The fastest way to fall behind on AI is to decide you must build it yourself. The frontier keeps moving, and the bill for keeping pace is paid by people who do nothing else. For most enterprises, that is not a fight worth picking.
Buy the capability, build the advantage
Foundation models are infrastructure now, like compute and storage before them. You do not build your own database engine to run a business, and you do not need to train your own model to use one. The honest default is to buy the capability and integrate it well.
Where building earns its keep is narrow and specific: the workflow that encodes how your company actually wins. Your data, your edge cases, your sequence of decisions. That is the layer no vendor can ship, and the only layer where building creates durable advantage rather than maintenance debt.
If a capability is improving faster than you could improve it, buy it. If it captures something only you know, build it. Most of what feels strategic is the former wearing the costume of the latter.
The real cost is not the licence
Teams compare a vendor invoice against the salaries of an internal team and conclude that building is cheaper. They are reading the wrong line. The cost of building is not the first version; it is the second, the retraining, the evaluation harness, the on-call rotation, the slow drift into a system nobody fully understands.
Buying frees that attention for the work that compounds: redesigning the workflow, earning adoption, measuring whether anything improved. I would rather spend a team's best months there than on reinventing a capability the market will commoditize anyway.
A test you can apply on Monday
For each AI ambition, ask one question. Does this differentiate us, or does it merely enable us? Enablers, you buy. Differentiators, you build, and you build only those.
The companies that win with AI are rarely the ones that built the most. They are the ones that knew the difference.