Field Guide · hiring
When Not to Hire an AI Agent
Four scenarios where the right answer is to keep your money in the bank or hire a human. The cost of getting an AI agent wrong is real; the cost of admitting when not to buy is small.
Saying when not to buy is the most useful thing a vendor can say. It is also the rarest. The Field Guide is mostly an argument for why agents work. This piece is the argument for when they do not. Four scenarios. Each one is a real shape the platform has talked a buyer out of. Each one ends with the same advice: do not buy. Save the money. Hire a human, or wait, or use a tool one rung simpler. The honest version of the agent pitch includes naming the cases the agent is not for. This is that list.
Why it matters
Buyers in the early phase of an AI buying decision are bombarded with vendors who will say yes to every use case. The buyer who has already seen three demos, watched two sales decks, and read four landing pages cannot tell which case is real and which is a stretch. A vendor who names the four cases they are not for gives the buyer a frame that holds across vendors — if the case is on this list, no vendor's agent fits, and the right move is something else. That frame is more valuable than another sales pitch. It is also the only way to ship the AI hire that the team will not resent six months in.
1. The team needs accountability somebody can shake hands with.
If the work the buyer is hiring for is the kind that requires a human in the room — a regulator meeting, a key-account renewal, a press call, a recruiter conversation that decides someone's career — the agent is not the answer. The answer is a human. The agent does not pretend to do this kind of work, and the buyer who tries to use an agent for it ends up apologizing to the counterpart on the other side. The honest version of the platform's pitch acknowledges this on every Roster entry's limit list. If the work is shake-hands work, hire a person.
2. The team's work is too unstructured for a constitution to capture.
Some roles do not have a stable shape. Early-stage founders, experimental product roles, creative direction in seasons of brand reinvention — the work changes week to week, and the constitution that would govern an agent built for it would have to be rewritten faster than the agent could deploy. The right move in these roles is to use a chatbot as a thinking partner, keep the work in human hands, and revisit when the role's shape has settled. Trying to ship an agent into an unstable role produces a constitution that is constantly out of date and an agent the team distrusts.
3. The cost of a private failure is a public failure.
Some categories of work look private but are one button-press away from public — drafting a regulator letter, writing the announcement for a layoff, responding to a security incident. A draft an agent produces in Slack is private until somebody clicks send. If the cost of an accidental send is significant — the kind of incident that ends up in the news — the right move is not to use an agent for that work, even with strict review gates. The agent is not for the work whose private draft can become a public crisis with one click.
4. The team is too small to integrate the agent.
A two-person team does not have the surface area to integrate an agent. The agent's value comes from working a stream of work in shared sight. A team of two does not have enough volume in the stream and does not have enough sight to share. The right move at two is to use a chatbot as a tool, hire the third person whenever the team can afford it, and revisit the agent question after the team is five or six and the stream of work has shape. Trying to deploy an agent before the team can absorb it produces an agent nobody opens.
What the cases share.
Each of the four cases is recognizable. Each can be by the buyer in advance — they know if their work needs a human in the room, if their role's shape is stable, if a private draft can become a public crisis, if the team is large enough. The platform's Roster constitutions all name some version of these cases on their limit lists. The buyer who reads the limit list and recognizes their case in it has been told not to buy. That is the right outcome. The buyer who buys anyway is the buyer whose review six months later starts with the word disappointed.
The edge
The reason this piece is on the platform — and not just on a competitor analyst's blog — is that it is more valuable to be right about when not to sell than to be lazy about when to sell. Saying yes to every case sells more agents in month one. It produces churn in month four, refund requests in month six, and reputation damage in year two. Saying no to the four cases that don't fit produces a customer base that fits, talks well about the product, and stays. The math on costly signals favors the second strategy. The competitive advantage is a vendor who is willing to be honest about when their product does not work.
Honest take
There are edge cases inside each of these four where the agent is closer to right than wrong. Some shake-hands work has a draft surface that benefits from agent input even though the meeting itself is human. Some unstructured roles are stable enough at the seam where an agent can help with one bounded slice. Some public-failure work has a review gate strict enough that an agent draft is safe. Some two-person teams have a third person in week eight. The honest version of this list does not pretend the boundaries are sharp. It pretends only that the boundaries are recognizable, and that the buyer who recognizes their case in the list deserves the answer this list gives.
Four cases where the answer is not to buy. The platform exists because the cases where the answer is yes are larger than the cases where it is no. Naming the no cases is the cheapest way to keep the yes cases honest.