Imagine an employee about to book a flight for a conference. Her manager has asked the team to watch costs, and the travel policy is open in another tab. The fare rules are hard to parse. The deadline is close. She asks an internal GPT whether a particular ticket is allowed, and the answer comes back in a clean sentence. She books it.
If the company later rejects the expense, a small label saying the GPT was unofficial will matter less than the answer the employee relied on. The tool was available inside the work. It spoke about a company policy. Someone still has to explain whose answer it was.
In the Moderna case, one small example carried the larger problem. Several travel-and-expense GPTs had appeared, and none had been built by the travel-and-expense team.
The example in Chapter 9 is modest by design. A GPT that creates a background for a presentation and one that answers questions about travel expenses may share a technology. Their work carries different consequences. The interface can make both look like small conveniences. The person using one of them may experience something closer to an instruction.
What job is the tool doing?
The tool’s job is the specific act it performs inside the workflow. “Travel” is a department or process label. It does not tell a leader whether the system is retrieving a policy, summarizing it, interpreting an exception, recommending a booking, or authorizing a cost. Those are different acts, even when they happen in the same window.
This distinction matters because responsibility tends to follow the work rather than the name on the product. The travel team may own the policy. A local employee may have built a helpful GPT. A manager may approve the expense. If the system explains an ambiguous rule, the person who reads its answer may not know which of those people stands behind it.
Each choice can make sense locally. The employee wants a quick answer. The builder sees a recurring question and makes it easier to handle. The policy team has not created the tool and may not know employees are relying on it. Yet together these choices create a new voice inside the organization, without a clear owner for what that voice can say.
Leaders often begin by asking whether a tool is accurate or useful. Those questions matter, but they leave the job description unfinished. Accuracy does not say whether an output is a draft, a recommendation, an instruction, or an official answer. Usefulness does not establish who is accountable when a person acts on it.
When does help start to carry authority?
Help starts to carry authority when a person has reason to rely on the output and the organization has not made its limits visible. A travel assistant that points to a current policy page is doing one kind of work. A tool that tells an employee a fare is reimbursable has moved closer to a decision, especially if the employee has no clear way to check or challenge the answer before booking.
The organization may think of the GPT as an experiment. The employee sees a tool inside the company, answering a question about the company’s rules. Neither description is necessarily false. The gap between them is where expectation forms.
This is why naming a human reviewer is not enough. A reviewer who cannot see the source material, question the interpretation, or stop the recommendation is not the owner of the answer in any meaningful sense. The person using the tool needs to know when the system is preparing information and when a person must take responsibility for what follows.
Chapter 9 describes three possible roles: execute, support, and stop short. A system can execute a stable, bounded task when exceptions are visible and the result can be corrected. It can support a person by finding relevant policy language or organizing information for review. It should stop short when the work requires an accountable interpretation, or when the answer itself could become a commitment the organization must defend.
That boundary is specific to the act, not the label on the tool. The same assistant may retrieve a standard mileage limit and still be unable to decide whether an unusual expense deserves an exception. The first task applies a known rule. The second asks someone to interpret what the rule means in a particular situation.
What belongs in the job description?
An AI job description should name what the system may do, whose authority it uses, and where a person must take over. It should describe the acts the tool can perform, the information it may rely on, and what happens when the case falls outside those limits. “Answer travel questions” is too broad to settle any of those points.
For a travel assistant, a narrower assignment might allow it to retrieve the current policy and point to the relevant passage. If the source is unclear, the assistant should route the employee to the travel team rather than turn uncertainty into a confident answer. If the tool is permitted to interpret or recommend, the organization needs to name who owns that interpretation and how the employee can ask for a correction.
These decisions cost time. Someone must keep the source material current, watch for questions the tool cannot answer, and give employees a real path back to a person. The cost does not disappear when the tool is called unofficial. It moves to the employee who must decide whether to trust it, and to the manager who explains the expense afterward.
The modesty of the Moderna example is useful. Nothing dramatic has to happen for role confusion to matter. Several small tools can appear around a shared process, each made to remove a little friction. Before long, employees may be taking answers from systems whose responsibilities no one has actually assigned.
The employee still has a booking open. The fare looks right. Somewhere in the company, a policy owner could answer the question, if the employee knows to ask and can reach them before the deadline. The GPT has already answered. The organization still has to decide what that answer was allowed to mean.
This essay draws on Chapter 9, “Give the Machine a Job Description,” of AI in the Org Chart.