Barclays expands Claude: lessons from an enterprise rollout for smaller businesses

According to Anthropic, Barclays UK's knowledge assistant, in service since 2025, relies on a retrieval-augmented search architecture. More than 16,000 employees have adopted it, and it has processed over one million queries. The press release also describes the handling of about 120,000 emails per day in Global Markets, to classify, enrich, and direct the requests.

Barclays also expects Claude Code to be adopted by half of its developers by the end of 2026. This is an announced goal, distinct from an achieved result. The article comes from the provider; Daillac has not audited these figures nor tested the bank's systems.

A manufacturer looking for its procedures in multiple shared folders can start with a limited document assistant. A service company receiving incomplete requests can experiment with classification and extraction of missing information. These examples are hypothetical: they illustrate processes, without attributing Barclays' results to these companies.

The initial scope should have a reference source, a responsible person, and a verifiable output. For the document assistant, the output can include the answer and the document used. For email processing, it can be a proposed category and a list of missing information, validated by an employee before a sensitive action.

The number of connections or requests is not enough. Measure whether users find the correct information, how many responses need to be corrected, and which requests reach the right person. A quick response that refers to an outdated procedure can increase verification work.

The AI implementation must include this evaluation. The test scenarios must contain difficult situations, not just well-written messages or the simplest documents.

A document assistant must not make accessible information that a user is not authorized to see. A filing tool must not automatically have the right to pay, delete, or transmit files. Link capabilities to responsibilities and keep records that allow understanding of a proposal or action.

For legacy systems, AI can help analyze documents or code, but a migration remains a task of architecture and validation. The guide for progressive modernization complements this point: preserving operations requires defining interfaces, tests, and rollback conditions.

The press release does not provide a full comparable cost for an SME, a resolution rate for all requests, or an independent performance. It does not allow promising a given saving to your organization. The right signal is to ask for evidence by use, then test a limited process before choosing the model and expanding access.

Example of a fictitious framework for an SME: the customer service department is looking for the applicable procedure for a return. The expected output contains the relevant passage, its date, and the source document. If two procedures contradict each other, the assistant points it out and forwards the question to the manager. The first deliverable is a list of twenty questions accompanied by the expected documents, rather than general access to all the documentation.

Anthropic — Barclays scales Claude. Sources consulted on October 1, 2026; the grids and decision examples are an editorial analysis by Daillac.