When broker emails sit unread in a shared mailbox, the clock on your client response KPI is already running. A pilot with GraceAI showed how AI triage can close that gap, processing each email in under 3 seconds from arrival.
We recently ran a pilot with a major insurance services company to see whether AI could improve the way they handled broker emails. The core problem was timing. Broker emails came into a shared mailbox, and by the time staff manually worked through the queue to find the right message, the window to get a response back to the client was already in jeopardy. Consistently meeting their KPIs depended on speed they couldn't guarantee with a manual process. The 4 hours spent on review each week was a symptom. The real issue was that every hour of delay carried a business consequence.
The process was inconsistent too. Different brokers had different naming conventions and communication styles. Reports were produced manually, with no standardisation, and there was no easy way to connect what was happening in the mailbox to what management needed to see.
What We Built
At the core of the pilot was GraceAI, our industry-specific AI model trained to understand insurance language and context. GraceAI runs in a private environment, with no data passing through external public models. That matters in a regulated industry where broker and client data is sensitive.
For each incoming email, the system identified the broker, classified the request, and routed it to the right account manager. End to end, that process took under 3 seconds per email.
We also standardised the reporting output so that management received consistent, audit-ready summaries rather than manually assembled reports.
What the Pilot Showed
During the pilot, the system processed emails automatically, demonstrating that the 4 hours of manual mailbox review each week could be eliminated. It handled communications consistently regardless of how each broker had chosen to write their message.
What stood out was that the constraints we had to work within are exactly what make off-the-shelf AI difficult to apply in this context: varying broker styles, insurance-specific terminology, the need for a clear audit trail. A model trained on the right domain performed better than a generic one would have.
This was a pilot, not a full deployment. But it confirmed that the technology works at the task level. The goal was never to replace the team reviewing that mailbox. It was to take the repetitive, time-consuming part of the job off their plate so they could focus on the work that actually requires judgement. That's the distinction that matters when applying AI in a regulated environment. If you're exploring what this could look like for your organisation, we're happy to talk.