. Its stated purpose is helping people buy a home, save for the future and improve their financial wellbeing. Its own site describes it as a UK building society offering mortgages, savings and financial advice, though that quote comes from another page's wording, so the member-ownership and branch details here rest on Microsoft's account and Skipton's About page.
Three things stand out in Microsoft's account:
- Broad Copilot access. Skipton started with 30 colleagues from across the business who got licences to explore the tool. Microsoft's account says this was not a formal pilot. The group compared notes and surfaced use cases. Group CIO Jenny Wood says that gave the Society confidence it had "both the governance and the innovation lens in place" before it opened Copilot to the whole organisation.
- An adoption figure. Wood says 80% of colleagues used Copilot Enterprise in July 2026. No definition of "used" is given. It could mean one prompt in the month or daily use, and the story doesn't say which.
- Training outside head office. Skipton ran dedicated AI days, from introductory Copilot sessions to more advanced talks on building AI-powered tools.
Q: a retrieval tool for the core banking system
Q is the most technical part of the story. Two Skipton engineers built it as a multi-agent retrieval tool on Azure Foundry, using GitHub Copilot. It runs as a chatbot. Business analysts, architects and developers use it to ask what a business rule in the core banking system means, how parts of the system fit together, or where a proposed change might need to go.
The headline example is a review of how a particular type of building address was handled in the core banking system. In early 2026 that took several days. In July, in an internal accuracy comparison, Q returned the answer in four minutes.
Treat that carefully. The story doesn't say what the address type was or how the comparison was run. It also doesn't say whether the answer was independently checked, or what Q's overall accuracy is. One anecdote does not show Q replaces a multi-day process in general.
Charlie Moran, Head of Data Capability, is upfront about that limit. He says accuracy "is never going to be 100%", but that human accuracy isn't either. The tool can generate code and suggest solutions. Staff still decide which problems are worth solving and how to solve them.
Several things are not disclosed:
- the models behind Q
- its retrieval sources
- how access is controlled
- how it was evaluated
- whether it can touch production systems
Microsoft's story also mentions an Azure-based Databricks data environment as part of the wider transformation. It doesn't say Q uses it, so nobody should assume it does.
Guardrails
Wood describes "freedom with limits". Existing controls mean colleagues cannot put data or IT security at risk. A change can't be rolled out beyond 10 or 15 people without talking to the right team. In her words, staff can all have approved tools, but there are rules on how they're used, "and we will monitor".
The story says approved AI tools sit inside the Society's wider framework for data protection, governance and third-party oversight. Teams keep visibility of emerging use cases.
The story doesn't say whether the 10-to-15-person threshold is a formal written policy or a technical control. It doesn't say which team must be consulted either. For IT admins, the pattern is a familiar one:
- Approve the tools.
- Allow small-scale experiments.
- Require a conversation before anything grows.
That is a sensible way to avoid shadow AI. The cost is that the checkpoint depends on people actually raising their hand.
Cultural framing: "Enabling Curiosity"
Under a programme called Enabling Curiosity, staff are asked whether a task should be eliminated, reduced or redesigned before deciding to automate it. Automation is one possible answer, not the default. That ordering is good practice. Automating a bad process just produces bad results faster.
The "not about cost-cutting" claim
Moran says the work was never approached from a cost-saving angle. He contrasts a team of 30 producing the same output with 15 against the stated aim. That is a statement of intent, not a staffing guarantee.
Skipton's implementation partner, Version 1, published its own account in July 2026. It says over 650 colleagues from across Skipton attended a Copilot internal enablement day. It also describes work from readiness and governance through to enablement and rollout. In that account Moran says the Society wants to talk about "hours invested, not hours saved". The aim is for the same number of people to deliver more output and value for members.
Version 1 is a vendor with a commercial interest, so that account corroborates the training effort rather than the outcomes.
Wider context
Search results show other Skipton technology work with Microsoft:
- KPMG describes a Microsoft Dynamics customer engagement platform at Skipton, including Copilot AI, Power Automate and DevOps tooling.
- CGI describes work to transition appropriate services from on-premises infrastructure to Microsoft Azure.
Those are partner case studies with their own marketing angles. They suggest Copilot sits on top of a long-running Microsoft-heavy modernisation, not in isolation.
What's proven and what isn't
| Claim | Status |
|---|---|
| About 80% of staff used Copilot in July 2026 | Company-reported, no usage definition |
| Q answered one address-handling question in four minutes | Single internal example, method undisclosed |
| 650+ attended an enablement day | Reported by Skipton's partner Version 1 |
| Customer or productivity gains | Not measured in the sources |
| Cost-saving was not the goal | Executive statement of intent |
Takeaways for IT teams
- Start small and informal. Skipton's 30-person group helped it find use cases and test governance before a wider rollout.
- Train beyond the technology teams. Skipton ran AI days for staff across the country, not only at head office.
- Define adoption metrics. "80% used it" says little without a threshold. Track depth of use, not just licence activity.
- Keep humans in the loop for internal knowledge tools. A retrieval assistant can speed up investigation, but its answers need validating, especially for code and business rules in core systems.
- Set an escalation threshold. A clear point where small experiments need review is easy to explain and easy to audit.
Skipton's story is a credible outline of a measured enterprise Copilot rollout. The measurable proof, such as accuracy rates, member outcomes and financial effects, isn't in the public record yet.