Plain-English Briefing
Barclays is using AI to sort emails. Start with one handover.
On 1 October, Anthropic announced an expansion of its work with Barclays. Alongside software development and internal knowledge tools, the announcement describes a familiar office job: making sense of incoming email and getting it to the right person with the right information.
According to Anthropic, a platform in Barclays’ Global Markets business processes approximately 120,000 emails a day, using Claude to classify messages, add useful information and decide their processing route. Those are supplier-reported figures about an existing use case, rather than an independently verified savings study. Read the announcement.
For a smaller business, that is a useful prompt to look at what happens after an enquiry arrives. How many times does somebody read it, copy it or chase the detail that lets the next person start?
Follow the email into the work
Take a fictional sign-making business. A customer emails a request for six workshop signs, attaches an old drawing and asks for delivery next Friday.
Someone reads the message, finds the customer, checks which drawing is current, enters the quantity into a spreadsheet and forwards it to the workshop. The workshop needs dimensions before it can price the job. Another email goes back. When the quote is accepted, somebody copies the agreed details into the job sheet.
There are several different tasks here. Software could reliably carry agreed customer details and quantities between records. AI might help interpret a loosely written request or identify information in an attachment. A person still needs to resolve an ambiguous drawing or agree a delivery promise.
A useful first system might create a draft enquiry, attach the original message, flag the missing dimensions and put it in a review queue. Once somebody checks it, the agreed details become the record used for the quote and the job.
The improvement to test is straightforward: does the handover leave the next person with enough correct information to continue?
Decide what should happen when it is wrong
Before trying that workflow, decide where uncertain information goes. A guessed dimension should stay visibly unconfirmed. A message with two possible customer matches needs review. An unreadable attachment should remain attached, with a clear reason the draft is incomplete.
Keep the original message available beside the draft. Make corrections easy, and make the person responsible for the handover clear. That gives the team a way to use the tool when the request is awkward, rather than only when the demonstration is tidy.
For the access side, our earlier briefing on AI permissions covers what a tool can read, change and record. The workflow still needs its own agreed rules about what counts as ready.
A subscription does not settle the handover
It is worth checking the tools you already have. Your current package may support an enquiry form, a shared queue, required fields or an integration that removes the copying. A standard product may already handle the whole process well.
If those options fit, use them. If an important step remains missing, a small connection between existing systems may be enough. A bespoke system becomes worth considering when the job needs rules, views or handovers that the available products cannot usefully support.
AI can earn a place where the input needs interpretation. Predictable calculations, record updates and approval rules can remain ordinary software. The choice follows the work you have measured.
Worth doing this week
Choose twenty recent enquiries of one type and review them within your existing authorised systems. You do not need to upload them to an AI service to do this exercise.
- Trace each one from arrival to the point where someone could act on it. Note who handled it and where information was entered again.
- Record the missing details, corrections and exceptions. Include the awkward enquiries as well as the clean ones.
- Identify one handover you could improve. Write down what a complete, checked record would contain and who would approve it.
- Test the smallest change that addresses that handover. Compare handling time, correction work and missed requests with the original process.
Released staff time can be valuable without immediately becoming money saved. Measure the checking and correction work as well as the initial entry; otherwise a quick draft can look better than the working day it produces.
If your spreadsheet is coordinating quotes, jobs and stock, that exercise may point to a clearer business system. If one repeated task is the problem, a small automation may be sufficient.
Bring us one enquiry and show us where it goes next. We can help work out whether the sensible answer is a better process, an existing product, a connection or a build.
Talk to Inferred about one repeated task →
Where this came from
- Anthropic: Barclays scales Claude to upgrade operations and improve client experience — 1 October 2026. This is the technology supplier’s announcement. The practical workflow example and suggested review above are Inferred’s editorial interpretation, not a description of Barclays’ internal system or an Inferred customer case study.