Which part of the queue should a manager redraw first?

A generative assistant entered the support chats at a Fortune 500 software firm. A study from the Stanford Institute for Economic Policy Research followed it through the chats of nearly 5,200 agents. During the chat, the assistant suggested how to reply and linked internal documents on the technical issue. Agents with the assistant resolved 14 percent more issues per hour than agents working without it. They closed conversations faster, and they resolved a slightly higher share of the problems.

Where the gain landed

The gain landed unevenly, which is the part a manager can use. The least skilled and least experienced agents improved the most, by up to 35 percent. The assistant had digested transcripts from the agents who were already doing the job well, and it passed those moves on, including moves that are hard to write down. Agents with two months on the job performed like unsupported agents with six months of experience.

The suggestion arrived inside the conversation the customer was already in. The agent could edit the line or leave it while the chat was live. That arrangement changes the workflow. A window in another tab, with the old script left as it was, is a different installation. The institute notes that new information technology often dips before any gain, because the firm has to change the process and the way people work. Firms are often happy with a 1 percent or 2 percent productivity gain from a new system. This rollout showed no such dip. Performance rose over a few months. I read that as a sign the firm put the assistant where the work already happened. The process change was the suggestion sitting in the chat.

Customers came out happier, on surveys and in the language of the conversations. The lift did not come from a peppier tone. Agents already sounded upbeat before the assistant arrived. They solved the problem more often. Agents with access also quit less often, which belongs in the operating result.

Who the system still has to keep

The same account warns firms that treat the finding as a reason to cut the highly skilled agents. The assistant spread what those agents were already doing well. Remove them and you remove the practice the assistant learned from. Keep them, pay them, and let the assistant carry their practice to the people who are still learning the queue. Redesigning the workflow around the assistant starts there. The new sequence has to include the person whose work is worth copying.

A manager who wants this kind of gain has to change the task itself. Decide where the suggestion appears, and whether the agent can ignore it. Put the document one click from the reply. Who is the user when an AI tool never leaves the building? asks whether that employee can dismiss a bad suggestion without leaving the task. The Stanford setup left the agent responsible for the conversation, and free to ignore or edit what the assistant offered. If the agent cannot refuse the suggestion, you should measure your own queue before you borrow this study's numbers.

Collective Campus runs AI design sprints so a team can redraw one workflow and leave with a decision. The value is the redraw of a live task. A licence on its own leaves the old script in place.

What to redraw before you add another tool

1. Put the assistant inside the task the person already has open, at the moment they have to respond.
2. Measure completion per hour, or the equivalent, for newer staff and for your best staff as separate numbers.
3. Name the skilled people whose practice the assistant should spread, and keep them in the work so the system still has something to learn.
4. Give the person on the task a way to ignore or edit the suggestion before it reaches the customer.

If you cannot point to the moment in the live task where the suggestion appears, you have added a tool beside the job. The Stanford result came from a change in how the chat ran. Start with one queue, redraw that moment, and read the two completion numbers before you buy a second tool.