Talk to us
Talk to us

Starter guide

AI and automation for a Kenyan business: where to start

Plain automation and AI are different tools, and a job often needs only the first. How to choose one process to start with, and what to ask before you pay anyone.

For any business9 min read

Read the guide Talk to us about automation This guide as Markdown

Rules, AI, or both

Automation and AI are often sold as one thing. They are two, and knowing which a proposal uses tells you what can go wrong with it.

Automation with rules
If this happens, do that. It is predictable and gives the same result every time. Examples: copying a form's details into your accounts, sending a reminder, creating a task when an order arrives. A chain of such steps is called workflow automation.
AI
Reads, writes, sorts or predicts from examples. Examples: reading an invoice and finding the amount, drafting a reply, sorting enquiries by topic. It can be wrong, so it needs a check.
Both together
A useful system is often mostly rules, with AI in one or two steps where a person used to read something and decide.

A useful test

Ask of any proposal: which steps are rules and which use AI? If the answer is vague, the vendor is not yet clear about what it will build.

Where it pays, and where it does not

Look first at work that is repetitive, frequent and easy to check.

Good first candidates
Kind of work Example Why it suits
Sorting what comes in Enquiries from email, a web form or a phone line sent to the right person. Many items, the same decision each time, easy to correct.
Copying details between systems Invoice details read into your accounts, or an order into a delivery schedule. Slow by hand, easy to make slips, and a person can check the result.
Approvals that wait A request that sits in somebody's inbox until it is passed on. The delay, not the decision, is the cost.
Matching two lists M-Pesa payments against invoices, or deliveries against orders. A clear rule, a large number of items, and a short list of exceptions.
Reports made by hand A weekly summary collected from the same few sources. The steps do not change, only the data does.

Poor first candidates

  • Rare tasks, which cost more to automate than to do.
  • Decisions that depend on judgement or on a relationship.
  • Work whose rules nobody can state.
  • Anything where an error is costly and nobody will check the result.

A chatbot on WhatsApp

A WhatsApp chatbot is a program that answers customers' messages on your business number. It can run on rules, on AI or on both, and setting one up is not only a software job. These are the questions that decide whether it works and what it costs.

  • Which route: the WhatsApp Business app is made for a person answering by hand. Software that answers by itself connects through the WhatsApp Business Platform, often called the WhatsApp Business API, either directly or through a provider. Ask which is proposed, and who the provider is.
  • Who owns it: the business number, the account and the customers' messages should be in your company's name, not the provider's.
  • Permission: how does a customer agree to receive messages from you, and how is that recorded? Ask what the platform's rules let you send first, and what Meta, the owner of WhatsApp, has to approve before you can send it.
  • What it costs: the provider's fee and the platform's own charges for messages are separate, and both can change. Ask for both in writing, and how each is worked out.
  • A person to take over: can a member of staff see the conversation and answer when the bot cannot, and how does the customer reach them?
  • Languages: does it understand the languages your customers write in, such as English and Swahili? Test it with real messages from your own inbox.
  • Personal data: what is stored, where, for how long, and who at the provider can read it. The Data Protection Act, 2019 applies to what you keep about customers.
  • Orders and payments: if the bot takes an order or a payment, which system records it, and how is an M-Pesa payment matched to the order? The guide to M-Pesa integration goes through the ways to collect.

Choosing the first process

One process, done well, teaches more than five started together.

  1. List the candidates

    Write down the tasks people repeat, how often, and who does them. Ask the people doing the work, not only their managers.

  2. Choose a frequent, rule-based, low-risk one

    If it goes wrong, the harm should be small and easy to spot.

  3. Write down how it is done today

    Every step, including the exceptions that people handle without thinking. The cost often hides in the exceptions.

  4. Decide what a person approves

    What the system may do alone, and what it must show a person first. Start with more checks and remove them as trust grows.

  5. Decide how you will know it worked

    A number you can read before and after, such as time to answer, errors found or items waiting. Measure the old way first.

What can go wrong

None of these is a reason not to start. All of them are reasons to plan.

What can go wrong
Risk What it looks like What to do
A wrong answer An AI step states something with confidence that is not true. Keep a person's check on anything that touches money, customers or records. Keep a log of what it did.
Data going somewhere it should not Customer details pasted into a public tool, or sent to a provider whose terms you have not read. Know where the text is processed and stored, and whether it is used to train a service. Keep personal data out of public tools.
Cost that grows A tool that charges by use, and use that grows faster than planned. Agree a monthly limit and an alert before you reach it.
Depending on one provider A service changes its price, its terms or goes away. Keep the rules in your own system and make the AI step replaceable.
Staff who do not trust it People work around it, or double-check every item. Involve the people doing the work from the start, and show them what it does and what it does not.

Customer data is your responsibility

Under the Data Protection Act, 2019, a business is responsible for the personal data it holds, including what it sends to a tool or a provider. Check where it goes before it goes.

Questions to ask a vendor

These work for any proposal, from any company.

  • Which steps use plain rules and which use AI?
  • What happens when it is wrong, and how will I find out?
  • Where does my data go, and who can see it?
  • What will it cost each month at my volume, and if the volume doubles?
  • Can I see a log of every decision it made?
  • What happens if the AI provider changes or stops its service?
  • Who looks after it, and who fixes it when a connected system changes?
  • Can I switch it off and go back to doing the work by hand?

Start with a pilot

Run the first process on real work for a defined period, beside the old way. Compare the numbers you chose, listen to the people who use it, fix what they found, then decide whether to widen it. A pilot that shows nothing is still worth having, because it has cost little and told you something true.

Keep reading

Common questions

How can AI help a small business in Kenya?

It can read and sort what comes in, such as enquiries or invoices, draft replies for a person to check, and spot patterns in your own records. It helps most where the work is repetitive, frequent and easy to check.

What is the difference between automation and AI?

Automation with rules does the same thing every time something happens. AI reads, writes, sorts or predicts from examples and can be wrong, so it needs a check. A useful system is often mostly rules with AI in a few steps.

Which business processes should I automate first?

Start with work that is frequent, follows rules and does little harm if it goes wrong: sorting incoming enquiries, copying details between systems, approvals that wait and matching payments to invoices.

Is it safe to put customer data into AI tools?

It depends on the provider and the data. Check where the text is processed and stored, and whether it is used to train a service. A business is responsible for personal data under the Data Protection Act, 2019, so keep customer records out of public tools.

How much does business automation cost?

It depends on how many steps and systems are involved, how much work goes through it and whether an AI service charges by use. Ask for the build cost, the monthly running cost at your volume, and what happens to the cost if the volume doubles.

Can a chatbot answer my customers on WhatsApp?

Yes. A chatbot can answer common questions on your WhatsApp business number and hand over to a person when it cannot. Software that answers by itself connects through the WhatsApp Business Platform, so ask who the provider is, who owns the number and the messages, and what each message costs.

How do I know if automation is working?

Measure the old way first: time taken, errors found, or items waiting. Then read the same number after a pilot beside the old process, and ask the people who use it.

Have a process in mind?

Tell us how it works today and where it gets stuck. We will say what could be automated, what needs a person, and whether AI has a part in it.

Talk to us about automation