Should Your Business Build Its Own AI Model? Almost Certainly Not
A confession from the Philsoft office: at least once a month, someone asks us to build them their own AI model. Not use one, not connect one, build one. "We want our own, like the big ones," said a client recently, in the same tone you might use to order a second plate of nyama choma. We love the ambition. But when it comes to custom AI development in Kenya, the honest answer to "should we build our own model?" is almost always no, and the reasons are worth understanding.
What building your own model actually means
The famous AI assistants run on enormous models trained on vast amounts of text, using huge clusters of specialised computer chips running for weeks or months. The organisations that build them employ large teams of specialist researchers and spend sums that would make most Kenyan CFOs need a quiet lie-down.
Building a model of that kind from scratch is not a project for a manufacturer in Athi River, a SACCO in Nyeri or a law firm in Upper Hill. It is not even a project for most large companies anywhere in the world. Many of them, too, are using models built by someone else.
Why it is almost certainly a bad idea
Even smaller custom models, trained for one narrow job, come with real costs that are easy to underestimate:
Data: you need a lot of clean, well-labelled examples, and most businesses have less usable data than they think
Computing power: training needs expensive hardware or cloud capacity, and the bills arrive whether or not the model turns out well
People: you need specialists who are hard to find and harder to keep
Maintenance: models drift as the world changes and need retraining, monitoring and updating
Speed: by the time your model is ready, off-the-shelf models may have improved past it
Kiprop, who runs a distribution business, once asked whether a custom model could predict which shops would reorder next week. Fair question. But a well-built report on his existing sales data answered it faster, cheaper and with no AI at all. Sometimes the smartest AI decision is a spreadsheet.
Smarter alternatives to custom AI development in Kenya
The good news is that you can get most of the value of "your own AI" without building a model. The usual options, roughly from simplest to most involved, are:
Use existing AI tools well, with good instructions and clear processes
Connect an existing model to your own systems through an API, so it works inside your apps
Give the model your own documents to answer from, such as policies, product catalogues or past reports, so its answers reflect your business
In rarer cases, fine-tune an existing model on your examples to adjust its style or behaviour for a specific task
The third option is where many businesses find real value. Your staff can ask questions in plain language and get answers drawn from your own information, without anyone training a model from scratch. Your data stays the special ingredient, and the heavy lifting is done by a model someone else has already spent the fortune on.
When building something custom does make sense
There are genuine exceptions. If you have a very specific, high-volume task, a large amount of unique data that nobody else has, and a clear reason off-the-shelf tools cannot do the job, a smaller custom model may be worth it. Examples might include specialised image recognition for a particular crop disease, or processing a document format no general tool understands.
Even then, the usual path is to start from an existing open model and adapt it, rather than building from nothing. And it should come after cheaper approaches have been tried and measured, not before.
The questions to ask first
Before any AI project, ask yourself: what specific problem are we solving, how will we know if it worked, and what is the simplest thing that could possibly work? If the answer to the last question is not "train our own model", you have saved a lot of money and several sleepless months.
If you have a problem you suspect AI could solve and want an honest view on the simplest way to get there, talk to Philsoft. We will happily tell you when you do not need us to build anything big, and point you to the right software development services when you do.