Marshmallow built an AI customer service agent, Marsha, and put it to work on its hardest job first: keeping customers at renewal. Vik Lakhotia, Group Product Manager at the UK car insurer, explained how at aiblLIVE London 2026.
Built with Sierra and Marshmallow’s own engineers, Marsha now talks to almost everyone who starts a chat. It is judged on retention and profit per chat, not automation alone, and it now matches the median human agent.
All figures as stated on stage by Vik Lakhotia.
Because its summer renewal peak outran its human team. Marshmallow sells half its policies over summer, so renewals bunch between July and September. Customers sometimes waited three or four days for a reply, long enough to buy elsewhere on a price comparison site. In winter, agents had too little work.
Give the vendor your hardest problem, and put the agent where the business feels it. Marshmallow tested partners on retention negotiation, which needs its internal systems. Marsha went live just before summer and was business critical within a month, so leaders and compliance stayed invested well beyond the launch buzz.
We wanted to make sure that we’re not eroding business value for the sake of a vanity metric of automation.
Vik Lakhotia, Group Product Manager, Marshmallow
Measure what the business earns. A human chat cost Marshmallow around £10, while a retained customer was worth an order of magnitude more. So it tracks retention success rate and profit per chat. That revealed Marsha was over-discounting by 10 to 20% at first, which the team traced by reading all of those chats before improving it.
A second AI agent checks that what the customer asked, what Marsha said and what it did all match. Human agents labelled 300 to 400 chats to train that judge, and tuning lifted its accuracy from about 83% to about 98%. Remaining mismatches trigger a Slack alert for a human to fix.
Vik Lakhotia is Group Product Manager at Marshmallow, the UK car insurer, where his team builds its AI agent, Marsha. He spoke on the Customer track at aiblLIVE London on 20 October 2026.
If you are weighing an AI agent for a high-stakes customer workflow, this is the playbook to borrow. The full PDF goes deeper than this page, in Lakhotia’s own words.
For customer service, operations and product leaders, and anyone running AI in a regulated business. Six pages, about a nine-minute read.
