What AI automation costs a business in South Africa, and what drives the cost

An AI agent costs what its six lines cost: the build, the platform and model usage, messaging and call fees, integration with your own systems, maintenance, and the people who still own the work. Only model usage, platform fees and Meta's messaging fees are published. The rest depend on your systems and your volumes, so a single rand figure quoted without those inputs is a guess.

Most of the public rates are set in US dollars. A South African business pays them through the exchange rate, so the rand cost moves even when nothing else changes.

A large swing factor is the share of cases that still need a person, because each carries a human cost on top of the machine cost.

• Build: designing the agent, writing its instructions and testing it on real cases before launch. Usually a one-off fee or a block of hours, with no public rate.
• Platform and model usage: what the model provider charges for the text the agent reads and writes, or a monthly fee for each person using a packaged assistant such as Microsoft's Copilot. Public, and in US dollars.
• Messaging and calls: what Meta charges for WhatsApp messages, and what a phone carrier charges for call time. Meta publishes its rates; carrier rates depend on your contract.
• Integration: connecting the agent to your CRM, ticketing, finance system or document store. The cost depends on whether those systems already have usable connections.
• Maintenance: changing instructions as products and policies change, moving to a new model when the old one retires, checking quality, and absorbing rate changes.
• People: the person who owns the agent, the people who review its output, and the people who take over when it hands a case on.

Model providers charge by the token, a fragment of a word. The agent pays for the tokens it reads (the question, its instructions, any documents it looks up) and for the tokens it writes, which cost more on every rate card we checked.

The spread between models is wide. Anthropic lists Claude Haiku 5.5 from USD 0.10 a million input tokens and USD 0.50 a million output tokens, against Claude Opus 5.5 at USD 4 and USD 20. OpenAI lists gpt-6-luna at USD 0.10 and USD 0.50, charged at twice the input rate and one and a half times the output rate when a prompt runs past 272K input tokens, and gpt-6-astra at USD 10 and USD 50, a hundredfold gap within one provider.

Three things follow for a buyer.

1. The model is a cost decision. Routing or sorting a message often runs well on a small model; a long task that needs judgement may need a large one. Ask the builder which model each step uses, and why.
2. Work that does not need an instant answer can run in batches. Anthropic states that Batch API requests are 50% off its standard rates. Overnight document processing is the usual fit.
3. Models retire on dates the provider sets. Anthropic gives the retirement of Claude Sonnet 5.5 as not sooner than 28 September 2027. A move to a newer model is therefore a planned cost, and it belongs in the maintenance budget.

Compare models on cost per completed task: tokens per task, measured on your own cases in a pilot, times the rate.

Microsoft's packaged assistant is charged by the person, whatever the volume of work. Microsoft's South African page lists Microsoft 365 Copilot Business at USD 18.00 a user a month paid yearly, or USD 25.20 billed monthly. The yearly figure is a first-year promotional rate, down from USD 21.00 and offered between 1 July 2026 and 31 December 2026. The page says the figure does not include VAT, and South Africa's standard VAT rate is 15%.

That suits staff who use it through the day in email, documents and meetings. For agents, Microsoft offers Copilot Studio, which it bills in Copilot Credits. Credits replaced messages as the unit on 1 September 2025. Microsoft says the credits a response uses depend on how complex the task is, that purchased capacity is enforced monthly, and that unused credits do not carry over.

Meta's changelog records charging by the message going live on 1 July 2025. Free-form replies inside the 24-hour customer service window that opens when a customer messages or calls you were free from 1 November 2024. Meta's documentation now bills them as service messages. An agent that answers many inbound chats can carry a Meta fee on its replies, and a quote written before the change may leave it out. Ask your provider what Meta charges on your number and which free allowances apply.

Check three things in any WhatsApp quote:

• Who pays Meta. A builder may pass Meta's fees through at cost, add a margin, or fold them into a monthly fee. Ask which.
• The currency. Ask which currency Meta's fees are billed in, and who carries the exchange-rate risk.
• Messages per conversation. A fee for each message rewards an agent that settles a question in fewer, fuller replies.

Meta's calling documentation describes WhatsApp calls as free for your customers and does not say what the business pays, so ask. On ordinary phone lines, call rates come from your carrier contract, so a voice quote should show them separately.

These two lines have no public rate, and they are the easiest to leave out of a quote.

Integration cost depends on your systems: a CRM with a documented connection is quicker to connect than an in-house database with none. Ask the builder to list each system the agent will read from or write to, and to say for each whether the connection exists or must be built.

Maintenance covers the work that starts after launch:

• updating instructions and knowledge when products, policies or rates change
• moving to a newer model when the current one retires
• reviewing a sample of conversations each week for errors
• absorbing a change in a model or Meta rate, and agreeing who carries it

A quote with no maintenance line has either hidden it inside another line or left it out.

Someone owns the agent, someone reviews its output, and someone takes the cases it hands on. Those hours belong in the cost, so use your own payroll. The public floor, for most workers, is the national minimum wage of R30.23 for each ordinary hour worked from 1 March 2026, plus the employer's UIF contribution of 1% of remuneration on earnings up to R17,712 a month, and the Skills Development Levy of 1% of the total paid in salaries once salaries are expected to pass R500,000 over the next 12 months. Leave, equipment, supervision, recruitment and turnover sit on top.

Work it out for one completed task, then for a month.

1. Today's cost of a task: the full monthly cost of the people doing it, divided by the tasks they complete in a month.
2. The agent's running cost of a task: model usage (tokens per task times the rate), plus messaging or call fees for the task, plus the monthly platform fee and maintenance divided by the month's tasks.
3. The handover cost: the share of tasks a person still finishes, times today's cost of a task.
4. Monthly saving: today's cost of all the tasks, minus the agent's running cost and the handover cost.
5. Payback: the build cost divided by the monthly saving.

Run it for a low, a base and a high case, varying the handover share and the tokens per task. If the payback works only in the high case, the case for building is weak. Convert the dollar lines at an exchange rate you would be willing to budget for, rather than today's.

Quotes for the same agent often cover different lines, so their totals cannot be compared until they are split out. Ask every bidder to itemise:

• the build fee, fixed or by the hour, and what testing it includes
• model usage: which models, the measured or assumed tokens per task, and whether usage is passed through at cost or marked up
• Meta and carrier fees: included or extra, and whether they reflect Meta's current charge for replies
• the currency of each line, and who carries the exchange-rate risk
• the handover share the quote assumes
• maintenance: what it covers, and what happens to the monthly fee when a model or Meta rate changes
• what you own on exit: instructions, workflows, knowledge files, conversation logs and data
• what is excluded: extra integrations, channels, languages, or volumes above a stated ceiling
• how automated decisions about people, and transfers of personal information abroad, are handled

A single bundled monthly fee can hide any of these.

Section 71 of the Protection of Personal Information Act restricts decisions about people, such as on credit or staff performance, that rest solely on automated processing, so plan for a human review step; section 72 limits sending personal information abroad. Both add build and running cost, so put them to the builder before the quote is final.

This is not legal advice. Checked against the text of the Protection of Personal Information Act on 9 October 2026, and against SARS's pages on VAT, UIF and the Skills Development Levy the same day. Have an attorney check your own arrangement.

IAMX builds custom AI agents and copilots for a business's own tasks, connected to the tools, workflows and data it already uses, through a vetted network of specialist partners with one accountable lead. Our service page for custom AI agents and copilots sets out what that covers.

• How much does an AI agent cost in rand?

  The published anchors are in US dollars: model usage from USD 0.10 a million input tokens on a small model to USD 10 on a large one, and USD 18.00 a user a month paid yearly for Copilot Business before VAT. Build, integration and maintenance depend on your systems, so no single rand figure holds. Ask for each line separately, with its assumptions, and then convert.

• Is WhatsApp free for replying to customers?

  Not without limit. Replies inside the customer service window were free from 1 November 2024; Meta's documentation now bills them as service messages, so ask your provider what Meta charges on your number and which free allowances apply. Meta describes WhatsApp calls as free for your customers.

• Is Microsoft's Copilot cheaper than a custom agent?

  It depends on who uses it. Copilot Business is a fee for each licensed staff member, USD 18.00 a user a month paid yearly at the current first-year rate, suited to staff in Microsoft's own apps; an agent that answers customers is billed by usage, so its cost follows volume.

• Who owns the agent if we change suppliers?

  Whoever the contract says. Ask for the instructions, workflows, knowledge files and conversation logs to be yours, in a format you can move, and check that the model and messaging accounts are in your company's name.

• Anthropic: Anthropic lists Claude Haiku 5.5 from USD 0.10 a million input tokens and USD 0.50 a million output tokens, Claude Sonnet 5.5 at USD 2 and USD 10, and Claude Opus 5.5 at USD 4 and USD 20. Checked 9 October 2026.
• OpenAI: OpenAI lists gpt-6-luna at USD 0.10 a million input tokens and USD 0.50 a million output tokens, with higher rates for prompts over 272K input tokens. Checked 9 October 2026.
• OpenAI: OpenAI's gpt-6-luna page says prompts with more than 272K input tokens are charged at 2x the input and cache rates and 1.5x the output rate for the full request. Checked 9 October 2026.
• OpenAI: OpenAI lists gpt-6-astra at USD 10 a million input tokens and USD 50 a million output tokens. Checked 9 October 2026.
• Anthropic: Anthropic states that Batch API requests are 50% off the base rate for each million tokens. Checked 9 October 2026.
• Anthropic: Anthropic gives the retirement of Claude Sonnet 5.5 as not sooner than 28 September 2027. Checked 9 October 2026.
• Microsoft: Microsoft's South African page lists Microsoft 365 Copilot Business at USD 18.00 a user a month paid yearly, a first-year promotional rate down from USD 21.00 offered between 1 July 2026 and 31 December 2026, or USD 25.20 a user a month billed monthly, in each case excluding VAT. Checked 9 October 2026.
• South African Revenue Service: SARS gives the standard VAT rate as 15%. Checked 9 October 2026.
• Microsoft: Microsoft states that from 1 September 2025 the unit Copilot Studio agents are billed in changed from messages to Copilot Credits, that purchased capacity is enforced monthly and that unused credits do not carry over. Checked 9 October 2026.
• Meta: Meta's WhatsApp changelog records charging by the message going live on 1 July 2025. Checked 9 October 2026.
• Meta for Developers: Meta's WhatsApp Cloud API documentation says a 24-hour customer service window starts when a user messages or calls the business, free-form service messages can be sent only inside it, and template messages are needed outside it. Checked 9 October 2026.
• Meta: Meta's WhatsApp changelog records service conversations becoming free for all businesses on 1 November 2024. Checked 9 October 2026.
• Meta for Developers: Meta's WhatsApp Cloud API documentation says service messages, the free-form replies sent while a customer service window is open, are billed under the service category. Checked 9 October 2026.
• Meta for Developers: Meta's WhatsApp Business Calling documentation describes calls as free for the business's customers; it does not set out what the business pays. Checked 9 October 2026.
• SAnews, Government Communication and Information System: The national minimum wage rose from R28.79 to R30.23 for each ordinary hour worked, binding from 1 March 2026; Expanded Public Works Programme workers, under a special dispensation, have a lower rate of R16.62. Checked 9 October 2026.
• South African Revenue Service: The employee contributes 1% of remuneration to the UIF and the employer another 1%, a total of 2% that the employer pays over. Checked 9 October 2026.
• South African Revenue Service: UIF contributions are calculated on earnings up to R17,712 a month (with effect from 1 June 2021), so the most deducted from each side is R177.12 a month. Checked 9 October 2026.
• South African Revenue Service: The Skills Development Levy is 1% of the total paid in salaries, and an employer is liable once total salaries are expected to exceed R500,000 over the next 12 months. Checked 9 October 2026.
• Department of Justice and Constitutional Development: Section 71 of the Protection of Personal Information Act bars a decision with legal consequences or substantial effect based solely on automated processing that profiles a person, except where its conditions are met, including measures that give the person a chance to make representations and sufficient information about the underlying logic; the text was read on 9 October 2026. Checked 9 October 2026.
• Department of Justice and Constitutional Development: Section 72 of the Protection of Personal Information Act bars a responsible party in the Republic from transferring personal information to a third party in a foreign country unless one of its conditions, such as adequate protection at the recipient, is met. Checked 9 October 2026.