Off-the-shelf AI features versus building your own

Building your own AI model sounds like the serious option, but for most mid-market businesses it is the slow, expensive and unnecessary one. The question worth asking is not whether you could build it. It is whether the off-the-shelf alternative is close enough to already be winning.

There is a build-versus-buy decision hiding inside almost every AI ambition, and most mid-market businesses get it the wrong way round. They imagine they must build something bespoke, when most of what they want now arrives as a feature in tools they already pay for. The skill is telling the two apart: what to buy off the shelf, and the few things truly worth building yourself.

Most of what you want is now a feature, not a project.

A few years ago, putting natural-language questions, forecasting or summarisation into your analytics meant a project. Today much of it is a setting. Power BI Copilot, the Fabric data agent, built-in forecasting in Power BI, summarisation across Microsoft 365, all of it ships inside the platform. For a great deal of what a mid-market business wants from AI, the right move is not to build. It is to switch on what you are already paying for, once the foundations underneath it are sound.

When off-the-shelf is the right answer.

Buy when the problem is common, the solution is generic, and it lives inside a tool you already run. Summarising documents. Asking questions of a report in plain English. A standard forecast on tidy sales data. These are solved problems, and a platform feature will do them as well as anything you could build, cheaper, and with no maintenance burden of your own. Building your own version of a commodity feature is effort spent reinventing something Microsoft will keep improving for free.

When building your own earns its place.

Build when the question is specific to you, the value is real, and no off-the-shelf feature understands your business well enough to answer it. A churn model tuned to your customers and your definition of lost. A market-basket analysis on your own ranges. A demand forecast that knows the particular rhythms of your trade. These are not generic, and a model built on your own history will beat a general tool every time, because the edge is in the specifics. This is where bespoke work pays, and only here.

The hybrid that usually wins.

For most businesses the answer is not all one or the other. It is to buy the platform features for everything common, and build the handful of models that are truly yours. It is the same shape as the wider decision about capability: use what is ready and shared for the general work, invest your own effort only where it gives you something nobody else has. Buy the broad, build the narrow, and let the platform carry the rest.

The trap at both ends.

There are two ways to get this wrong, and they are mirror images. Building what you could have bought, pouring months into a bespoke version of a feature that was a click away, which is expensive and quickly outdated. And buying a feature and expecting it to know your business, switching on a generic assistant and being surprised it cannot answer the specific, data-hungry question that was the whole point. The off-the-shelf feature does the common thing well and the bespoke thing not at all. Knowing which is which is the entire decision.

So this week, write down everything you want AI to do for you, and sort each item into two columns: everyone needs this, and this is specific to us. The first column is almost certainly a feature to switch on. The second is the short list worth building. For most mid-market businesses the split lands around eighty to twenty in favour of buy, which is usually a relief and always a saving. Helping you draw that line, switch on what is ready, and build only the models that pay, is the shape of our Analytics Acceleration Programme and our machine learning work. The same instinct that says build, borrow or both for your team applies to your tools.