How much does AI automation cost in Kenya?

What builds cost

These ranges reflect competent delivery in the Kenyan market: a firm that measures a baseline, integrates with your live systems, handles exceptions properly, documents the result and hands over ownership. Cheaper work exists; it usually skips at least two of those.

Notice how wide each range is. That is not evasion — it reflects the single biggest variable in AI project cost, which is how many other systems the thing has to talk to. A WhatsApp agent that answers questions from a document library is genuinely straightforward. The same agent, if it must check live stock in one system, order status in another, and customer credit terms in a third, is a different project with three times the testing surface.

The running costs nobody quotes

This is where budgets get ambushed. A proposal gives you a build number; twelve months later you discover the total cost of ownership was materially higher. None of these items is unreasonable, but all of them should be on the table before you sign.

Model usage

Charged per unit of text processed. For a typical customer-service exchange the cost is somewhere between a fraction of a shilling and a few shillings, depending on conversation length, how much context is retrieved, and which model tier is used. It matters at volume: ten thousand conversations a month is a real line item, two hundred is not. Any competent firm will model this at your actual numbers rather than waving it away.

WhatsApp conversation fees

Meta charges through the WhatsApp Business Platform per 24-hour conversation window, with different rates by category — a conversation the customer starts is priced differently from one your business initiates. On top of that, the Business Solution Provider through which you access the API charges a platform fee, typically a monthly subscription. Rates and categories change periodically, so treat any figure you are quoted as needing verification against current Meta pricing at the time you sign.

Hosting and infrastructure

For most Kenyan mid-market workloads this is modest — a small server, a managed database, some object storage. It becomes significant only if you are running open-weight models on your own GPUs, which is a deliberate choice usually driven by data residency rather than cost.

Maintenance

The item most often omitted entirely. Model providers deprecate versions with limited notice. Source systems change their APIs. Your product range changes, your prices change, your policies change. Budget roughly 15 to 25 per cent of build cost per year for a system you intend to keep in service. A system nobody maintains does not stay still; it degrades.

Your own team's time

Not an invoice, but a real cost and the one most likely to determine whether the project succeeds. Expect two to four hours a week from the people who understand the process, plus more concentrated time during testing and training. Every stalled project we have been asked to rescue was under-resourced on the client side rather than the vendor side.

What actually drives the price

If you want to reduce a quote, these are the levers that matter, in order of impact.

1. Number of integrations. Each system adds discovery, authentication, error handling, testing and ongoing fragility. Going from one to three systems does not add 200 per cent to cost, but it frequently adds 60 to 100 per cent.
2. Data quality. If records must be cleaned, de-duplicated or reconciled before anything can be built, that is a project in its own right. Sometimes it is the whole project.
3. Exception handling. The happy path is quick. Deciding what happens to the 8 per cent of cases that do not fit — and building the queue, the interface and the audit trail for them — is most of the engineering.
4. Regulatory requirements. Audit trails, approval workflows, retention rules and explainability all add work. Necessary work, but priced work.
5. Number of decision-makers. Not technical, but real. Three approvers across two departments will lengthen every cycle and therefore every estimate.

Conspicuously absent from that list: how advanced the AI is. Model choice is rarely the cost driver people expect. The expensive part is almost always the plumbing around it.

Working out the payback

You can do this yourself in twenty minutes, and you should do it before speaking to anybody.

1. Hours. How many hours a week does the process consume across everyone who touches it? Ask the people doing the work rather than their manager; the honest number is usually higher.
2. Rate. Fully loaded cost per hour — salary plus employer contributions and overheads, divided by genuinely productive hours. Not the headline salary figure.
3. Errors. What do mistakes cost per year in write-offs, rework, penalties and lost customers? Be conservative; you need a number you could defend, not a maximum.
4. Delay. What is the cost of decisions made late or badly because information arrives late? This is often the largest figure and the hardest to pin down. Estimate it anyway, and mark it as an estimate.
5. Compare. Annual cost of the problem, against build cost plus one year of running cost.

A worked example. Two finance staff spend six hours each per week on reconciliation: 624 hours a year. At a fully loaded KES 900 per hour that is KES 561,600. Add a conservative KES 150,000 for errors and rework and the problem costs about KES 711,600 a year. A build at KES 280,000 plus KES 40,000 of first-year running cost totals KES 320,000, giving payback in roughly five and a half months.

Those inputs are invented to demonstrate the method. Use yours. And apply this rule: if the payback is longer than about eighteen months, do something else. Both the technology and your business will have changed before the investment returns, and you will have spent the budget on the wrong thing.

The cheapest useful starting point

If budget is genuinely tight, the highest return per shilling in this market is not a build at all. It is a half-day training session plus a written AI acceptable-use policy, typically KES 80,000 to 150,000.

The reason is simple. Your staff are already using these tools on personal accounts, with company documents, without guidance. One session makes the productive users measurably faster, brings the rest up to a useful baseline, and closes a data-protection exposure that currently sits on your risk register unrecorded. It requires no integration, no infrastructure and no maintenance, and it can happen next week.

The second cheapest useful thing is a scoped audit of one department. It will not change anything on its own, but it converts a vague sense that “we should be doing AI” into a specific list with numbers attached — which is what makes every subsequent spending decision defensible.

Comparing quotes fairly

When two quotes for apparently the same thing differ by a factor of three, the scope differs even if the words do not. Ask every firm the same questions and put the answers side by side:

• Exactly which systems will you integrate with, and have you integrated with them before?
• What happens to cases the system cannot handle? Who sees them, and where?
• What is measured before, and what will be measured after?
• What are the running costs per month at our volumes?
• Who owns the code, prompts and cloud accounts?
• What is included after go-live, and for how long?
• What is explicitly out of scope?

That last question is the most revealing. A firm with a clear out-of-scope list has thought about the project. A firm without one is planning to discover the boundaries at your expense.

Cost traps to avoid

• Per-seat pricing on a platform you cannot leave. Fine while you are small, painful at scale, and the switching cost is the point.
• “Unlimited” plans with soft caps. Read what happens when you exceed fair use.
• Free pilots that create dependency. A pilot built on a proprietary platform is not free if the only way to keep it is to buy the platform.
• Paying for a strategy you cannot execute. A roadmap requiring capability you do not have and cannot hire is an expensive document.
• Buying capacity before proving value. Start with one process. Expand once it works. The order matters more than the total.

If you would like these numbers applied to your actual situation, that is exactly what the free call is for. We will run the arithmetic with your real volumes and tell you honestly if the answer is no.