Build, buy or partner: the enterprise AI decision

The memo almost always arrives framed as build versus buy. Partnering — where an external team deploys and then hands over — is the option that gets mentioned in a sentence and modelled in none of the columns, despite being where a large share of enterprises actually land.

The four variables that decide it

• Proximity to core differentiation. If the workflow is how you win, building keeps the advantage inside. If it is overhead, building is an expensive way to own a commodity.
• Data sensitivity. Not whether the data is sensitive — almost all of it is — but whether the sensitivity survives a well-designed deployment boundary. Frequently it does.
• Internal capability today, not the capability on the hiring plan. A team you intend to build in nine months cannot carry a decision you are making this quarter.
• Time to first value. The variable executives weight highest and models weight lowest.

The honest cost picture

Building looks cheapest in the first year and is almost never cheapest by year three, because the model that nobody budgets for is maintenance: evaluation harnesses, drift monitoring, model upgrades, and the fact that the two engineers who built it will move on. Buying is predictable and caps your ceiling at whatever the vendor's roadmap allows. Partnering is the most expensive per week and the cheapest per unit of learning, because the capability transfers if — and only if — handover is contracted rather than assumed.

The hybrid most enterprises land on

Partner for the first workflow to compress time to value and to import the patterns. Buy for the commodity layers where a vendor's roadmap is genuinely ahead of anything you would fund. Build only where the workflow is proprietary and you have already proven the capability once with someone else's hands.

How to test the partner route cheaply

Before committing to any of the three, watch a partner run the capability on a problem shaped like yours, with your operators asking the questions. It costs a session and it resolves most of the build-versus-buy argument, because the argument is usually a proxy for uncertainty about whether the thing works at all.