Otter's call coaching helps new staff most, and Cadillac's F1 team expects AI to fix records fast

During a Zoom, Teams or Google Meet call, Otter's Live Assist shows small private prompts inside the meeting window, drawn from the company's playbooks and from earlier conversations. Enterprise customers have had it since July, and it reaches every paid plan in the coming weeks. Otter can also write a brief before the call and draft the follow-up for a person to approve.

Live coaching has been measured at scale before. When customer support teams got an AI assistant that suggested replies during chats, the newest agents resolved about a third more issues per hour, and the most experienced barely improved. The suggestions spread what the best people already do, so they teach a beginner and tell an expert what she already knows.

For example, a freight forwarder's new account manager has a renewal call with a customer who is unhappy about late containers. Her brief lists the customer's two complaints and how a colleague settled a similar dispute. When the customer mentions a cheaper rival, a prompt shows the approved reply and how far the price may move. She still decides what to offer.

That changes who should try the tool first. Most companies give new software to their strongest people, who would report, correctly, that it told them little. A trial with the newest hires shows what it is worth, and in a team where people leave often, the newest hires handle a large share of the calls.

The contractor allegedly let a personal Claude Code agent sign in with the contractor's own login, so that it could write code and set up cloud servers. Everything the agent did then looked like the contractor's own work, and it appears to have sent logins and a network map to a high-risk website before anyone noticed. Accenture rejects the allegations.

Most agent software gives an agent an ID that lasts for one run and keeps no record of who approved its work, so a borrowed login is the only name it leaves behind. If an agent merges 400 customer records under a sales analyst's CRM login, nobody can later tell its changes from hers.

FedEx requires every agent that its employees build to run on FedEx's own platform and be registered, with each action traced to a named person.

Some parts that the new team made itself got wrong serial numbers in its ERP system. Fixing one means undoing every delivery, assembly and service booked against that part, in order, then correcting the record and booking everything again. By hand, that took two weeks per part. An AI tool from IFS, the maker of the ERP, can run that sequence, and the team expects two to five minutes per part.

Most companies work around errors like these because fixing them cost too much. At minutes per fix, unclean data stops being a reason to delay an AI project.