What is an AI agent, and how is it different from a chatbot?
A chatbot answers. An agent acts. A chatbot takes a question and returns text. An agent takes a goal, such as getting a customer a quote or matching this month's payments to invoices, and carries it through several steps, using the systems where the work already lives: your price list, your diary, your CRM, your accounting system.
Three things make something an agent and not a chat window:
- Tools. It can read from, and write to, the systems the job runs through.
- Judgement inside limits. It chooses the next step itself, within rules you set.
- A way to stop. It knows which decisions are not its to make, and hands those to a person with the history attached.
What do South African businesses use AI agents for?
The jobs that suit an agent share a shape: they repeat every day, they follow rules a good employee could write down, and they cross more than one system.
- Enquiries to quotes. Read the WhatsApp message or the email, ask for what is missing, price it from your rules, send the quote and follow it up. See conversational systems.
- Document intake. Read the application, the invoice or the contract, pull out the fields, check them against your records and flag what does not match.
- Finance operations. Match payments to invoices, chase what is overdue and prepare the pack for a person to approve. See automation.
- Lead research. Read the market every week, score each business on why it would buy, and brief the sales team. See lead intelligence.
- Internal answers. Staff ask a question, and the agent answers from your own documents and names the page. See knowledge systems.
The jobs that do not suit an agent: anything where a mistake cannot be undone, anything with no rule a person could explain, and anything that happens twice a year.
What does building an AI agent involve?
- 01Pick one jobOne job with a number attached: hours a week, days to quote, rand written off. Not "use AI in operations".
- 02Map the workEvery step a person takes today, the systems they touch and the decisions they make. The decisions become the agent's rules.
- 03Connect the toolsThe agent gets access to exactly the systems the job needs, with its own logins and the narrowest permissions that work.
- 04Set the limitsWhat it may do alone, what needs a person's yes, how much it may spend or discount, and when it must stop.
- 05Test on real casesRun it against your real past cases, including the awkward ones, and measure it before a customer ever sees it.
- 06Run itWatch every result after launch, review the failures, and move to a better model when one ships. This step does not end.
The model is the small part. Most of the work, and most of the difference between one company and the next, sits in mapping the work, setting the limits and running it afterwards.
What makes an AI agent safe to put in front of customers?
- Approvals. Anything that reaches a customer, moves money or changes an important record can wait for a person's yes.
- A log of everything. Every message read, every decision and every action is kept, so it can be reviewed.
- Hard limits in code, not in the instructions. A discount cap or a refund limit the agent cannot talk itself past.
- A test set that is run again. The same real cases, repeated every time anything changes.
- Saying "I don't know". When the answer is not in your data, it says so and hands over.
- Personal information handled lawfully. See POPIA-compliant AI in South Africa.
Should you build a custom AI agent or buy a platform?
| Off-the-shelf platform | Custom build | |
|---|---|---|
| Fits when | The job is common and your process is standard | The job crosses your own systems, or your rules are your edge |
| First result | Sooner | Later, after the work is mapped |
| Your systems | The integrations the vendor ships | Whatever you run, including the old ones |
| Who owns it | The vendor. You rent it | You: the code, the accounts and the data |
| Cost shape | Per seat or per conversation, every month | A build price, then a running cost |
| Main risk | Prices and features change under you | Someone has to run it |
A fair test: if a platform does the whole job the way you do it today, buy the platform. If you would have to change how you work to fit it, or it stops at the edge of your own systems, build.
How do you choose an AI agent development company in South Africa?
Ask these, and listen for specifics.
- Who writes the code, and will I talk to them?
- Show me something you built that is running in production today.
- What happens when the model is wrong? Walk me through one failure.
- What will it not do?
- Who owns the code, the accounts and the data if we part ways?
- What happens when a better model ships next quarter?
- How will I see what it did yesterday?
- Where does our data go, and which companies see it?
ZAIQ's own answer to the second question is public: the products we build and run ourselves are on the Work page. For a wider view of the market, see the top AI companies in South Africa and how to choose an AI engineering company.
What does an AI agent cost?
ZAIQ prices on application: one price to build and one to run, both in writing before any work starts. Four things move the number:
- How many systems the agent must work in, and how old they are.
- How many decisions need a person's approval, and how that approval is given.
- The volume: messages, documents or records a month.
- What it costs to run: model usage, hosting and monitoring, which continue after launch.
Be wary of any price quoted before someone has asked how the work is done today.
Short answers
- Which companies build custom AI agents in South Africa?
- ZAIQ builds custom AI agents for established South African businesses and runs them after launch. Other firms that build AI systems are listed, with what each is best for, in our guide to the top AI companies in South Africa.
- What is the difference between an AI agent and a chatbot?
- A chatbot answers questions. An agent is given a goal and tools, carries out the steps itself inside limits you set, and hands over to a person when the rules require it.
- Can an AI agent work on WhatsApp?
- Yes. An agent can answer on your own WhatsApp number, read voice notes, photos and PDFs, quote and book from your rules, and hand serious enquiries to your team with the history attached.
- Which AI model does an agent run on?
- Whichever current model does the job best at a sensible cost. A well-built agent is not tied to one model: it can be moved when a better one ships, after being tested against the same real cases.
- Will an AI agent replace our staff?
- It takes over the repeat work: reading, typing, matching and chasing. People keep the decisions that need judgement, and approve anything that reaches a customer or moves money.
- How long does it take to build an AI agent?
- It depends on how many systems it must work in and how many approvals it needs. A plan with the scope, the price and the order of work comes before any building starts.