Kenyan startup Fastagger builds offline AI for small businesses on basic phones
Fastagger's assistant Auni reads mobile-money records and turns them into sales insights for salons and restaurants. About 6,000 businesses have signed up since it was added to M-Pesa's business app.
Fastagger, a three-person startup in Kenya, builds AI models that run offline on low-power phones. Its main product, Auni, is a business intelligence assistant. It reads a small firm's mobile-money transaction records and turns them into charts and marketing suggestions. Auni launched in 2024. Its reach grew when it was built into the business app of M-Pesa, Kenya's largest mobile money service. IEEE Spectrum reports that about 6,000 businesses have signed up since then.
The company is small, but its approach shows something about how AI reaches people. Most AI products are designed for fast connections, capable devices and clean digital records. Fastagger started from the opposite case: cheap phones, unreliable internet, and businesses whose only records may be their mobile-money transactions.
Starting from the constraints
Cofounder Mutembei Kariuki says he first thought about AI in Kenya in 2019, at a conference on the future of AI at the University of Tokyo. "But nobody mentioned Africa at all," he told IEEE Spectrum. Even then he saw the obstacles. Internet access was slow and unequal. There was little compute for training models and little structured, high-quality training data. Most people went online with older, low-memory phones, not computers.
Kariuki founded Fastagger with Jude Mwenda and Stephanie Njerenga. They chose to focus on mobile money apps because they are so widely used: more than 80 percent of Kenyan adults use them, according to the article. That made transaction records the most practical data source available.
The first technical step was document reading. The team adapted an open-source optical character recognition model to read financial statements exported from mobile money apps. Existing models struggled with these PDFs. They were low quality, oddly formatted, and sometimes had stamps and errors on them. Because many prospective users had poor connectivity, the team made sure key steps, such as arranging the extracted data into tables, could run on the phone itself.
What runs on the phone and what does not
Some of Auni's functions work offline on low-cost phones without querying the cloud: reading sales statements, analysing the data and visualising the results. These combine conventional software and AI. Standard OCR software extracts the text, and a small neural network works out how each document is structured. More detailed analysis and answering questions about the data still need an internet connection.
Before building the product, the founders spent whole days watching small-business owners at work. One constraint stood out: owners had almost no time to think beyond daily operations. "They are standing up the whole day cutting hair," Kariuki says of salon owners, who are now among Auni's main users.
Paper ledgers to charts
Peter Chege runs a salon and barbershop in Nairobi. He started using Auni last year. Before that he recorded all his sales on paper. "We were not able to translate that information into a graph," he says. The app now shows him his busiest days and hours and his loyal customers. When a regular stops coming, he can send them a personalised message.
A simple chart showed him that mornings were his slowest period. He now runs errands in the morning or keeps fewer barbers in the shop until around 1 p.m., when business picks up.
A fish restaurant in Nairobi offers another example. To find loyal customers and give them discounts, staff used to go through SMS sales receipts one by one. That made retention marketing impractical. With Auni, the restaurant identifies its top customers automatically.
Fastagger has also built a workforce management tool for Kenya's security companies, which together employ hundreds of thousands of people. Guards often work in remote places on low-end phones with poor internet. The company's AI agent talks to them over SMS, tracks their routes, detects anomalies and sends alerts to operations teams.
The startup has had support from accelerator programmes run by Google, Microsoft and NVIDIA. It still has only three employees, the founders. They are now working on agents for Auni that would carry out marketing actions for business owners automatically.
A risk of two tiers
Daniel Björkegren, an economist at Columbia University who studies AI in developing economies, told IEEE Spectrum that the region's reliance on basic phones points to different kinds of use. "The history of technology should suggest that this is not going to be people running Claude Code or Codex on their laptop," he says. "You will have people like that, but for much of the population, we're going to have to figure out different usage modalities."
He also warns of a split. He sees value in systems built for people Silicon Valley has served poorly, but is wary of a clear division between small models for emerging economies and ever larger models for wealthier regions. "I am worried that that could lead to a bifurcation," he says. "We need to enable a large menu of different options and really make sure that we have access and then see what works."
The article lists similar projects elsewhere: an app for rural schools without internet in Chile, a medical triage chatbot for rural patients in India, and translation models for African languages that run on basic phones.
What it means for work
For small firms like salons, restaurants and small manufacturers, the tasks automated here are bookkeeping analysis and customer retention: turning a stream of payment records into busy hours, top customers and lapsed regulars. These tasks were possible before but too slow to do by hand, so they mostly went undone. The gain is less about replacing staff and more about giving owners information they never had, such as when to schedule fewer workers. For security guards, an SMS agent now handles part of the route tracking and reporting that supervisors did. What remains open is how far the offline approach can go, since deeper analysis still needs a connection, and whether small, constrained models end up as a permanent second tier next to the large models used in richer markets.
Sources
This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.
More from māchine
All stories
OpenAI says the European landlord cut the cost of a staged listing photo by more than 99% and shrank a climate-risk project from five people over a year to two people over two months.
A revised IPO filing says Anthropic's internal estimates point to AI soon dramatically accelerating its own research, a point it links to recursive self-improvement. No current model has crossed it, the company says.
The suits say Brave copied paywalled journalism, sold access to it through its search index to AI companies, and let browser users skip subscriptions. Brave has not commented yet.