Custom AI development, South Africa
Custom AI development, from a team that runs it in production.
IndunAI is a South African AI development company in Durban. We develop conversational AI software that talks to customers on WhatsApp, the web and email, knowledge assistants that answer from a business's own documents, and the integrations that let them book, bill and file work. We run what we build ourselves, on our own servers in Johannesburg, and a project is scoped in a conversation and priced per project.
Want an agent rather than a project? If what you need is one assistant that qualifies, books and hands over on your own rules, start with custom AI agents. If it is the everyday admin after an enquiry, that is AI automation. This page is for work that needs developing.
online, replies in seconds
Do you work on weekends, and is parking included?10:12
We are open Saturdays 08:00 to 13:00. I do not have the parking details in front of me, so I will check with the team and come back to you.10:12
While they look, shall I find you a Saturday time?10:12
Yes, pleaseAn illustration of the pattern. The hours come from the business's own settings, and a fact it has not been given is checked with a person rather than invented.
What we develop
Three kinds of AI software, all of it running today.
People searching for AI developers in South Africa are after everything from a chatbot to a forecasting model. We are specific about our part of that: generative AI software that holds a conversation and then does something, built on an engine we already run for paying businesses.
Conversational AI agents on WhatsApp, the web and email
The core of what we develop is conversational AI on the WhatsApp Business Platform, in a website chat widget, and over email. It keeps one record of each customer, asks the questions a business needs asked, and uses tools to act: check a diary, place a booking, request or send a quote, attach a document, hand the conversation to a person. What an agent should decide and do for your business, and why the server rather than the model makes the call, is set out on custom AI agents, so we will not repeat it here.
Knowledge-base and document assistants (LLM and RAG)
Plenty of "RAG development" is a vector database in front of a pile of PDFs. What we actually ship is simpler and easier to trust at small-business scale. The owner uploads a price list, brochure or FAQ, and one model call drafts the settings from it: services, prices, hours, qualifying questions and question and answer cards. A person reads every row before anything is saved, and the documents are treated as untrusted text, never as instructions. The approved knowledge then goes into the assistant's instructions each turn, and it is told to use those words, not its general knowledge. The same engine can deliver an owner's own document, such as aftercare instructions, at the right moment. If your knowledge base is thousands of pages that genuinely need search and retrieval, say so early; that is a different build and we will tell you whether we are the right people for it.
Integrations into calendars, CRMs and payment
An AI app that cannot reach your systems is a demo. The integrations that exist and run today:
Calendars
Google Calendar, CalDAV and GoHighLevel calendars, read before a time is offered and written when it is booked, with per-staff read-only calendar feeds where a team member keeps their own.
CRM
GoHighLevel: contacts, notes, pipeline stages, appointments and tags, plus an inbound webhook so the CRM's own automations can pause or resume a conversation.
Payment
PayFast payment links and deposits into the business's own account, or banking details with a reference, and invoices computed by the server.
Web forms and email
A website quote form that feeds the assistant directly, forms the assistant hosts and files as PDFs, and the business's own mailbox for sending and receiving.
A CRM or system not on that list is a build question, and the honest answer depends on what its API allows. We check before we quote.
How a build works
Conversation, fixed scope, then it runs.
Enterprise AI development often starts with a discovery phase you pay for. Ours starts with a conversation, and the step-by-step version is on custom AI agents. In short:
A conversation
You tell us the job: what arrives, who handles it now, and what should happen by the end. We tell you what already exists and what would need building.
A fixed scope
What it will do, what it will refuse to do, and the price, in writing before work starts. We commit to a timeline here, not earlier.
Then it runs
On its own instance in Johannesburg, tested by you on a real number before a customer sees it, and looked after by us once it is live.
Production, not a demo
What makes production AI different from a demo.
Anyone can build a chatbot that impresses in a meeting. The difference shows up at 22:00 on a Saturday when a customer asks something odd, the model has a bad day, or the bill starts climbing. These are the parts of our software development that exist for exactly that.
Verification
Model output is a proposal. Slots, prices, ownership of a booking and payments are re-checked by the server before anything happens.
Logging
Every completed model call is one row in a ledger: when, which part of the product spent it, which model, the tokens and the price in rand. Faults land on an errors list somebody reads, not only in a log a restart wipes.
Cost caps
Model keys sit in a proxy the client's instance cannot edit. Each client's daily calls are metered there, with an alert well above a normal day and a hard cap we can switch on.
Tested against the real model
Scripted golden conversations run through the real prompt, tools and model, because unit tests cannot see what the model actually says.
Safety that outranks everything
A customer disclosing self-harm is never silenced, delayed or gated, and always reaches a person. Every image a customer sends is screened before a model sees it.
Isolation and recovery
One isolated instance per client, backups that are restored and verified rather than assumed, and an erasure that removes what it says it removed.
The honest scope
What we do not build.
If you are looking for any of these, you will be better served elsewhere, and we would rather tell you now than after a scoping call.
Computer vision systems
Camera analytics, object detection, inspection lines. Not our work.
Predictive models
Demand forecasting, credit scoring, churn models or anything trained on your historical data.
BI and dashboards over your data
Data warehouses, reporting stacks and analytics on business data outside the conversations we handle.
General "ask it anything" bots on WhatsApp
Why we will not put one on your number is explained on custom AI agents.
Pricing
Built on IndunAI from R1 450 a month. Anything else, per project.
Most custom AI development we do is built on the IndunAI engine, because the expensive parts already exist and run every day. That keeps a build smaller than starting from nothing.
R1 450 / month, including VAT
+ R2 500 once-off setup
- The IndunAI assistant as it is, set up around your business, on its own instance
- Up to 300 conversations a month
- 300 to 500 conversations: R2 250 a month
- Managed, fully looked after: R4 995 a month, setup included, 2,000 conversations, build items within reason; bigger work is scoped in writing and priced once you have signed up
- Software that does something IndunAI does not do at all is priced per project
- Meta's WhatsApp message fees are Meta's, and separate
FAQ
Fair questions.
Do you build AI apps from scratch?
Rarely, and on purpose. Most projects are built on the IndunAI engine, which already handles WhatsApp, web chat, email, calendars, payments and hand-offs in production. Starting from that is faster and cheaper than starting from nothing. Where a project genuinely needs something the engine does not do, it is scoped and priced as its own project.
Can you build on ChatGPT, Claude or open models?
The engine talks to models through the OpenAI-compatible chat-completions protocol, so it works with OpenAI, with a self-hosted open model served by vLLM, and with gateways such as OpenRouter that offer other providers' models through the same protocol. A client can also use their own AI key, in which case that provider's terms apply. Which model fits is part of the scoping conversation.
Who owns the code and data?
Your business data, your customers, your WhatsApp number and your instance are yours, and your data can be exported or erased if you leave. The engine is ours, which is what makes a build affordable in the first place.
How long does a custom AI project take?
It depends on how much of it already exists. A build on the IndunAI engine is usually configuration plus new work on top, which is much shorter than a project from zero. We commit to a timeline in the written scope, not before.
How is a custom build priced?
Built on IndunAI, from R1 450 a month including VAT for up to 300 conversations, plus a once-off R2 500 setup, for the assistant set up around your business. A custom agent runs on the Managed plan, R4 995 a month including VAT with setup included and build items within reason; bigger work is scoped in writing and priced once you have signed up. Software that does something IndunAI does not do at all is priced per project. Meta's WhatsApp message fees are separate.
Tell us about the project.
What arrives, who handles it now, and what has to be true when it is done. Send it on WhatsApp or to angus@indunai.co.za, or talk to the assistant in the corner of this page first.
Where the rest of the detail lives.