AI agents for business

How do you build AI agents that work in a real business?

Map how the work really runs first. Build the simplest agent that fits into the systems you already use, then test it on real work. People keep the decisions.

Cape Town, South AfricaHow we build them
Start with the work

Why many agent projects stall

An example prompt can make a useful demo. It doesn't tell you how an agent will handle changing data, missing records or an exception in your process.

When teams start building before they understand the work, they can end up with answers they can't check and connections that need constant attention.

We start by asking what should happen, what can go wrong and who handles it when it does. Rules first. AI only where it helps.

Our approach

How we build them

Critical thinking first

Before writing code, we ask what must be true for the agent to help, what could make it fail and how we will know it works.

First principles

Break the task down. What does the customer need? What can the technology do today? Build from capabilities we can check.

Simplest thing that works

Use plain rules where rules work. Add AI where it helps with the task. Keep each part small enough to understand and maintain.

Tests before building

Agree what a good result looks like. Test output, difficult cases, response time and what happens when the agent needs help.

Each build teaches the next

Use what we learn from testing and running an agent to improve the next build. Keep making small changes based on real work.

Our own tooling

We build and maintain internal tools for testing and deployment. They help us reuse what works and see where something needs attention.

Real work, anonymised

A real example: a fibre infrastructure company

Customers weren't getting outage updates, so they flooded support with the same questions. The fix was connecting outage data to proactive WhatsApp notifications, rather than smarter AI.

A first version in 2 weeks handled about 60% of inquiries on WhatsApp.

By week 6 it handled about 80% of tier-1 support, and passed complex cases to a person with full context.

The tools

What we build with

Multi-channel

WhatsApp, Telegram, Slack, Discord, web and voice. Choose the channels that fit how people already do the work.

Memory and knowledge

Keep conversation context and retrieve information from your knowledge sources so the agent has the context for the task.

Tool integration

Connect CRM, ERP, helpdesk tools and databases through Model Context Protocol (MCP). Give the agent access to the tools its task needs.

Testing

Test individual parts, system connections and agent responses. Check the output and the handover when the agent cannot finish a task.

Deployment

TypeScript and Python, Docker containers and automated deployment. Plan for updates and maintenance as part of the build.

Privacy

Handle personal data with care. People sign off decisions that affect people.

Working with us

Map. Build. Run.

Three stages, each with a clear price and a clear result. You can stop after any of them.

Map

First we ask: is this still the best way to do it? Then we map how the work really runs and put a number on it.

Build

Senior engineers build it into the systems you already use.

Run

We hand it over so your team can run it without us.

Fixed fee, agreed upfront.

This page is the method. It leads to an Automation Audit. Grok Bot setup is the managed assistant you can buy.

See how it works Grok Bot setup
Start here

Give your people their time back.

What you get
Your process map, the cost, a fixed build quote
Price
Fixed fee, agreed upfront

Or send us a short note

Tell us where the repetitive work piles up. We reply by email.