AI can sort 120,000 emails a day. The hard part is what lands on a person’s desk.

On October 1, Anthropic said Barclays was expanding its use of Claude across the bank. The company says its Global Markets teams use the models to classify and add context to client enquiries, decide the best route, prioritize them and check whether operations staff have enough information to act.

The same announcement says the platform processes roughly 120,000 emails a day. Barclays also says its UK Colleague Knowledge Assistant has been used by more than 16,000 employees and handled more than one million searches since 2025. These are company-reported deployment figures, not an independent study of accuracy, customer satisfaction, or mistakes.

That distinction is worth keeping. A large daily volume proves the system is doing real work. It does not tell us how often the right case reached the wrong queue, how often a missing detail was invented rather than flagged, or whether a colleague had to repair a clean-looking summary before a client felt the difference.

AI inbox tools are often sold as though every message should end in a reply. Most good inbox work happens earlier. It separates a real request from an FYI, spots the missing account number, recognizes a deadline, and sends the case to someone who can actually help.

That can give a small team real time back. It is far better to receive a case with the relevant attachment, stated deadline and a plain reason for its route than to spend fifteen minutes reconstructing a customer’s situation from a long thread.

But triage has a quiet failure mode: the system can make a weak guess feel settled. A case labeled ‘complete information’ may still omit the one fact the next specialist needs. A confident priority label can pull attention from a less flashy request. If the original email and the basis for the classification vanish behind a neat summary, the person handling it inherits the risk but not the evidence.

Start with a narrow, ordinary test. Pick one queue with repeated work: quote requests, customer renewals, vendor paperwork, or internal IT requests. Let the tool sort and prepare context, but keep people responsible for sending, changing records or closing a case until the handoffs hold up.

Then inspect a small sample of the unglamorous cases. Can the recipient see the original message, files and routing reason in one place? Can they mark the route wrong without writing a little essay? Does that correction affect the next similar case? And when the tool decides information is missing, does it ask for the missing thing—or fill the gap with a plausible guess?

Those checks sound fussy until a queue gets busy. They are how a team discovers whether the system reduced busywork or merely moved it downstream to the person who now has to undo a decision they did not make.

Mara Vale’s standard is blunt: if a routing guess is wrong, correcting it should be quicker than explaining it to a coworker. A triage screen that makes people fight its categories will teach them to work around it, then the system will keep learning from its own unchallenged story.

Ivy Chen looks at the receiving desk. The person covering the queue should be able to tell what arrived, what the AI added, what is still missing, and what they are expected to do next. If that takes a thread archaeology session, the tool has not removed admin. It has just made the mess look more organized.

Theo Marlow keeps the evidence boundary clear. Anthropic’s announcement supports the scope of Barclays’ rollout and the company’s reported volumes. It does not establish an error rate or a customer outcome. Buyers should ask for that evidence from their own queue before calling an AI inbox rollout a success.

The best AI inbox tool may never produce a spectacular answer. It may just stop five people from rereading the same message, preserve the one missing question, and send the work to the right place without turning a customer into a case number.

That is enough. Work automation earns its place when it makes the next human’s day less repetitive without making them responsible for invisible assumptions.