What AI Costs a Small Business in South Africa

The thirty-second version

A sensible first AI budget for a 50-person firm is a pilot of 10 to 15 seats at roughly $20 to $30 a seat a month. At R18 to the dollar that lands between about R3,500 and R8,500 a month. Not a platform project, and not 50 licences. The people who skew expensive are the ones who buy for everyone on day one.

What the tools list for

List prices at the time of writing (October 2026), in US dollars because that is how you will be billed. Check the vendors' pricing pages before you buy; these move.

Tool class	Typical list price	Notes
General assistant, individual tier (ChatGPT Plus, Claude Pro)	About $20 per user per month	Consumer-grade data terms; fine for evaluation, not client data
General assistant, business tier (ChatGPT Team, Claude for Work)	About $25 to $30 per user per month	No training on your data by default, admin controls, data processing terms
Copilot for Microsoft 365	About $30 per user per month	On top of your existing Microsoft 365 licence
API usage (automation behind the scenes)	Pay per use	Fractions of a rand per task at SME volumes; scales with usage, not headcount

A worked example

50-person professional services firm, first year

12 pilot seats on a business tier at $25 a seat: $300 a month. At R18 to the dollar, that is R5,400 a month, about R65,000 a year.

Break-even: if a loaded admin hour costs the business R250, the stack pays for itself at 22 saved hours a month across the whole firm, which is roughly half an hour per seat per week. For comparison, a single automated process in our results removed 50 to 100 hours a month. The maths is rarely the problem. Adoption is.

Note what is not in that budget: 38 seats for people who have no AI-shaped work yet. Add seats when the pilot group's numbers justify them, not before.

The costs nobody budgets

Overlapping tools. Three subscriptions that all summarise meetings is the most common waste we find in spend reviews.
Unused seats. A seat nobody opens costs the same as one used daily. Review monthly and cancel in the same meeting.
The exchange rate. Dollar pricing means a 10% move in the rand is a 10% price change you never agreed to. Budget with a buffer, and evaluate every tool on its ZAR total cost of ownership.
Rollout without training. A tool nobody adopts is 100% waste. The fifteen-minute rollout session costs less than one month of one unused seat.
Variable API costs. Automations that call AI models are cheap per task but need a monthly cap and an owner, like any utility.

Rules that keep the spend honest

Evaluate on free tiers, but never with client or personal information.
Buy business tiers for the few, expand on evidence. The tier difference is the contract, not the features.
One tool per job. A second tool for the same job needs a reason the first cannot answer.
Name a spend owner and review monthly: what are we paying, who uses it, what did it save?
Put the usage rules in a one-page policy before the seats arrive, not after.

Prices are vendor list prices in US dollars at the time of writing and will change; the exchange rate used in the worked example is stated so you can redo the sums at today's rate. Tracking this spend against results is the "Manage spend" outcome of our AI Adoption program.

Common questions

How much should a small business budget for AI?

Budget per seat, not per headcount. A sensible first budget is 10 to 15 seats at roughly $20 to $30 a seat a month, between about R3,500 and R8,500 a month at R18 to the dollar. Start there, measure what the pilot group saves, and expand on evidence rather than optimism.

Are free AI tools good enough for business use?

Good enough to evaluate whether a tool fits a task, as long as no client or personal information goes into them. For real work the business tiers are worth the money: data processing terms, admin controls, and a commitment not to train on your data.

When does custom AI work beat subscriptions?

When the same task repeats at volume. A person using an assistant pays per seat; a workflow calling an AI model through an API pays fractions of a rand per task and runs without anyone at a keyboard. The crossover is usually a high-volume, rule-based process such as document intake, enquiry handling or reporting, which is process digitisation territory.

What AI costs a 50-person firm, in rands.