Greenfield #01 - They already know how to count

A vendor at the Grand Marché of Bamako cannot read a single digit. She will still tell you instantly what 3 bags at 750 CFA cost, and exactly what change to give from 5,000. Research has a name for this: street mathematics. Terezinha Nunes and her colleagues studied child vendors in Brazil and found strong mental arithmetic sitting right next to weak symbolic skills.

So when the question comes - could AI teach merchants like her to count? - the premise is already wrong. And the way it is wrong is exactly where the product lives.

She does not need to be taught to count. She needs a bridge from oral numeracy to symbolic numeracy.

The idea

The real gaps are narrow and specific:

• Reading and writing numerals - a price tag, a scale, a phone screen
• Mobile money - confirming you received 15,000 and not 1,500, reading the SMS confirmation (a huge fraud and loss vector)
• Record-keeping - the carnet de crédit where customer debts live, stock, daily profit
• Operations beyond the memorized - percentages for margins and loan interest, comparing price per unit

So the product is an AI tutor that speaks the merchant’s language - literally - and anchors every lesson in transactions she already masters. Not school for adults. Her own market day, played back as curriculum.

And if you want a single wedge to start with, it is mobile money: help her confirm, record and eventually read amounts on her own - no agent, no nephew, no neighbor. Bookkeeping becomes the daily habit. Numeracy is the hidden curriculum.

The part that makes this a venture and not a one-market project: the pedagogy is universal. Counting change, tracking credit, reading a mobile money screen - the lesson is the same in Bamako, Abidjan, Kigali or Nairobi. Only the language layer changes.

So think in trade languages, not countries. Bambara is one face of the Manding continuum that carries commerce across Mali, Côte d’Ivoire, Burkina Faso and Guinea. Hausa crosses Nigeria, Niger and their neighbors. Swahili moves goods for well over 100 million people in East Africa. The map is not clean - markets are multilingual, and a single transaction can mix a local language, French and a counting convention - but the corridors are real. Build the numeracy engine once - every new language is a deployment, not a rebuild.

Why now

The historical blocker was always the same: a learner who cannot read needs voice, in her mother tongue, and languages like Bambara, Soninke, Songhai or Fulfulde had no speech tech. Three things changed:

1. Speech recognition and synthesis for African languages is finally emerging - Lelapa AI’s Vulavula in South Africa, GhanaNLP’s Khaya for Twi, Ewe and Ga, Spitch for Yoruba, Hausa and Igbo, Digital Umuganda for Kinyarwanda, RobotsMali building Bambara speech datasets and models in Bamako, and the Masakhane research collective that laid much of the groundwork. Coverage is uneven - Bambara is still low-resource - but the trajectory is clear.
2. LLMs are infinitely patient tutors. Endless practice problems generated from the learner’s actual trade (“you sold 4 measures of rice at 600…”), difficulty that adapts, and no humiliation. The marginal cost of a personal tutor went from impossible to cents.
3. Multimodal models close the loop. Point the camera at the coins on the table and the AI counts along with you. Photograph a price tag and hear it read aloud in Bambara. Trace a digit with your finger and get corrected.

And one more thing, badly underrated: dignity. Adult illiteracy carries deep shame. A voice tutor on a private phone removes the single biggest barrier to adult education - being seen learning children’s material. No classroom can offer that.

There is a clock on this. Sub-Saharan Africa is the world capital of mobile money - 40% of adults now have an account, up from 27% just three years earlier, the highest rate of any region, moving roughly two-thirds of global mobile money value. Tourism pulls in the same direction: Africa welcomed 74 million international visitors in 2024, above pre-pandemic levels, with governments from Kenya to the African Union investing to grow the number. Every one of these trends converts spoken numbers into screen numbers - QR codes, SMS confirmations, digital price tags.

Meanwhile UNESCO counts 225 million adults in Sub-Saharan Africa who cannot read or write - a number that grew over the past decade even as literacy rates improved, because population grew faster. Worldwide, nearly two-thirds of non-reading adults are women - women like the vendor at the Grand Marché. Adult numeracy is not even properly measured: the global assessments barely reach beyond rich countries.

The development world has a name for what is at stake: “leave no one behind”, the central promise of the UN’s 2030 Agenda - whose education target names adult numeracy explicitly. So do not read 225 million as a market-size slide. It is millions of working adults watching their own economy switch to a language they cannot read.

Product shapes

Three tiers, by device reality:

1. Feature phone, IVR + AI. The merchant calls a free shortcode and a voice in Bambara runs a five-minute daily lesson: “A customer buys 2 sodas at 300 francs. She gives you 1,000. Say the change out loud.” Speech recognition checks the answer. Viamo’s 3-2-1 service already reaches tens of millions across Africa on plain voice calls - AI upgrades IVR from “press 1 to hear that again” to an actual conversation. Widest reach, and telcos can zero-rate it.
2. WhatsApp voice-note bot. Nothing to install, because WhatsApp is already the operating system of African commerce. The merchant trades voice notes with the tutor; photos teach digit recognition (“what does this price tag say?”). Cheapest to build, easiest to pilot.
3. Offline-first smartphone app. The full experience: camera-based money counting, finger-traced digits, lessons that work with no signal. This is the onebillion model - their numeracy app co-won the Global Learning XPRIZE after a field trial with out-of-school children in Tanzania, and their randomized trials in Malawi are the strongest evidence base in this space. On-device models matter here, because connectivity cannot be assumed.

Start with the WhatsApp bot: one market, one language, fifteen merchants. But pick a language that travels - prove it in Bambara at the Grand Marché, and the same build is a deployment away from Bouaké or Bobo-Dioulasso, where the markets speak Dyula, Bambara’s close cousin. The feature phone tier is the scale play. The app comes only once the pedagogy is proven.

Who pays

Do not sell learning. Sell a tool.

Package it as a voice bookkeeping assistant: it tracks customer credit and daily sales (“I sold 3 bags at 500”), and that is worth money on day one. The teaching happens as a side effect - every amount appears on screen while it is spoken aloud, until the merchant starts reading before listening. Adults show up for utility, not curriculum.

And the budget exists. The World Bank’s Findex survey found that one third of mobile money account holders in Sub-Saharan Africa cannot use their account without help from an agent or a family member. Operators have a direct business case - independent customers transact more and are harder to defraud - and their footprints already map onto the trade languages: Orange Money across the Sahel, M-Pesa across the Swahili zone, MTN MoMo in between. That answers the “who pays?” question that NGO-funded edtech never solves.

Honest reckoning

These are not reasons to walk away. They are the questions the right founder will enjoy answering - and the ones I would ask before writing a check.

• The verification paradox is the hard problem at the core: a bookkeeping tool for someone who cannot read the screen must help her catch the day the model hears 750 and writes 75. Reading every entry back is safe but slow, and a pencil tally takes one second. Whoever cracks that confirmation loop owns this product.
• The cheap rival is hardware. India’s Paytm Soundbox - a small speaker that announces each payment out loud - put audio confirmation in front of millions of merchants with zero learning curve, and nothing stops African operators from shipping the same box. But a box that shouts amounts teaches nothing; it makes dependence comfortable. Decide early whether you compete with it or build on top of it.
• Speech recognition for Bambara or Soninke is still early - and a benchmark score means little on a market floor. Open-air markets are loud, dialects drift between Bamako and rural Ségou, and money is spoken in counting conventions the model must convert, not just transcribe. A pilot may need to start in a better-resourced language, or make voice data collection part of the project itself - the Digital Umuganda playbook, where the dataset you gather becomes an asset.
• “Language-agnostic” is where edtech goes to die if it means lowest common denominator. The engine can be shared; the lessons cannot be machine-translated. In Bamako markets, prices in Bambara are quoted in units of five francs - say “eighty” and you may mean four hundred CFA. Every new language needs a linguist and a market walk, not just a new synthetic voice.
• Market women are time-poor in both directions - selling all day, running a household all evening. A lesson that asks for spare minutes loses. The tool has to pay for its seconds inside the transaction itself.
• Edtech pilots die at distribution, not technology. What works is human-anchored: market women’s associations, tontines and microfinance groups as the onboarding channel, one facilitator per group. Pure self-serve apps for first-time smartphone users mostly fail. And do not make a telco deal load-bearing: operator procurement cycles outlast seed runways - treat zero-rating and operator money as upside, not as the plan.
• Many women merchants do not control a device. Shared phones push toward voice calls over WhatsApp.
• Hold the evidence bar high: success is observable behavior - “can read a mobile money confirmation unaided” - not lesson-completion counts.

Monday morning

1. Pick one market, one trade language, one skill. My suggestion: reading mobile money amounts - highest pain, clearest win.
2. Spend a week in the market before building anything. Watch twenty transactions, time how long a debt entry takes on paper - your voice flow has to beat it - and find where counting actually breaks down.
3. Record real transactions on a basic phone in the market noise and run them through the available Bambara speech models. The word error rate you measure there - not the one in the papers - is your roadmap.
4. Build the dumbest possible WhatsApp voice-note pilot with fifteen merchants, recruited through a vendors’ association.
5. Talk to GhanaNLP, Lelapa or Spitch about language coverage, and to a mobile money operator about who funds the pilot.

The merchants of Bamako already know how to count - better under pressure than most of us. The question is whether someone will finally build the thing that meets them where they are.

Greenfield ideas are free to take. No strings attached.

But if you are building seriously in this space, I would love to compare notes - reach out on LinkedIn. I mentor founders who commit, and I invest in the ones who ship.