Skip to main content
010 447 2445Contact
Arke Technologies home
Request a hardware quoteTalk to us
Arke Technologies home
Two colleagues discuss a trend line on a large wall screen.

Chatbots, virtual agents and AI voice agents

AI voice agents and chatbots answer customers and staff in natural language, by phone, on your website, on WhatsApp or in Teams, and hand over to a person when they should. Arke Technologies, in Johannesburg, designs, builds and runs virtual agents for organisations across South Africa, with logging, approvals and POPIA controls.

Book an AI readiness sessionSee how we build them

What a virtual agent is, and how it differs from an AI agent

A virtual agent is software that holds a conversation in natural language to answer a question or complete a simple task, in text or by voice. The older generation followed scripted decision trees. Today's generative AI chatbot reads your approved documents and systems, understands questions phrased in many ways and answers in your organisation's tone.

TypeHow it worksBest forWatch for
Rules-based chatbotMenus and scripted decision treesA small set of fixed questionsBreaks when people phrase things differently
Generative AI chatbotA language model answering from your approved content (retrieval-augmented generation)Policy, product and service questions in any wordingMust answer only from approved sources, and say when it does not know
AI voice agentSpeech recognition, a language model and speech synthesis on a phone lineCall triage, after-hours calls, status queriesDelay on the line, accents, and telling callers how their information is used
AI agentA model that takes actions in your systems through approved toolsLogging tickets, making bookings, updating recordsNeeds permissions, approvals and a full action log

AI assistants vs AI agents comes down to action. An AI assistant for business answers and drafts; an AI agent also does things, such as opening a ticket or changing a booking, so it needs tighter permissions and approvals. We usually start AI virtual agents as assistants and add actions one at a time. Our guide to AI agents vs generative AI explains the difference, and our AI agents service covers agents that carry out multi-step work under approval.

For organisations comparing virtual agent providers in South Africa, the model matters less than the controls around it: what the agent may see, what it may do, and how quickly a person can take over. Designing those controls is where our AI consulting and implementation practice starts.

Where an AI chatbot for business earns its keep

  • Customer service

    Order and account status, fault logging and product questions, answered at once and handed to a person when the query needs judgement. This is AI customer service automation that shortens queues without closing the door to a human.

  • Internal service desks

    Staff questions about IT, HR policy, leave or procurement rules, answered from your own documents in Microsoft Teams or on the intranet.

  • Municipalities and public bodies

    Residents' questions about accounts, service requests and office hours, with requests logged against the right department and every conversation kept for the record.

  • Schools, colleges and universities

    Questions from parents and students about admissions, fees, term dates and applications, with learner data handled under POPIA.

  • Sales and quoting

    Product and availability questions answered before a salesperson picks up the enquiry. The customer assistant we built for Server Hub, our hardware division, runs in off, draft or automatic mode, with staff approving drafts; the Server Hub commerce and quoting platform case study shows the wider system.

  • Contact centres

    AI voice agents that answer, establish the reason for the call, collect the details and route the caller to the right queue.

A manager presents results on a wall screen to colleagues in an open office.

AI voice agents and voice AI automation

An AI voice agent answers a phone call, understands what the caller says, finds the answer or the right department, and replies in a natural voice. Voice AI automation suits calls that follow a pattern: where is my order, what is my balance, log a fault, book or move an appointment, reach the right person.

We connect voice AI solutions to the telephony you run, or that we deploy for you: VoIP, hosted PBX, SIP trunks and Teams Phone, through our Microsoft Teams, VoIP and telephony service. Before go-live we test with recordings of your real callers, because recognition quality varies by language, accent and line quality, and we keep a route to a person one key-press away.

Voice brings its own rules. POPIA section 18 requires callers to be told that their information is being collected and why, and we add a clear notice that they are speaking to an AI. Outbound calls made by automated means for direct marketing fall under section 69, which allows them only with consent or, for existing customers, under strict conditions.

A finance clerk reviews a printed document beside her laptop.

Channels: website, WhatsApp, Teams and phone

  • Website. A chat window on your site or customer portal, for signed-in or anonymous visitors.
  • WhatsApp. Conversations on the WhatsApp Business Platform, built through our WhatsApp Business API integration service.
  • Microsoft Teams and the intranet. Internal assistants for staff, inside the tools they use every day.
  • Phone. AI voice agents on your existing lines or on a new telephony platform.

One knowledge base and one set of rules serve every channel, so an answer on WhatsApp matches the answer on the phone. That consistency is what turns separate bots into conversational automation you can govern.

Custom chatbot development: how we build and run virtual agents

  1. Discovery

    We study your ticket, email and call logs to find the questions worth automating, and agree which conversations must stay with people.

  2. Knowledge and integrations

    We connect the virtual agent to approved sources, such as your website, policies and SharePoint or Google Drive documents, and, where it needs live data, to your CRM, ERP or ticketing system through APIs.

  3. Conversation design

    Tone, languages, greetings, escalation rules and the data the agent may see, written down and approved.

  4. Model choice

    We build with leading models from Anthropic (Claude) and Google (Gemini, through Vertex AI), or with open-weight models on private AI infrastructure when data must stay in South Africa, choosing the model for each job.

  5. Testing

    A test set of real questions with expected answers, plus deliberate attempts to mislead the agent (prompt injection), run before launch and after every change.

  6. Pilot in draft mode

    The agent drafts, staff approve, and we measure accuracy before it answers on its own.

  7. Run and improve

    Monitoring, logging, a monthly review of conversations and costs, and updates as your services change.

Our engineers work in TypeScript, Node.js and Python, and every build follows the principles set out in our AI governance service. Our AI chatbot development services continue after launch, with the agent run under a service agreement.

Chatbot development platforms and tools: configure, build or host privately

OptionSuitsLimits
The bot built into your helpdesk or CRMSimple questions about that one productKnows only that system's data
Low-code builders, such as Microsoft Copilot Studio (formerly Power Virtual Agents)Internal Teams bots for organisations already on Microsoft 365Licensing and message capacity need checking before you commit
Custom chatbot development on a leading modelSeveral channels, live data from your systems, your controls and logsNeeds design, testing and someone accountable for running it
A private deployment on open-weight modelsSensitive data that must stay in South AfricaHardware and running costs are yours

Acting as your conversational AI consultant, we start by telling you which of these fits. If a tool you already license will do the job, we say so and help you configure it.

Risks, and how we prevent them

  • Wrong answers. The agent answers only from approved sources, shows where an answer came from where the channel allows, and says when it does not know.
  • Prompt injection and data leakage. It sees only what the user is entitled to see, holds no passwords or secrets, and any action beyond answering needs approval.
  • Personal information. A POPIA notice at the start of each conversation, the minimum data collected, retention limits, and processing outside South Africa only where section 72 allows it.
  • Failed handover. A clear route to a person on every channel, with the conversation history passed across so nobody repeats themselves.
  • Runaway cost. Spending limits per conversation and per month, and a cost report in every service review.
  • Brand risk. Tone rules, blocked topics, and draft mode before automatic mode.

Governance, reporting, procurement and continuity

  • Service and reporting. The running agent is covered by a service-level agreement, with a named service lead, an escalation path and a monthly service review showing conversation volumes, handover rates, a sample of answers checked for accuracy, and costs. Changes to prompts, knowledge sources and permissions go through change control.
  • Compliance. POPIA throughout, King IV Principle 12 for boards, and PFMA and MFMA record-keeping for public bodies. Conversation logs are kept for the period your records policy sets.
  • Procurement. Arke Technologies (Pty) Ltd is a Level 1 B-BBEE contributor (sworn affidavit available on request), which gives buyers 135% procurement recognition. Company documents, our CSD supplier number and SARS tax compliance status are available on request. The build is bought against a fixed-scope proposal and the running service under an SLA; we complete standard bidding documents for tenders. See our company facts.
  • Continuity. You own the knowledge base, prompts, configuration and conversation logs, and we hand them over in open formats if you move the service elsewhere.

What a chatbot or AI voice agent costs

Last reviewed: September 2026.

Chatbot development cost has two parts: the build, which we quote as a fixed-scope project after discovery, and the running cost, which you pay monthly. What drives each:

  • Build: the number of channels, the systems the agent must read or update, the languages, whether it speaks or types, and how much conversation design and testing the risk demands.
  • Run: model usage, billed per token by the model provider; telephony minutes for voice; WhatsApp Business Platform messaging charges set by Meta; hosting; and support under the SLA.

We set spending limits from day one and show the cost per conversation in each monthly review. For context, published rates for IT project work in South Africa run from about R550 to R750 an hour ex VAT (matnet.co.za, September 2026).

A contact-centre lead in a headset checks a printed page while a colleague works behind him.

Frequently asked questions

What is a virtual agent? Show answer

A virtual agent is software that holds a natural-language conversation, in text or by voice, to answer questions or complete simple tasks for an organisation. Modern virtual agents use generative AI to understand questions however they are phrased, answer from approved documents and systems, and hand the conversation to a person when a query needs judgement or authority.

What is the difference between an AI assistant and an AI agent? Show answer

An AI assistant answers questions and drafts content; an AI agent can also act, for example by logging a ticket, updating a record or booking an appointment. Because agents change things in your systems, they need tighter permissions, human approval for riskier actions and a full log of what they did. Many projects start with an assistant and add actions gradually.

How much does chatbot development cost in South Africa? Show answer

Chatbot development cost depends on the channels, integrations and languages involved and on whether the agent handles voice, so we quote each build as a fixed-scope project after discovery. Running costs follow usage: model charges, telephony minutes and WhatsApp messaging fees. Published rates for IT project work run from about R550 to R750 an hour ex VAT (matnet.co.za, September 2026).

Can an AI voice agent understand South African accents and languages? Show answer

It can, but recognition quality varies by language, accent and line quality. Before go-live we test with recordings of your own callers, measure accuracy in each language they use, and design the fallback, usually a quick route to a person who speaks the caller's language. We do not switch a language on until the tests show it works.

Will a virtual agent replace our contact-centre staff? Show answer

No, and we do not design it to. A virtual agent takes the repetitive, patterned queries so that your people can handle the ones that need judgement, empathy or authority. Our principle is that AI recommends and people decide: complaints, exceptions and decisions with a real effect on a customer stay with staff, and every conversation can reach a person.

Can an AI chatbot comply with POPIA? Show answer

Yes, if it is built for it. POPIA requires telling people their information is being collected and why, collecting only what the conversation needs, protecting it with security safeguards and meeting section 72 before it leaves South Africa. We build those controls in, keep logs for the period your policy sets, and can host models locally when required.

Can a chatbot work with our existing systems? Show answer

Yes. Through APIs, a chatbot can read and update your CRM, ERP, ticketing, billing or student-administration system, so it answers with live information rather than generic text. We give it only the access each task needs, log every call it makes, and require approval before it changes anything that matters, such as a payment arrangement or an account detail.

Put a virtual agent to work, with people in charge

Tell us which conversations take up most of your team's time, and we will show you which of them a virtual agent can handle safely.

Book an AI readiness session