How to choose an AI development partner in South Africa

First: do you need a partner at all?

Three routes, three different bills. Pick the route before you pick the vendor.

The evidence leans towards getting outside help. MIT's NANDA initiative studied more than 300 enterprise generative AI deployments in 2025 and found that buying from specialised vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded only a third as often (MIT NANDA via Fortune). Hiring your own team is also slow here: demand for AI skills in South African job ads grew 352% between 2019 and 2024 (BusinessTech / Pnet). You are competing with the banks for the same people.

Not sure you are ready for any of the three? Take the AI readiness assessment first. It takes three minutes and saves you a pitch meeting.

The ten criteria that actually matter

These are the rows in the scorecard below. The default weights are our view of a typical mid-sized SA enterprise project. Change them to fit yours.

1. POPIA posture and data residency. Your AI vendor is an operator under POPIA, and section 21 requires a written contract that holds them to your security measures and makes them tell you immediately about a breach (POPIA s21). Personal information sent to a model hosted overseas is a cross-border transfer under section 72 (POPIA s72). A good partner raises both before you do and can tell you which region your data will sit in. Our POPIA and AI guide covers the detail.
2. Delivery evidence. Production systems with real users, not pilots and not demos. Ask to see one running. A slick prototype proves they can build a prototype.
3. Team seniority. Who writes the code after the partner who pitched goes back to pitching? Get names and CVs of the actual delivery team, and a clause that stops them being swapped out without your say.
4. IP ownership and handover. South African copyright law has no automatic "work for hire" rule for contractors, and an assignment of copyright only has effect in writing, signed by the assignor (Copyright Act 98 of 1978, s22(3)). Get the assignment in the contract, plus the repositories, infrastructure accounts and documentation at handover.
5. Pricing model. Fixed price, time and materials, or a monthly retainer are all fine. Vague is not. You want a clear scope, named milestones and a change-request process you understand before you sign.
6. Support and running costs. Hosting, model usage and maintenance are real monthly line items. Our cost guide puts maintenance at 15 to 20% of the build cost a year. A vendor who quotes zero running costs is hiding them.
7. Model and vendor lock-in. Can you switch the underlying model when a cheaper or better one ships? Can another team take over the code? Proprietary platforms you rent forever are a pricing strategy, not an architecture.
8. References you can phone. Two clients, ideally in your industry, who will take a 15-minute call. Logos on a slide are not references.
9. Local context. Load shedding, WhatsApp-first customers, customers who do not all speak English, rand budgets. A partner who has shipped here designs for it without being asked.
10. Change management. The best system your staff refuse to use is worth nothing. Ask how they handle training, rollout and the people whose jobs change.

Red flags: walk out politely

• They promise accuracy numbers before they have seen your data.
• "Operator agreement" gets a blank stare.
• No production work they can show you, only case-study PDFs.
• The pitch team and the delivery team have no overlap, and they will not name the delivery team.
• The contract gives you a licence to use the code, not ownership of it.
• Running costs are "minimal" or not mentioned at all.
• Everything runs on their proprietary platform and there is no exit plan.
• They say yes to every feature in the first meeting. Good partners push back.
• They want the full fee upfront for a project with no milestones.

Questions to ask in the pitch

1. Show us one system you built that is live today. Who uses it, and how many?
2. Who exactly will work on our project, and what percentage of their time is ours?
3. Where will our data be stored and processed, and which models will see it?
4. Will you sign an operator agreement under POPIA section 21? Can we see your template?
5. Who owns the code, the prompts, the trained models and the data pipelines when we are done?
6. What will this cost us a month to run in year one, and what drives that number up?
7. If we want to change model provider in a year, what breaks?
8. What would make you tell us this project is a bad idea?
9. How do you measure whether it worked, and who agrees the measure?
10. Can we speak to two clients, one of whom had a project go wrong?

Question eight is the best one on the list. A partner who has never talked a client out of a project has never been honest with one.

What it should cost

These are the same bands we publish in our AI development cost guide, where a calculator explains what moves you inside them. If a quote lands far outside the band, ask why. Sometimes there is a good answer.

Plus running costs: roughly 15 to 20% of the build cost a year in maintenance, and model usage on top.

The vendor scorecard

Name up to three vendors, set how much each criterion matters to you, then score each vendor from 1 (poor) to 5 (excellent). Totals update as you go. Nothing leaves your browser: no email gate, no tracking of your scores. Download it as a CSV for the procurement file, or print it for the evaluation meeting.

How the maths works: each score is multiplied by its weight, summed, and divided by the best possible total (every criterion scored 5). Unscored criteria count as zero, so finish the card before you trust the number. A weight of 0 removes a criterion.

Where to find vendors to score

Start with our free, editorial directories. Nobody pays to be listed. AI companies in South Africa tracks the full national list, with city cuts for Johannesburg and Cape Town. The directory tells you who exists. The scorecard tells you who fits.

Questions people actually ask