Build or buy an AI assistant: how to choose

A subscribed product answers from general knowledge; your customers ask about your prices, your policies, and your promises, and those live only in your material.

Once a company decides it wants an assistant, the market offers two doors. One is to assemble one from components you control: a language model, your own documents, your own rules for what it may promise and when it hands a conversation to a person. The other is to subscribe to a ready product and configure it in an afternoon. Most buyers pick by monthly cost, which is how they end up measuring the wrong thing.

The honest framing is that each door buys something real and charges for something real. Buying gets you speed to a first answer and someone else's maintenance, and it cannot give you grounding in your own material, your own tone, or a handover path you designed. Building gets you exactly those, and bills you for them in ownership: the work of feeding it your documents, defining its boundaries, and reviewing what it says, indefinitely. Neither door removes that last part; the review stays yours either way.

The failure versions are easy to recognize because both feel reasonable at the start. The subscribed bot answers confidently from generic knowledge, and when a customer asks something only the company can answer, it guesses instead of saying that a person will take over, because nobody configured a handover at all. The self-built project consumes months, because its scope quietly grew from answering questions to running the company, and the founder is now maintaining a second business instead of running the first.

The practice is a decision rule, not a preference. Buy when the questions are generic and the risk is low, where a good general answer is genuinely a good answer. Build, or have it built, when the answers must come from your documents, your pricing, your rules for promises, and your discipline for handing a conversation to a person. And whichever door you walk through, the same four laws apply: grounding in real material, boundaries on what it may claim, human review of what it says, and honesty that it is AI.

We chose the build door, and the reason was not ambition. Our answers had to be ours: the fees, the jurisdictions, the handover to a human colleague, the plain admission that the client is talking to AI. No product on a shelf carried our material or our rules, so we assembled the assistant ourselves and pointed it at our own records. That story is an example, not a recommendation: the same choice would be waste for a company whose questions are generic.

We build assistants for companies in exactly that second case, and we tell the first case to go buy something and keep the subscription small. Guidance, not guarantees: the technology choice decides how the answers are produced, and the four laws decide whether customers trust them; whichever route you take, review the answers before your customers do.