A global insurer's service team opened its day to more than 10,000 emails, claims, renewals, policy questions and complaints, all in one shared inbox. People spent hours just reading and forwarding each one to the right desk, so genuinely urgent messages waited behind routine ones and replies routinely ran past the promised window.
Every email landed in one shared inbox, and a person had to open it, work out what it was about, judge how urgent it was, and forward it to the right team. At more than 10,000 a day, that reading and routing swallowed the hours before anyone had actually helped a customer.
Urgency was invisible until someone opened the message. A claim that needed acting on today sat in the same queue as a routine address change, so the messages that mattered most were the ones most likely to wait, and the promised response window kept slipping.
The cost climbed with the volume. Handling more email meant adding more people to read and sort it, and every attachment, a policy document, a claim form, a statement, was one more thing a person had to open, understand and file by hand.
Let AI understand each message and its attachments, then route it and answer the routine ones, so people spend their time on customers instead of the inbox.
A natural-language model reads each incoming email and its attachments, a claim form, a policy document, a statement, and works out what the message is actually about, rather than matching on a subject line or a keyword.
The system classifies each email by intent, a new claim, a renewal, a policy change, a complaint, and by how urgent it is, so the queue is ordered by what needs attention first instead of by arrival time.
Each message is routed on its own to the team or specialist who should handle it, with the relevant policy and history already attached, so nothing sits in a shared inbox waiting for a person to forward it.
For the routine, well-understood requests, the system drafts a reply that pulls the customer's details and policy status from the CRM, so an agent starts from a written answer to check and send rather than a blank message.
No drafted reply leaves on its own. An agent reviews and sends it, and anything the model is unsure about, an ambiguous request, a sensitive complaint, is escalated to a person rather than answered automatically.
Because the model reads the attachments alongside the email, the information needed to act, the claim details, the policy number, the change requested, is captured up front instead of an agent opening and re-keying every document by hand.
The reading and sorting happens on its own, and people spend their time answering customers.
First response times fell from as long as 48 hours to under 30 minutes. Urgent claims and complaints now surface immediately instead of waiting behind routine mail, and the team's operations lead described the change as getting their people out of the inbox and back onto the work that needs a human.
The company handles its growing email volume with the team it already has. Service costs fell by around 30%, and the desk scales with the business rather than needing a new hire for every jump in volume.
Service slows down when people spend the day reading and routing instead of answering. The value here is an AI that reads every email and its attachments, understands what each one needs, sends it to the right desk and drafts the routine replies, with a person sending every answer, so urgency is acted on in minutes and the team's time goes to customers. It is the same intent-detection discipline behind our ticket triage work, aimed at the insurer's inbox rather than delivery tickets.
A delivery platform's agents juggled four systems to resolve one late order. AI triage with a ready credit and a hard guardrail cut handle time 15%, lifted first-contact resolution 30%, and handled 1.5× the tickets.
Read the case → After-call work · FinanceA financial firm's agents burned a third of the day on after-call summaries and CRM entry. AI auto-summaries approved in a glance cut average handle time up to 49% and lifted capacity 35%.
Read the case →
Reef TRH