The decisions an AI assistant should never make alone

The line is not how clever the assistant is: it is which decisions are reversible, and who owns the few that are not.

Give an assistant real work and within a week it will be quoting, reassuring, promising. Most of that is exactly what you wanted: speed in places where silence used to be the answer. Then one day you read back a conversation and find the assistant made a decision you would have wanted on your desk: a discount offered, a date committed, a scope accepted. Nothing about that moment is mysterious. The assistant was never told where the lines are, or it was told once, in a document nobody reads.

Our list of decisions an assistant never makes alone has three headings, and they are short enough to memorize. First, money that leaves the company: pricing exceptions, refunds, credit terms, anything that reduces what a customer owes or what the business earns. Second, commitments beyond what the sources say: delivery dates, custom scopes, promises about outcomes that no operator has confirmed. Third, anything that needs the weight of a signature: terms of business, disputes with a legal dimension, statements that could bind the company to something.

None of this is a verdict about AI being weak. It is a statement about reversibility. An assistant that drafts, proposes, and prepares is a genuine force, because almost everything it produces can be undone cheaply by a human with a keyboard. The few decision types that cannot be undone, the ones that move money or create obligations, are exactly the ones a named human should own. That division is the point: draft at machine speed, decide at human speed, and keep the deciding where accountability lives.

The failure versions are all real conversations somewhere. The assistant that approved a refund path it invented mid-conversation, because the customer was upset and a solution lay within reach. The quote that went out carrying a delivery promise no operator ever confirmed, which operations met as a complete surprise. The cheerful yes-we-can to a request sitting entirely outside the service line, discovered weeks later as a project nobody had priced. In each case the machine did what it is built to do: it was helpful. Nobody had written down where helpfulness has to stop.

The practice is to write the list down as standing rules, not as nagging appended to every message. Rules that hold for every conversation need no reminder to work; they are the shape of the system. Route those decision types to a named human, with a fast path, so escalation is a handoff and not a queue. And let the assistant say honestly: this one goes to a colleague. Customers accept that sentence remarkably well when it arrives consistently. What they do not accept is an assistant that says yes to everything, followed by a company that cannot deliver what was said.

We run our own funnel on exactly this split, and our assistant says plainly that it is AI, and that certain answers always route to a human, because we would rather be trusted than mistaken for something we are not. Where your own lines sit, and how your tools enforce them, depends on your systems and the rules that bind your industry. Guidance, not guarantees: write the short list, hand the few irreversible decisions to a person, and let the machine do the rest of its excellent work.