The Challenge
Support volume was growing faster than the team could hire, with messages arriving from email, CRM, and ERP platforms in no particular order, no way to tell an urgent escalation from a routine question. Every reply meant an agent manually checking dispatch timetables, delivery verification, and billing logs before writing back by hand. Reply quality became a matter of luck, whoever picked up the ticket, and how much they remembered to check. And it wasn't confined to one desk, the same problem repeated across every brand and business unit the company ran, each with its own inbox and its own institutional memory.
The Problem
Support volume grows faster than support teams can hire, and it arrives with no built-in triage:
- From everywhere. Email inboxes and CRM/ERP platforms like Salesforce, NetSuite, or SAP, all in no particular order.
- With no way to tell urgency at a glance. An escalation looks identical to a routine question about a return, a delivery, or an invoice, until someone opens it.
Every reply required an agent to read the message, work out what the customer actually needed, and manually open the right system to check:
- Dispatch timetables, to answer "where is my order."
- Delivery verification, to confirm what actually arrived.
- Billing logs, to resolve invoice and payment questions.
Only then could they write a response by hand, every single time.
The cost wasn't just slow response times. It was inconsistent quality: the right answer depended on which agent picked up the ticket, how much they remembered to check, and how many other tickets were already sitting in their queue.
Our Solution
We built CARE, an AI system that sits between the support channels and the company's internal systems.
CARE reads every incoming message, works out what the customer needs, pulls the matching data from the company's own records, orders, delivery status, billing history, and drafts a reply that reads as if a well-informed agent wrote it.
It works the same way across multiple brands and business units at once, without a separate system for each one.
The Result
A draft response is ready before an agent ever opens the ticket, written using the same dispatch timetables, delivery verification, and billing logs an agent would have looked up manually, so reply quality doesn't drop as volume grows.
- Multi-tenant from day one. Runs across multiple brands and business units from a single deployment, with each tenant's data, connectors, and AI provider kept fully isolated, no shared infrastructure, no cross-contamination.
- Not locked into one AI provider. Defaults to a model like GPT-4o today, can pivot to Anthropic Claude or a custom fine-tuned model tomorrow, without a rebuild, which matters because the AI landscape changes every few months.
- Cost-safe by design. Idempotent processing guarantees a message is never billed to the AI provider twice, even after a retry or an infrastructure restart. Smart filtering means only new messages get processed, never a full re-scan of the inbox.
- Gets sharper with use. When an agent edits a draft before sending, that correction feeds back into the system, so accuracy keeps improving the longer CARE runs, not just on day one.
Business Value
Speed Without Headcount
Faster response times without growing the support team at the same rate as message volume.
Consistent Quality
Every reply is grounded in the same account data, regardless of which agent, or whether one is available at all.
Future-Proof Investment
Swap AI providers as the technology evolves, without rebuilding the system around it.
This is the business summary of a platform we call CARE. For the full engineering breakdown, including the architecture, the multi-tenant design, and how we control AI costs at scale, read Building CARE: A Multi-Tenant, AI-Driven Email and CRM Automation Platform →
Jovan Dimitrijević