On this page
From answering questions to getting things done
A chatbot tells you how to do something. An intelligent virtual assistant does it. Ask it to “book a site visit for Mr Shah on Saturday morning” and it checks the calendar, finds a free slot, creates the booking and sends the confirmation. Ask “what did we quote this customer last month?” and it searches the CRM and returns the answer with a link to the record.
Technically, the assistant is a language model given a set of tools — functions it can call in your systems. The model decides which tool to use and with what details; your software checks whether that action is allowed and carries it out. The model never gets direct, unchecked access to your data.
Where assistants earn their keep
We usually recommend starting with internal teams. Staff are forgiving of the occasional misunderstanding, the cost of a mistake is low, and the time saved is immediate. Common starting points:
- Sales teams at a real estate agency asking for lead details, logging call notes by voice or text, and scheduling follow-ups.
- Clinic front desks checking doctor availability, booking and rescheduling appointments.
- Operations managers at a manufacturer asking for today’s pending orders or stock levels without opening the ERP.
- Founders at a SaaS startup getting a summary of yesterday’s signups, churned accounts and open support tickets.
- Restaurant managers checking reservations and daily sales from their phone.
Customer-facing assistants come later, once the actions and limits have been proven internally.
What’s included
- Tool design — the specific actions the assistant can take, each built as a safe, narrow function in your system.
- Permissions — per user or role, deciding what the assistant can read and change on that person’s behalf.
- Confirmation steps for actions that matter, such as sending a message to a customer or cancelling a booking.
- An action log recording every request, every tool call and its result, with a way to reverse changes where possible.
- Channels — a web chat, Slack or Microsoft Teams, Telegram, or WhatsApp, depending on where your team works.
- Summaries and reports generated on request or on a schedule.
How we build it
- List the requests your team makes most often and decide which ones the assistant should handle.
- Design the tools — for each request, the exact function, inputs, checks and output.
- Set the permissions and decide which actions need confirmation.
- Build and test each tool on its own, then test the assistant with realistic phrasing, including vague and tricky requests.
- Pilot with a small group, review the logs together and refine.
- Expand to more users and more actions once the first set is reliable.
Technology
Assistants are built on models with strong tool use (also called function calling) from OpenAI or Anthropic. Your systems are connected through their APIs — for example Google Calendar or Microsoft 365, your CRM, your database or a helpdesk. Where a system has no API, we may build a small service in front of it. The orchestration layer is usually written in Python or Node.js, and the Model Context Protocol (MCP) can be used where it simplifies connecting tools.
What affects timeline and cost
- Number of actions the assistant can perform, and how many systems they touch.
- Quality of your systems’ APIs — clean, documented APIs are much faster to work with.
- Permission complexity — many roles with different access take more design and testing.
- Channel — adding WhatsApp, for example, brings Meta’s platform rules and approval steps.
- Usage volume, which drives ongoing model costs.
Mistakes to avoid
- Broad tools like “run any database query”. Narrow, specific actions are far safer and easier to test.
- Relying on the prompt for security. Permissions must be enforced in code; instructions to the model can be overridden by clever input.
- No confirmation for irreversible actions.
- Launching to customers first, before the assistant has been proven with your own staff.
Frequently asked questions
Can the assistant make mistakes?
Yes, which is why actions are narrow, permissions are enforced in code, important actions need confirmation and everything is logged. Starting with internal users keeps the cost of mistakes low while the assistant is refined.
Which systems can it connect to?
Most systems with an API: calendars, CRMs, helpdesks, databases, accounting and inventory tools. For older systems without an API we assess what is possible case by case.
Can it work on WhatsApp or Telegram?
Yes. Telegram uses the official Bot API. WhatsApp requires the official WhatsApp Business Platform and follows Meta’s messaging rules, which affect what an assistant can send and when.
Does each staff member get different access?
Yes. The assistant acts on behalf of the signed-in user and only has the access that user’s role allows.
How is this different from a chatbot?
A chatbot answers questions from content. An assistant also takes actions in your systems — creating a booking, updating a record, raising a ticket — through specific tools with permissions and logs. Many projects start as a chatbot and add actions later.
Talk to us about intelligent virtual assistants
Assistants that book, look up, summarise and take action — not just reply with canned text.