AI in Zambian languages
Bemba, Nyanja, Tonga, Lozi and the country's other languages need their own data, models and tests, so that AI also serves people who don't use English every day.
Applied AI AI product development in Lusaka, Zambia
Zambian organisations that learn to use AI well will get more done with the people they already have.
I bring product management and forward deployed engineering to AI. I start with the workflow, choose where AI can help, then build, integrate and evaluate it with the team. At FigureFlow, this includes document workflows, analysis and decision routing with human review.
My work in practice
At FigureFlow I lead AI-assisted analysis, chat and document workflows inside a financial SaaS platform. Each tool the AI uses has a limited scope, a person reviews its output, and the platform records every action.
I test decision models on thousands of real cases before they run in production.
More about meMy view on AI
AI is moving from answering questions to doing parts of the work itself. Most organisations get little from it because they skip the first step, which is to understand how their own work gets done. I test every AI feature against real cases before anyone relies on it.
How AI works
-
What a language model is
A large language model predicts the next piece of text, one token at a time, using patterns it learnt from a very large amount of writing. That is why it writes fluent explanations and summaries.
It is also why it can sound certain and still be wrong. The model tracks what is likely, and likely is different from true. I design every AI system with checks for that.
-
Choosing a model
The most capable models run on the provider's servers, and you call them through an API. Open-weight models run on your own servers, which matters when data has to stay in Zambia. Small models run on a laptop or phone and cost almost nothing per request.
I compare models on your actual task, then weigh speed, cost per request, privacy and how well each one handles the languages your users speak. The top model on a public leaderboard is often a poor fit for a specific job.
-
Answering from your own documents
A model knows only its training data and what you send with the request. Retrieval-augmented generation, or RAG, searches your documents for the relevant passages and sends them to the model with the question.
When it works, each answer cites its sources and a person can check them. Most of the effort goes into cleaning and splitting the documents, because a model given the wrong page answers from the wrong page.
-
Tools and agents
Tool calling lets a model run a database query, look up an invoice or send a draft for approval. An agent chooses which tool to call next, reads the result and continues until the task is done.
Agents suit open-ended tasks. A process with known steps is cheaper and easier to audit as a fixed workflow with AI at specific steps. I decide which one a problem needs before building either.
-
Workflows as graphs
Most AI systems run several steps, such as extract, check, decide, call a tool, wait for a person and retry. LangGraph and similar frameworks describe those steps as a graph with shared state that every step can read and update.
With the steps written out this way, I can see which step failed, with what input, and run it again.
-
Testing before trusting
Every AI feature I build has a test set drawn from real cases. I measure precision, how often the system declines to answer, the cost of each request and the response time, and I rerun the tests whenever a prompt, model or tool changes.
On one routing system I tested on more than 10,000 cases, 99.5% of the decisions it accepted were correct. It reached that because it could answer "none of these" and pass the request to a person.
-
Permissions, records and people
An AI system should reach only the data and actions its user is allowed to reach. It should record what it saw, what it decided and who approved it.
A person must always be able to take over, so I build the hand-over to a person before I build the automation.
Workflow or agent?
Here is the same supplier invoice handled two ways. Run each one, watch the shared state change and see where a person steps in.
A fixed workflow is predictable, cheap and easy to audit. Use it when you already know the steps.
Shared state
Run log
- Press Run to process invoice INV-3307.
Know when not to answer
A good AI system acts alone only when it is confident, and hands everything else to a person. Move the bar and see the trade-off.
Handled by AI
0%
Sent to a person
0%
Mistakes per 1,000
0
- Handled correctly
- Handled wrongly
- Sent to a person
Simulated requests from a well-calibrated model, for illustration.
Where AI helps
Good AI projects start with a slow or error-prone process and a person who knows it well. Pick an industry to see examples.
Agriculture
- Spotting crop disease from a phone photo
- Farming advice by SMS or voice in local languages
- Forecasting yields and input needs for outgrower schemes
Mining
- Predicting equipment failures before they stop production
- Reading safety reports and flagging risks early
- Summarising technical and compliance documents
Banking and finance
- Reconciling transactions and explaining the exceptions
- Detecting fraud and unusual activity
- Drafting management commentary with sources
Health
- Triage support for community health workers
- Turning handwritten records into structured data
- Answering common patient questions in local languages
Justice and government
- Sorting new court cases and planning hearing schedules
- Helping citizens find services in their own language
- Processing applications faster, with a person approving each one
Education
- Tutors for exam preparation that students can use at any hour
- Marking support that gives teachers time back
- Learning material in Zambian languages
Zambia can lead in using AI on its own problems without building the biggest models.
Zambia now has a national AI strategy. Open-weight models and falling prices mean a country can get value from AI without owning the largest models, and the work that matters is applying them well.
AI infrastructure pays off only when people use it. Across the region, expensive data centre capacity already sits idle, so skills, data and real use cases should come before hardware.
Public services first
Courts, clinics, farms and schools have backlogs that AI can help clear, with a named person responsible for each decision.
Skills before hardware
Engineers who can build, test and maintain AI systems are a better investment than data centres that sit idle.
Protecting citizens' data
Every system should follow data protection law, keep sensitive data secure and explain how it reaches its decisions.
What we should do
- Train people first, then buy hardware.
- Start with one real public problem at a time.
- Build open datasets in Zambian languages.
- Publish measured results for every project.
- Name a person responsible for every decision that AI supports.
Your team has AI tools. Why hasn't the work improved?
- The problem
- A procurement team receives requests by email, checks documents, compares supplier quotes, confirms the budget and sends each request for approval. Everyone has AI tools, and the process is the same as before.
- How I think about it
- I start with how the team works today, including where information comes from, which rules apply and who makes each decision. Then I check whether ordinary automation can solve it without AI.
- What I build
- A workflow that connects requests, quotes and budget records. AI extracts details and flags missing information, agreed rules handle calculations and approval limits, and uncertain cases go to a person.
- How I know it works
- We run it on one process and compare time saved, error rate and review effort with the old way. It expands only after those numbers improve.
Questions people ask
Do you offer AI consulting in Zambia?
Yes. I work with organisations in Lusaka and across Zambia, in person or remotely, and with international teams remotely. We can start with product advisory or an embedded delivery engagement.
Where should a Zambian business start with AI?
Start with one slow, repetitive process that someone understands well. We map how it works today and check whether simpler automation is enough. Then we test AI on that one process and measure the result before expanding.
Will AI replace my staff?
The systems I build take repetitive work off people and give them better information. A person still approves every decision that matters.
Can AI work with our data privately?
Yes. Depending on how sensitive the data is, we can use a provider with strict data terms, run open-weight models on your own servers, or keep some steps away from AI entirely.
Do you build AI agents?
Yes, when the problem needs one. Many processes work better as a fixed workflow with AI at specific steps, which costs less and is easier to audit. I recommend whichever fits the problem.
Bring me a problem like this.
Tell me what you're trying to solve, your timeline and any constraints. I reply within two working days.