Seven Jobs an AI Agent Can Take Off a Small Business This Month

Picture the owner of a bakery in Nakuru at 10 pm. Her phone shows 43 unread WhatsApp messages, half of them "do you have a cake for Saturday". She replies until midnight.

An AI agent is software that reads a message like that, decides what to do from rules you set, and does the small job: checks the menu, quotes the price, takes the order, sends the M-Pesa prompt. It hands over to a person when it is unsure.

It is not magic and it is not a replacement for staff. It is a tireless helper for the jobs a good employee finds boring. This guide lists seven, with what each needs and what to keep human. It sits in our buyer series: the total budget for a build is in how much it costs to build an app in Kenya.

#	Job	Needs	Keep human
1	WhatsApp order taking	A menu or catalogue, prices, M-Pesa or card link	Odd requests, big orders, complaints
2	Answering common questions	A written list of answers: hours, prices, location	Anything not on the list
3	Receipt and invoice capture	Photos or PDFs, a place to store the data	Final tax and accounting decisions
4	Payment follow-ups	A list of who owes what and a polite message	Disputes and hard cases
5	Appointment booking and reminders	A calendar and opening hours	Changes to the schedule
6	Daily sales summary	Sales data from your till or sheet	Decisions based on the summary
7	Supplier reorder drafts	Stock levels and supplier contacts	Approving the order

If you already know which job hurts most and want it built, you can ask Mctaba Labs about an AI agent build.

1. WhatsApp order taking

The agent greets the customer, shows the menu, confirms the order and sends a payment prompt. For example, a customer writes "two chapati and a litre of milk to Kilimani", and the agent replies with the total and a pay link.

Needs: a clean price list, delivery areas and a way to take payment. Risk: a wrong item or price. Always send the order back for the customer to confirm. Keep human: large or unusual orders. Our note on WhatsApp bots for small businesses in Kenya covers the set-up in detail.

2. Answering common questions

Most messages ask the same five things: price, hours, location, availability and payment. The agent answers from a document you write.

Needs: that document, kept up to date. Risk: an invented answer. A good agent is told to say "let me ask the team" when the answer is not in the document. Keep human: complaints, refunds and anything legal.

3. Receipt and invoice capture

You photograph a supplier receipt. The agent reads the supplier, date, items and total, and files them in a sheet or your accounts.

Needs: a consistent place to send photos. Risk: misread digits on a blurry slip. Show the extracted numbers for a quick check before they are saved. Keep human: tax treatment. Our own Company OS product does receipt capture with KRA validation for small teams, so we know the clean-up work involved.

4. Payment follow-ups

The agent sends a polite reminder to customers whose rent, fees or invoice is overdue, and records replies.

Needs: an accurate list of who owes what. Risk: a reminder to someone who already paid. Match payments before sending. Keep human: any dispute, and any message to a customer you want to keep for years.

5. Appointment booking and reminders

For a clinic, salon or garage, the agent offers free slots, books one and sends a reminder the day before.

Needs: a calendar as the single source of truth. Risk: a double booking when staff also write in a paper book. Keep human: changes to opening hours, emergencies and special cases.

6. Daily sales summary

Every evening the owner receives a short WhatsApp: total sales, top items, cash versus M-Pesa, and anything unusual.

Needs: sales data from the till or a sheet. Risk: a summary built on incomplete data. State the data source in the message. Keep human: what to do about it.

7. Supplier reorder drafts

When stock falls below a set level, the agent drafts a message to the supplier. You approve it with one tap.

Needs: stock counts that are mostly right. Risk: ordering too much because of a counting error. Keep human: the approval, always. The agent drafts, you decide.

An agent has a build cost and a running cost. The running cost is mostly messages and the AI model itself.

On WhatsApp, Meta charges for business messages delivered, by category and country. It says the first 1,000 delivered service messages per business phone number each month are free, and that from 1 October 2026 service messages beyond that are charged (Meta WhatsApp pricing documentation, checked on 3 November 2026). Meta also describes a Business Agent service billed by token use. Read the page for the current rates before you plan a budget.

For example, imagine a shop that handles 30 conversations a day with about eight messages each. Count the messages you would send, subtract the free tier, and price the rest from the Meta table. That one sum often decides which jobs are worth automating first.

The same wider cost method, with seven drivers, is in our pillar guide on the cost of building an app. If an agent takes payments, what it costs a business to integrate M-Pesa adds the fee side.

• Wrong answers. Give the agent a written source of truth and tell it to hand over when unsure.
• Customer data. Kenya's Data Protection Act applies to personal data you handle, including phone numbers and order history. Tell customers what you keep and why, and ask the studio where data is stored and who can see it.
• Money mistakes. The agent may take an order but should never approve a refund or a discount on its own.
• Dependence. If the agent stops, can staff take over? Keep a simple manual fallback.
• Tone. A rude or odd reply goes out under your name. Read a sample of conversations every week for the first month.

A good rule: the agent may do anything you would let a new employee do on day one, and nothing you would not.

For more on what a customer chat looks like in practice, see our article on a WhatsApp AI assistant in Kenya.

1. Pick one job. The one you or your staff do most often and find most boring.
2. Write down the rules. If you cannot write them in a page, the agent cannot follow them.
3. Run it beside a person for two weeks. The agent drafts, a person sends. Compare.
4. Measure. Hours saved, mistakes caught, customers who replied faster.
5. Add the second job only when the first is calm.

When picking a studio, use the checklist in 12 questions to ask before you hire a software studio in Africa. For agent work, add one more: how does it behave when it does not know the answer?

If an agent is not what you need yet, custom software or off-the-shelf POS helps you decide between buying and building the underlying system first.

Choose the one job from the seven that costs you the most hours. Write the rules on a page. Then take that page to a scoping call and we will tell you what the agent can do on day one and what needs a person.

To begin, ask Mctaba Labs about an AI agent build.