Barclays, the British bank, is expanding its use of Claude across the bank, and it put numbers on what the technology is actually doing. The two jobs it describes with numbers are jobs your business also has, finding answers inside your own documents and sorting the email that arrives every day. That is the pattern worth copying.
The first is a system Barclays calls the Colleague Knowledge Assistant, which helps its UK staff find information and answers faster when they are supporting customers. It has been live since 2025, more than 16,000 colleagues use it, and it has handled over one million searches. It runs on retrieval-augmented generation, which means the AI looks things up in the company’s own documents and answers from what it finds there. For you, that is the important part: the answers come from your prices, your policies, your procedures, not from general knowledge.
The second sits in Barclays’ Global Markets business. Claude models read incoming client emails, classify them, add information to them, and decide where each one should go. They also check whether a request contains what the operations team needs in order to act on it. That runs at roughly 120,000 emails a day, and it cuts manual handling.
On the engineering side, Barclays expects Claude Code, Anthropic’s coding tool for developers, to reach 50% of its developer population by the end of 2026, and a majority of its software engineers in 2027.
What to copy
Neither of those two systems is a clever new product. One is lookup. The other is triage. Both are jobs where people spend their day moving information around instead of deciding anything. If you are looking for a first AI project, start there: the pile of repetitive requests that lands every morning in email, forms, or WhatsApp, and the same questions your team keeps asking you.
One caveat. Barclays runs all of this with governance, security controls, and human oversight. Scale that down for your size, but do not skip it. Anything that reads customer messages or answers from your documents needs a person who checks samples of the output, and a clear line on what it is not allowed to decide alone.
