AI agents for customer management
A system of AI agents that answers customers across channels, updates the CRM, prepares follow-ups and hands the conversation to a person when needed, with clear rules on what it may do on its own.
Customer care, AI and automation
- Method
- Process mapping before automation
- Control
- Handoff to a human always available
- Quality
- Automated tests on real conversations
The challenge
The team spent most of the day answering the same questions and copying information between email, chat and the CRM. Important requests got lost in the volume. The client wanted to automate without losing their tone or risking wrong answers.
How we worked
- Mapping existing processes with the team: which requests arrive, who handles them, where time is lost. Only what is already clear gets automated.
- Analysis of a sample of past conversations to define categories, edge cases and the brand’s tone of voice.
- A gradual start: first the AI suggests replies to operators, then it answers on its own only for categories with verified results.
- A weekly review with the team of conversations handed to a human, to improve rules and knowledge.
Architecture and choices
- Several specialised agents instead of one assistant: triage, reply, CRM update and follow-up, each with its own tools and permissions.
- The company knowledge base queried with semantic search, with answers grounded in official documents.
- Explicit rules on what the AI may do without confirmation (reply, classify) and what needs a human (refunds, exceptions, complaints).
- A set of test conversations with expected outcomes, run on every change to prevent regressions.
External services, and why these
Messaging
WhatsApp Business API
The channel customers already use, with approved templates for proactive messages.
Chosen over: Unofficial intermediaries
CRM
The client’s existing CRM, via API
No migration: the agents work inside the tools the team already knows.
Chosen over: A new “AI-native” CRM
AI models
Chosen by testing
The model is picked by measuring quality, cost and latency on the client’s own conversations, not by brand reputation.
Chosen over: A fixed vendor chosen upfront
Outcome
Faster answers to recurring requests, a CRM kept up to date without manual work, and a team that focuses on the cases that truly need a person.
Stack
- LLM
- Tool calling
- Python
- Node.js
- PostgreSQL
- pgvector
- WhatsApp Business API
- CRM API
- Evals