Agents and assistants

The problem

A generic chatbot will confidently tell your customers things that aren’t true: it makes up a price, promises a delivery date or hands out advice nobody asked for. An agent with broad access to your systems is a liability, not a helper. Meanwhile the day’s loose ends —the reminder, the call back, the heads-up to a colleague— still depend on someone remembering.

How we work

Every assistant we build is designed around a single source of truth. In the sales proposals we publish, the assistant knows the real prices and terms and is not allowed to stray from them; asked something outside its data, it says so and offers to pin it down with a person. The same rule applies on a storefront, in a client portal or on a support channel.

Agents that act on your systems do so through MCP servers with a deliberately narrow set of tools: they can search the catalogue or prepare a draft order, but they can’t run whatever they like. Anything that affects a customer comes out as a draft for a person to confirm, and every call is logged.

For schedules and to-dos we deploy digital personal assistants: agents with a defined remit —work, home, one specific person— that keep reminders and calendars, reply in their own Telegram topic and alert whoever needs to know.

What’s included

• Defining each assistant’s source of truth and ground rules.
• Connection to your channels: web, WhatsApp, Telegram, Instagram or email.
• Tool servers with least-privilege permissions and usage logs.
• Optional voice: transcription and spoken replies.
• Testing against real questions before launch, and regular reviews of answers.