30 AI use cases that actually work in Kenyan businesses

How to read this list

Each use case carries a difficulty rating based on what it actually takes to get into production in a Kenyan business, not on how impressive the technology sounds.

• Easy — ships in two to four weeks, one or two integrations, low risk if wrong.
• Moderate — four to eight weeks, several integrations or meaningful data cleanup.
• Hard — two months or more, either because the data is difficult or because the consequences of error demand serious controls.

Difficulty is not the same as value. Some of the hardest items on this list are the most valuable, and some of the easiest are the ones we recommend starting with precisely because a fast visible win makes everything after it easier to fund.

Finance and accounts

Where the shortest paybacks live, because the work is high volume, rule-based, and already measured in hours by somebody.

1. M-Pesa and bank reconciliation (Moderate) — match incoming payments to invoices automatically, queue only genuine exceptions. Usually the highest-return automation available to a Kenyan business. We wrote a full playbook on this.
2. Supplier invoice capture (Easy) — read invoices arriving by email, scan or phone photograph into structured data, validate against the purchase order, push into accounting. Handles bad lighting and creased paper better than most people expect.
3. Delivery note and GRN matching (Moderate) — three-way match between order, delivery note and invoice, flagging discrepancies before payment rather than after.
4. Expense claim processing (Easy) — staff photograph a receipt, the system extracts, categorises, checks against policy and routes for approval.
5. Debtor follow-up (Easy) — automated, polite, escalating reminders over WhatsApp and SMS with a payment link attached, stopping instantly when payment arrives.
6. Management reporting (Moderate) — the weekly and monthly pack assembled from source systems, anomalies flagged, delivered before anyone asks for it.
7. Cash flow forecasting (Moderate) — projections from your own receivables, payables and seasonality, updated daily rather than argued about monthly.

Customer service and sales

Where the revenue effect is most visible, because response speed converts directly into orders.

8. WhatsApp customer service agent (Moderate) — answers the common questions instantly, in English or Kiswahili, checking real systems. Complete guide here.
9. Order and delivery status (Easy) — often a third of all inbound messages, and among the simplest things to automate well.
10. Lead qualification (Easy) — ask the qualifying questions, score, book the call, write a clean record into the CRM so salespeople call people who are ready.
11. Quotation drafting (Moderate) — assemble a first-draft quote from a customer request, current pricing and their tier, for a human to check and send.
12. After-hours capture (Easy) — honest handling of messages that arrive at 22:00, with a realistic response commitment rather than a pretence of availability.
13. Complaint triage (Easy) — categorise, prioritise by severity, route to the right person, and never attempt to resolve a complaint autonomously.
14. Review monitoring and response drafting (Easy) — particularly valuable in hospitality, where an unanswered review is a standing advertisement against you.
15. Call summarisation (Moderate) — sales and support calls transcribed and summarised into the CRM, with action items extracted.

Operations and supply chain

16. Demand forecasting (Moderate) — built on your own history with Kenyan seasonality that actually matters: school terms, harvest, Ramadan, December, election periods.
17. Reorder suggestions (Moderate) — combine forecast with lead times and current stock to propose what to buy, for a human to approve.
18. Proof-of-delivery verification (Moderate) — driver photographs checked for legibility, correct document and matching order before the trip closes.
19. Dispatch and route information (Moderate) — driver instructions and customer ETA updates handled over WhatsApp, designed to work on a weak network.
20. Visual quality inspection (Hard) — genuine value in manufacturing, but be warned that a surprising share of quality problems are solved more cheaply by better lighting than by a model.
21. Predictive maintenance (Hard) — requires sensor history and service records that most organisations do not yet have. Worth starting to collect now even if you build later.
22. Tender and RFP analysis (Moderate) — extract requirements, deadlines and compliance criteria from long tender documents, and check your draft response against them.

HR and internal

23. Internal policy assistant (Easy) — staff ask “what is our policy on...” and get an answer with the source document cited. Must enforce access permissions in the retrieval layer, not in the prompt.
24. CV screening support (Moderate) — structure and summarise applications against stated criteria. Shortlisting decisions stay with a human, both for fairness and for legal defensibility.
25. Onboarding assistant (Easy) — answers the hundred questions a new joiner has, without occupying a colleague for two weeks.
26. Meeting notes and actions (Easy) — transcription, summary and extracted action items with owners, distributed automatically.
27. Document search across the organisation (Moderate) — find the contract, the specification, the previous proposal, without knowing which folder it was filed in.

Marketing and content

28. Product description generation (Easy) — consistent, accurate copy across a large catalogue, drafted from your specification data and reviewed by a person.
29. Social and campaign drafting (Easy) — first drafts in your brand voice from a defined brief, edited before publication. Never publish unreviewed.
30. Customer segmentation (Moderate) — group customers by real behaviour rather than assumption, and target accordingly. Requires the data work to have been done first.

Specialist and sector-specific

Beyond the general list, several sector-specific patterns come up repeatedly in Kenya:

• SACCOs — member onboarding and KYC capture, loan file assembly, arrears follow-up, regulatory report preparation.
• Insurance — claims intake from photographs and WhatsApp, policy document Q&A for agents, renewal and lapse prevention.
• Healthcare — appointment handling, insurance claim form preparation, clinical document summarisation for practitioners only.
• Agriculture — farmer support lines with SMS fallback, field form digitisation, aggregation and payment reconciliation.
• Professional services — contract review support, research summarisation with citations, proposal drafting from your own precedents.
• Education — admissions enquiries, fee balance queries, marking assistance, transcript assembly.

We go into the constraints for each of these on our industries page.

What to avoid

Some things are technically achievable and still a bad idea.

• Fully automated credit decisions. Significant legal exposure under the Data Protection Act, 2019, and a reputational problem the first time it is wrong about someone.
• Medical advice to patients. Summarising records for a clinician is useful. Advising a patient is not something to automate.
• Legal opinions without review. Models produce confident, well-formatted, occasionally fictitious citations. A qualified person must check.
• Unreviewed published content. The reputational cost of one obviously wrong published claim exceeds the saving from a hundred drafts.
• Automated pricing without guard rails. A pricing error at scale is expensive within minutes. Bounds and alerts are mandatory.
• Anything pretending to be human. Tell people they are talking to a machine. It is honest, and customers mind far less than executives fear.

Choosing your first one

Run every candidate through four questions:

1. Does it happen at least weekly?
2. Does it follow rules more than judgement?
3. Does the data it needs already exist somewhere, even messily?
4. Is a mistake recoverable rather than catastrophic?

Four yeses is an excellent first project. Three yeses means build it with human review on every case. Fewer than three, choose something else first — not because it cannot be done, but because your first project should be the one that proves the approach rather than the one that tests its limits.

Then commit to shipping it in six weeks. A modest automation running in production does more for your organisation's appetite for this work than any strategy document, because people can see it, use it, and argue about improving it rather than about whether to begin.