The last mile of AI work is local
AI is automating desk work, and paying people for the judgement models lack. Physical AI needs that judgement about real places. Here is what fair local AI work should look like.
Short answer: AI companies already pay people for the judgement models lack. As AI moves into the physical world, a new kind of that work appears: checking real places. It has to be done by people who live near them. Done well, it is clear, opt-in, reviewed work. Done badly, it repeats the worst of online data work.
AI already runs on human judgement
The biggest AI labs pay for human data at scale. In June 2025 Meta put $14.3B into Scale AI, a company built on people labelling and grading data. Newer platforms pay experts to grade model answers. Much of the work is remote and on a screen: reading, ranking, correcting.
That is the first layer of AI work. It is mostly desk work, and much of it will be automated by the models it trains.
Physical AI adds a layer that can’t be remote
A robot, a delivery fleet or a model of a township street needs to know what is really there. Is the shop open? Is the product on the shelf? What does it cost today? Did the billboard go up? No model can answer that from the web, because the answer is not on the web. Someone has to stand in the place and look.
That makes local knowledge a new kind of skill in the AI economy: knowing your streets, your shops and your area well enough to answer a specified question quickly, with proof.
The last mile of knowledge work is local. It is done by people who are already there.
Our thesis
Why it matters in South Africa
Unemployment among South Africans aged 15 to 34 was 47.4% in the second quarter of 2026 (Stats SA). Many young people know their area well, and local knowledge is exactly what this work needs. We do not promise jobs or income; we design each task to be clear, opt-in and reviewed.
What fair local AI work looks like
Online data work has a mixed record. In 2025 data labellers in Kenya formed their own association after reports of opaque, low pay and long hours. Local AI work should learn from that from the first task. At Timbuk2 we design it around four principles:
- Opt-in tasks with a clear, specified question. You know what is being asked before you start.
- Pseudonymous by default; consent for each use. Contributors are not identified to buyers.
- Answers reviewed and accepted against a stated rule. AI handles routine checks; people review anything unclear.
- Nothing licensed for model training without agreed consent.
Today, people take part through 2Klip, our local-knowledge game, where answers earn BUK2 rewards, and through FOONA, our field network for requested checks with a photo, location and time.
Where this goes
We are building the network AI can ask when the answer is not online. Every answer in it is a small piece of local AI work. For the investment case, read our investor thesis; for the market, read why physical AI in Africa needs ground truth first.
