Mewokrom Ghana Ltd · Accra
In Ghana, business already runs on assistants.
A shop has an attendant. A home has a help. Mewokrom gives every small business one that works inside WhatsApp, speaks the way they speak, and never forgets what it was told.
“Kwame took five bags yesterday. Has he paid?”
Kwame's order is recorded as GHS 1,250. Nothing has been marked as received.
Shall I remind him?
Mewokrom, from the TwiMe wɔ kromI am in town. I am in the community.
Why
Ama does not need software. She needs a good assistant.
Ama runs a provisions shop in Accra. She is a composite rather than a customer, put together from how these businesses are commonly run.
- Orders arrive by WhatsApp, by phone and in person
- Regulars buy on credit, and the ledger is partly in her head
- Stock is what she remembers plus what she can see
- She moves between Twi and English without thinking about it
A good assistant is a salary, and most small businesses cannot carry one. That is the whole opportunity.
A day with Mewokrom
Morning stock. Afternoon orders. Evening books.
One ordinary trading day, with the assistant present at each point in it. Every exchange is illustrative.
- 08:00
“What am I running low on?”
Cooking oil and Milo are below where you normally are for a Tuesday. Shall I put a reorder together?
- 13:30
A regular sends a voice note in Twi, asking for two crates after four.
Understood, matched to the right customer, written down as an order, and added to this afternoon.
- 17:00
“Who still owes me?”
Kwame, GHS 350, six days over. Abena, GHS 120, due today. Shall I send reminders?
- 21:00
“How did we do today?”
GHS 2,180 across nine orders. Your strongest line was rice. Five orders are still unpaid.
No app to open, no form to fill, no dashboard to learn. WhatsApp is already where the work happens.
And it knows when to ask.
Ama has already told it how far it may go on price. Inside that, it answers and she never hears about it. At the edge of her own rule, the decision goes back to her.
A customer asks
“Ten crates. Can you do 1,800 instead of 2,000?”
What the rule says
Ama allows up to 5 percent off for regulars. This is 10.
So it asks the owner
“Madam, Kojo wants ten crates at 1,800. Your rule lets me offer 1,900. Do you want to take 1,800?”
Language
Understanding the words is not the hard part.
People do not speak to an assistant the way they type into a form. They code-switch mid-sentence, use local shorthand, and leave out anything the listener should already know.
“Ama, tell Kwame say the five bags he take yesterday, make he momo me before evening.”
A model that only transcribes this has done almost nothing. To act on it, the system has to work out:
- Which Kwame, out of everyone who buys here
- Which order, and what is still outstanding on it
- That momo means a mobile money transfer, and to which number
- That sending a payment request needs Ama's say-so first
Chale,Pidginthe black shoe size 42English,ɛyɛ sɛn?Twi
“Friend, the black shoe size 42, how much is it?”
We are not training a foundation model.
We are building the layer that makes existing frontier models reliable for Ghanaian language and real economic work. A different problem, a more tractable one, and we think a more valuable one.
Local language and task data
Consented speech from real business conversation, not scraped text.
Evaluation that measures the job
Right customer, right amount, right action, and knowing when to stop.
Grounding and bounded action
Answers tied to the business's own record, with the owner asked before anything that commits it.
The interesting question is not which model. It is what has to be built around one before it can be trusted with somebody's livelihood, in their own language.
Ambition
One assistant is useful. Many become a network.
When a customer wants something Ama does not have, finding it currently means phoning around or sending someone out. That behaviour already exists. It has simply never been connected.
Agents that can ask each other.
Eventually a Mewokrom agent will be able to put a question to other business agents, source goods, coordinate services, and bring the decisions back to its owner. Not a directory to search. Businesses answering for themselves.
“I need 20 school chairs by Friday, near Madina.”
One message. It goes to the businesses that could actually answer it, and comes back with what is real.
- Furniture workshop, Madina20 in stock, GHS 65 each, delivered Thursday
- Supplier, Adenta20 by Friday, GHS 58 each, collection only
- Carpenter, Legon12 now and 8 by Monday, GHS 60 each
Illustrative. The businesses and the prices are invented.
Where we are building
Ghana
Twi, Ghanaian English and Pidgin. Traders and small businesses in Accra.
Where the same shape exists
West Africa
The same channels and commercial patterns, different languages to learn.
The long ambition
Africa
A language and action layer other people can build on.
The network does not exist yet. It is the reason the assistant is worth building first, and the reason the language work underneath it matters.
Start with the assistant. Build the network.
About
Built in Ghana, for problems we understand.
An early-stage company in Accra, preparing a first pilot with a small number of traders.
- Company
- Mewokrom Ghana Ltd
- Based in
- Accra, Ghana
- Stage
- Prototype, preparing first pilot
Businesses will pay for the assistant. The network it makes possible is where the larger opportunity sits.
Built and running
Listed because it exists, not because it is finished.
WhatsApp channel
Voice notes and text, handled so a repeated delivery cannot record the same thing twice.
Business memory
Prices, stock, customers, orders and balances, kept across conversations.
Permissions and audit
A defined boundary on what the assistant may do alone, with every action logged.
Consent and retention
Recordings kept for a bounded period and then deleted.
AI should meet people where they already are.
Language. Voice. WhatsApp. Real work.