Five processes every Kenyan SME can automate with AI this year

Most advice about AI for small business is either a sales pitch or a list of tools. This is neither. These are the five jobs we are actually asked to automate, why they work, and what each one realistically involves.

1. Reading invoices and receipts

What it looks like now: somebody opens an email or a WhatsApp photo, reads the figures and types them into QuickBooks, Sage or a spreadsheet. Maybe fifty a week, three minutes each.

Why AI is good at it: this is pattern extraction from messy input, which is precisely what modern models do well — including crumpled photographs and inconsistent layouts that defeated older OCR.

What you need: a place for documents to arrive, extraction, a check against the purchase order, and a review queue for anything the system is not confident about. That last part is not optional.

Realistic effort: two to four weeks for a working version.

2. Answering the same twenty questions

What it looks like now: the same enquiries arrive every day — opening hours, price lists, delivery times, stock availability — and a person answers each one individually.

Why AI is good at it: drafting a reply from a known set of facts is a language task, and it is fast.

What you need: your actual answers written down somewhere the system can read, a draft generated per enquiry, and a person pressing send. Over time you will find a category you trust enough to send automatically. Start with none.

Realistic effort: one to three weeks.

3. Routing and triage

What it looks like now: a shared inbox or WhatsApp number where everything lands and somebody manually decides who deals with what — so urgent things wait behind routine ones.

Why AI is good at it: classification with a reason attached. It is also the cheapest of these to run, because the output is one label rather than a page of text.

What you need: your categories, your routing rules, and a fallback for anything it is unsure of.

Realistic effort: one to two weeks. Often the highest return of the five.

4. Assembling the weekly report

What it looks like now: somebody exports from two or three systems, pastes into a spreadsheet, formats, and emails it round — on Monday morning, about data that stopped being current on Friday.

Why AI is barely involved: most of this is plumbing, not intelligence. The AI part is at most a written summary on top. We mention it because it is the request we get most often under the label "AI", and because it is usually the cheapest thing on this list to fix.

What you need: read access to the systems that hold the numbers, a schedule, and agreement on what the report should actually say.

Realistic effort: one to two weeks.

5. Keeping two systems in sync

What it looks like now: a customer changes their phone number, and somebody updates it in three places. Or does not.

Why AI is barely involved here either: this is an integration. When there is an API on both sides it is ordinary code, and ordinary code is what you want for anything that must be right every time.

Realistic effort: one to three weeks depending on how cooperative the systems are.

What not to automate

• Anything where a confident wrong answer is expensive. Pricing decisions, credit limits, medical or legal advice. Use rules, not a model.
• Processes that change every month. You will spend more maintaining the automation than doing the work.
• Low volume. Ten minutes a week is not worth a four-week project, however annoying those ten minutes are.

How to pick the first one

Pick high volume, clear rules, and a small blast radius if it gets something wrong. Run it alongside the person for a fortnight and compare the outputs before it takes over. If the first one does not pay for itself inside a quarter, you picked the wrong one — and that is a cheap lesson at this size.

If you want this done for your specific business rather than in general, our free AI audit asks seven questions and tells you which of your processes are worth it and which are not.