Barclays routes about 120,000 emails a day with Claude. Count your own work before you pick a process

Most owners pick their first AI process by instinct: whatever annoys them most this week. On October 1, Anthropic described how Barclays is using Claude. The announcement named a few specific uses and put a number on every one. If you are choosing your own first processes, our guide to what to hand to AI first holds the method that stays the same. This post covers what changed and what to copy from it.

What Barclays announced

Anthropic said Barclays is extending Claude across the bank “to accelerate software development, modernize legacy systems, and improve operational efficiency.” The announcement gives three concrete uses. All figures are company-reported:

• Software development. Barclays expects Claude Code to reach 50% of its developers by the end of 2026, and a majority of its software engineers in 2027.
• A knowledge assistant for staff. It has been live since 2025. More than 16,000 colleagues have used it, and it has handled over one million searches.
• Email routing in Global Markets. Claude classifies incoming email, adds details, and picks a processing route. The platform handles about 120,000 emails a day.

Anthropic says the bank applies governance, security controls, and human oversight to its AI use cases.

Why the count matters more than the bank

You are not Barclays. You do not have 120,000 emails a day. But look at how the announcement is built. Every use comes with a unit (developers, searches, emails) and a volume. None of them is “use AI more.”

That is the habit worth copying. A process you can count is a process you can check. If you know the inbox gets 200 messages a day, you can see on day three whether the routing is keeping up. If you only know the inbox “feels busy,” you cannot tell whether anything got better.

The email example is also the clearest fit for a small team. Sorting inbound messages is frequent, follows rules, and is easy to check: a person can glance at where a message landed and say right or wrong. That is the profile our guide puts first in line.

The count-it test

Run this on any process before you hand it to AI. It takes about 20 minutes per process. Write one line for each question:

1. Unit. What is the one thing that moves through this process? An email, an invoice, a support ticket, a booking.
2. Volume. How many units a day or a week? Count a real week, not a guess.
3. Routes. Where can each unit end up? List the destinations. If you cannot list them, the rules live in someone’s head.
4. Check. How would a person tell, in under 30 seconds, that one unit went to the right place?
5. Owner. Who looks at a sample of the work, and how often?

Decision rule: if you can answer all five, the process is ready to build. If you are missing the unit or the routes, write those down first. If you are missing the check or the owner, do not automate it yet.

A small worked example

Here is an illustration with made-up numbers, so you can see the arithmetic.

A 15-person firm has one shared inbox. Over one five-day week, the office manager counts 900 messages: 180 a day. They sort into five routes: new inquiries to sales, invoices to the bookkeeper, scheduling to the coordinator, client questions to the account lead, and everything else to the office manager.

Check: each person confirms the messages in their folder belong there. Owner: the office manager reviews 20 routed messages every morning for the first two weeks, then 20 a week.

If sorting takes the office manager about 20 seconds a message, that is 180 × 20 = 3,600 seconds, or one hour a day. The test does not promise that hour comes back. It tells you what to measure: after the build, count how many of the 20 sampled messages were misrouted, and how long the morning review takes. Those two numbers decide whether the process stays.

What not to copy

Barclays is a large regulated bank. The announcement does not say how many people built these systems, what they cost, or how errors are handled day to day. Do not read the developer targets as a benchmark for your team. Read them as a sign that the bank chose a measurable unit (share of developers) for each rollout.

What to do this week

• List the three processes that take the most hours in your business.
• Run the count-it test on each one. Count a real week of volume.
• Pick the one with all five answers, and the simplest check.
• Name the person who will review a daily sample for the first two weeks.
• Write down the two numbers you will judge it by before anyone builds anything.

If you want help building the first few once they pass the test, our Build engagement takes up to five processes a month and builds them with your team so your people run them afterward.