What Should You Delegate to AI? Start Small with Lessons from Barclays

Author: Lucy.·Thursday, October 1, 2026
What Should You Delegate to AI? Start Small with Lessons from Barclays

What should your company hand over to AI first? Consider the work that happens before someone can make a decision: reading incoming emails, finding the right colleague, or searching for information to answer a customer.

Anthropic’s October 1, 2026 announcement about British bank Barclays offers a useful starting point. The lesson is to break a workflow into steps, decide which parts AI should handle and how much responsibility it should have, and give people a clearer focus.

What work has Barclays given AI?

Two examples stand out: internal knowledge search and email processing. Both help employees gather information and prepare for the next action.

Barclays UK’s Colleague Knowledge Assistant has been live since 2025. According to the announcement, more than 16,000 colleagues have adopted it, and it has handled over one million searches in total. It combines Claude with retrieval-augmented generation, or RAG: a system first finds relevant material, then uses it to help the AI produce an answer. The assistant helps employees find information and answers more quickly when supporting customers.

In Global Markets, Barclays uses Claude to classify incoming emails, add information needed to process them, and determine where they should go next. The announcement says the platform processes approximately 120,000 emails a day. “Processes” here includes classification, enrichment and routing. It does not mean AI automatically replies to all 120,000 emails or resolves every request from start to finish.

Looking at these examples as workflows makes the opportunity clearer. “Customer support” includes several distinct steps: finding information, classifying a request, connecting it with the right person, and deciding how to respond. Look for repetitive steps, long waits, or repeated searches for the same material. That gives you a more specific place to begin.


Less preparation can leave more room for judgment

Before writing a customer reply, an employee has to understand the request, find supporting information, check what is missing and decide who should handle it. When that preparation takes too long, it crowds out time to understand the customer’s situation and consider a solution.

If AI gathers and organizes the relevant information, people may be able to give more attention to exceptions, priorities and the explanation a customer needs. They may also have more room to spot recurring problems and plan improvements to guidance or procedures. That is a useful direction to take from this case.

A small business, for example, could start with “read a quote request and extract the requirements, preferred dates and questions we still need to ask.” An employee would review the result and decide on pricing, delivery dates and the response. This is an illustration of the approach, not a description of Barclays’ implementation.

Decide how far delegation should go, too. At first, AI could suggest a category and a person could confirm it. For request types where results prove reliable, the next step might be routing them to the right colleague. Keeping separate approval steps for external replies and decisions affecting costs or contracts makes responsibilities clear.

Anthropic reports faster access to information, quicker customer support and less manual handling at Barclays. However, the announcement does not disclose a time-saving percentage, classification accuracy or monetary savings. More time for human judgment and planning should therefore be a goal to test in your own pilot, rather than an outcome assumed from this announcement. Barclays also says it applies security controls and human oversight.


Start small, then expand what works

One workflow in one team is enough to begin. A good candidate repeats often, has an outcome you can describe clearly, and can be reviewed by someone responsible for the work. A narrow scope, such as classifying inquiry emails or searching internal guidance, makes it easier to see what has improved.

First, record how long the current process takes. Then test AI using materials and tools your company permits, and track three things together:

  • Total processing time: Include human checking and corrections, not just the AI’s running time
  • Correction effort: Record which categories or summaries need frequent fixes and how long those fixes take
  • Missed requests: Check separately for overlooked inquiries and missing essential information

Average handling time alone can hide unusual cases. Test both routine and ambiguous requests, and distinguish the types AI handles well from those that need a person. Compare subscription or usage fees alongside the time needed to operate the process to see whether it fits your budget.

Once the trial demonstrates a benefit, widen the scope gradually: from one request type to a similar one, or from suggesting categories to routing requests. Define approval authority, permitted information sources and the existing process people can return to when an exception occurs. This is a suggested approach informed by the case, not a reproduction of a pilot method Barclays disclosed.

It also helps to look beyond “the number of items AI processed.” Check whether employees can actually focus more on complex customer problems and planning the next piece of work. That gives expansion a clearer purpose.


Find a first step that fits your work and budget with AIB

AIB is a new startup helping people choose AI options that fit their work and budget, so differences in access to information and financial resources create fewer barriers to using AI. It aims to be an independent comparison platform without favoring a particular AI provider.

Improving how you work with AI, often called AI transformation or AX, can begin with reviewing your current ChatGPT or Claude plan, or trying AI on one repetitive task. If you are unsure where to start, whether the cost makes sense, or how to get more from AI you already use, talk to AIB. Other questions and concerns about using AI are welcome, too.

Contact AIB about AI adoption, costs and practical use


Source