Why Faster AI Doesn’t Always Mean Less Work

Imagine a broker asking AI to draft a reply to a lender. The lender needs an updated loan amount and an income document. The draft takes a few seconds. Writing it by hand would have taken longer.

Before asking, though, the broker had to find the borrower’s latest email and explain which amount it replaced. Afterward, they still have to check the draft, choose the attachment, and make sure the message goes out. If the lender replies with another question, they may have to explain the file all over again.

The faster draft helps. But it is only one part of the job, and much of the effort sits around it. As AI gets faster, that surrounding work deserves more attention.

The work around the answer

A person who delegates a task still has responsibilities. They have to explain what needs doing, supply the right information, decide whether the result is usable, and connect it to whatever happens next.

With AI, some of that work is easy to overlook. Copying an email into a chat takes a moment. Explaining which document is current takes another. Moving the answer into the right place seems minor. Spread across a day, those moments can leave someone spending much of their time briefing and supervising software.

More output can add to the load. Three alternative replies give the broker a choice, but also three replies to assess. A summary may be useful, yet still need comparing with the original email. Each result arrives quickly; someone has to decide what to do with it.

This is how individual tasks can get faster without the whole job getting much easier. The person remains responsible for keeping all the pieces together.

When attention becomes the limit

Anthropic’s October 16 study of its own workplace describes a version of this problem. As agent output grew, human attention became a bottleneck. The researchers estimated that reading all their agents’ output would take even non-technical employees half their working week.

That is one company’s experience, but the question it raises reaches further. What happens when producing work becomes much faster than people can absorb and act on it?

For a mortgage broker, reading every new summary is competing with returning a borrower’s call, resolving a lender’s question, and deciding which file needs attention. A tool that saves writing time can still demand more of the broker’s attention elsewhere.

The answer cannot simply be to check less. Our previous article explored why evidence and meaningful review matter. The next challenge is to make that review easier to carry out as AI takes on more work.

What would make the job easier?

Return to the broker’s reply to the lender. The broker has already connected the borrower’s email to the file and established which amount applies. A useful assistant should be able to carry that understanding into the next task. The broker should only need to revisit it when something changes or a new question arises.

That expectation is shaping our work on Clara. We’re developing ways to connect email conversations to a deal so Ask Clara can use the correspondence when answering questions about it. The aim is to spare the broker from repeatedly finding and explaining information they have already provided.

Review also becomes easier when the draft arrives with the relevant evidence. Put the proposed amount beside the email that supports it. Identify the document being attached. If something is unresolved, explain the specific question. The broker still reads the response and exercises judgment, but has less detective work to do first.

Then the task needs a clear outcome. Was the message sent? Did it fail? Is that still uncertain? We’ve been working on scheduled-email alerts in Clara for this reason. An alert should bring the broker back to the message that needs attention, with an accurate account of what happened.

These are modest examples of what we’re building toward. They address the effort of keeping a task moving: remembering the context, finding the basis for a decision, and following up when something needs help.

Count the whole task

A useful way to evaluate AI is to follow a task from the moment a person starts preparing for it until they can move on. Include the time spent gathering information, explaining the request, reviewing the answer, making corrections, and checking the outcome.

For Clara, we want to understand how often a broker has to repeat information, how much a draft needs changing, and how many unfinished tasks require another visit. Those questions would tell us more about the working day than the speed of the first response alone.

Faster answers matter. They become more valuable when the work around them gets easier too. For the broker in our example, success means answering the lender’s question and getting back to the borrower, with fewer loose ends to remember.