When Bolted-On AI Is Actually the Right Call

I argue hard for AI-native, so let me argue against myself. Bolted-on AI is sometimes the right call, and pretending otherwise wastes money and time. Not every product should be rebuilt around a model. When AI genuinely sits on the edge of your value instead of at the center, a bolt-on is not a compromise, it is the correct engineering decision. The zealotry that says everything must be native is as wrong as the laziness that bolts a chatbot onto everything. Here is when bolting on is actually smart.

When AI is a convenience, not the value

The clearest case: the AI is a nicety on top of value that stands on its own. A project management tool where the core value is coordination and the AI writes summaries. A design tool where the value is the canvas and the AI suggests colors. Strip the AI and the product is still a product people pay for. That is bolt-on territory, and rebuilding around the model would be throwing away a working foundation for no gain.

The mistake native purists make is treating every AI feature as a reason to rebuild. If the model is a convenience, bolt it on, ship it, and put your rebuild budget where it actually matters. The question is always whether AI is the value or a garnish on it, which is the same test I use in rebuild vs retrofit.

When you are still learning the market

Early on, you do not know if the AI feature matters to customers. Bolting on is the cheap way to find out. Ship a thin version, watch whether people use it, learn what they actually want. Rebuilding native before you have that signal is betting a year on a guess. Bolt on to learn, then rebuild if the signal says the model should be core.

This is the honest path from wrapper to product. The failure is not shipping a wrapper, it is treating the wrapper as finished. Ship thin to learn, then build the spine once you know it is worth building. I have done this repeatedly across the portfolio: prove the demand cheaply, then invest in native only where the demand is real.

When the deterministic core must stay deterministic

Sometimes the core of your product must not be probabilistic, ever. Financial calculations, legal filings, anything where the answer has to be exactly right every time. You do not want a model in that path. You want the model bolted on around it, assisting, drafting, and flagging, while the deterministic core stays deterministic.

This is not bolted-on as a failure, it is bolted-on as a design choice. The model helps at the edges and the trustworthy core stays untouched. Even in my most AI-native products, the commit step is deterministic on purpose, which is the whole point of why reliable agents beat capable ones. Knowing which parts to keep away from the model is as important as knowing where to put it.

When native would out-run your team

There is a resourcing reality too. A native rebuild is a serious commitment. If your team cannot execute it well, a badly built native product loses to a cleanly built bolt-on. Ambition you cannot deliver is worse than a modest thing done right. Be honest about what you can build and maintain before you commit to the harder path.

I run the twenty companies behind Girard Media solo, so I feel this constantly. Native is the right long-term call for the products where the model is core, but I still bolt on where the model is peripheral, because my time is finite and I spend the hard rebuild budget where it compounds. Choosing where not to go native is how the native bets stay fundable.

The rule that keeps you honest

Bolt on when AI is a convenience, when you are still learning, when the core must stay deterministic, or when native would exceed what you can execute. Rebuild native when the model is the value and you have the signal and the capacity to do it right. The failure is not choosing bolt-on. The failure is choosing bolt-on for a product where the AI is the whole point, because you wanted to avoid the hard build.

Most products that should be native get bolted on out of convenience, which is why I spend most of my writing pushing the other way. But the reverse mistake is real. AI-native beats AI bolted on when the model is core. When it is not, bolting on is just good judgment.