Nobody Hires an Employee Who Cannot Be Held Accountable
In February 2024, a Canadian tribunal wrote down a sentence that should be taped to the monitor of anyone about to buy an AI employee. An airline had deployed a chatbot. The chatbot gave a grieving customer wrong information about bereavement fares. When the customer sued, the airline argued, in writing, that its chatbot was a separate legal entity responsible for its own actions.
The tribunal member called that a remarkable submission, which is judicial language for you cannot be serious. The airline lost. It owned the chatbot, so it owned what the chatbot said.
Now hold that ruling next to the pitch decks. A wave of products arrives under the banner of the AI employee, and the word is doing almost all of the selling. Employee means reliable. It means answerable. It means you can hand it a task and stop thinking about it, the way you stop thinking about work you gave a competent colleague. That is the whole appeal. It is also a promise, and most of these products cannot keep the part of the promise that counts.
Here is what you can do with an actual employee. You can read back their work and find the mistake. You can ask why they made it. You can stop them before a bad decision becomes a permanent one, and reverse it if it already has. Strip the branding from most AI employee products and none of that is on offer. What remains is a worker whose errors land on you, reasoning you cannot inspect, and actions you often cannot take back. That is not an employee. It is a liability with a subscription.
The word does the selling, and the word is a promise
Vendors reach for the word employee because it does an enormous amount of work in a single syllable. It says reliable. It says accountable. It says you can hand this a task and stop thinking about it, the way you stop thinking about a task you gave a competent colleague. That is the entire appeal, and it is why the label spread so fast.
But a word that borrows the trust of employment also borrows the obligations. A real employee comes with a whole apparatus you never itemize because you take it for granted. There is a record of what they did. There is a person who answers for it. There is a way to stop them before a mistake becomes permanent, and a way to reverse it after. When a vendor uses the word and ships none of the apparatus, the word is not a description. It is a costume.
The gap is not academic. It shows up the first time the software is confidently wrong.
The blame does not transfer, no matter what the box says
The details of that case matter, so sit with them. Jake Moffatt booked a last-minute flight after his grandmother died and relied on the chatbot’s answer in good faith. The correct policy sat on a different page of the same website. In Moffatt v. Air Canada, the tribunal found the airline liable for negligent misrepresentation and ordered it to pay $812.02 in total, of which $650.88 was the fare difference and the rest interest and fees.
The dollar figure is trivial. The principle runs in exactly the wrong direction for the AI employee pitch. A company deployed an automated worker, the worker was confidently wrong, and the company tried to hand the blame to the software. The tribunal handed it straight back. You do not get to enjoy the labor of an automated system and disown its mistakes. If you put it to work, you own what it does.
Read that through the lens of the marketing one more time. The AI employee you are being sold is a worker whose errors are yours, running on reasoning you cannot see, taking actions you frequently cannot reverse. The word promises a colleague. The arrangement hands you unlimited downside with the evidence locked in a box.
The market already senses the problem
You can see the hesitation in the numbers, even from buyers who badly want this to work. Research from Workato and Harvard Business Review found that 86% of organizations plan to increase their investment in agentic AI, while only 6% trust agents to run a core process end to end.
Sit with the distance between those two figures. Nearly everyone is spending more. Almost nobody trusts the thing they are spending on to run a real process without a human standing over it. That is not confusion. That is an accurate read of the accountability gap. People are buying capability and quietly refusing to grant it authority. Authority without accountability is how you end up explaining a mistake you cannot see, to a customer you cannot appease.
Gartner sees the same fault line from the other direction, one we track in our industry coverage. It expects more than 40% of agentic AI projects to be canceled by the end of 2027, and among the named causes is inadequate risk controls. Not weak models. Not slow performance. Controls. The projects that die are the ones nobody could govern, and a thing nobody can govern is a thing nobody can hold accountable.
The lesson underneath both numbers is the same one the tribunal delivered. Capability is not the scarce resource anymore. Accountability is. And a system that cannot be held to account is not a bargain at any price, because the price does not include the day it is wrong.
What AI accountability actually looks like in software
Here is the part vendors skip, because it is harder to build than it is to name. Accountability is not a feeling or a compliance checkbox. It is four concrete properties, and a system either has them or it does not.
A visible plan. Before acting, the system shows you what it intends to do, in language you can read, so you can catch a bad decision while it is still a proposal rather than a consequence. A worker that acts first and explains later, if it explains at all, has removed your only chance to say no.
Confirmation before anything irreversible. Not every action needs a checkpoint. Filing something reversible does not. But sending a message to a new recipient, moving money, deleting a record, anything with a blast radius that cannot be walked back, should stop and ask. The line sits at reversibility, not at importance, because importance is subjective and reversibility is not.
Undo. When something can be reversed, it should be, on your command, without a support ticket. A worker you cannot correct is a worker you have to supervise constantly, which defeats the entire purpose of having one.
A decision record. Every action leaves a trail you can actually read, sitting where the work happened rather than buried in a log you will never open. It is the same instinct behind our public trust center: the record should be something you can retrieve, not something we ask you to take on faith. When you need to answer for what the system did, and the Air Canada case is a reminder that you will need to, the answer should be one you can retrieve in seconds, not reconstruct from wreckage.
Those four are not a wish list. They are the minimum for the word employee to be honest. Miss any one and you have reintroduced the gap. A plan with no confirmation still lets the mistake through. Confirmation with no record still leaves you unable to explain what happened. The properties only work as a set.
How Pax approaches it
This is the standard we hold Pax to, and it is worth being precise about what that means in practice rather than in adjectives.
Pax is a governed AI worker. It reads, reasons, and acts across the tools a professional already uses, and it does real work end to end. But it works inside those four properties by design, not as an afterthought bolted on for the enterprise tier. Before Pax takes an action that cannot be undone, it shows the plan and asks. Routine, reversible work moves without friction. Higher-stakes actions, the ones aimed at a new recipient or carrying real consequence, stop for your confirmation. When an action can be reversed, you can reverse it. And every decision Pax makes is recorded in the conversation itself, next to the messages it responded to. The reasoning sits where you are already looking, not in a separate audit view you have to go hunting for.
None of that makes Pax slower in the ways that matter. It makes Pax answerable, which is the thing the word employee was promising all along. You can see what it intends. You can stop it. You can reverse it. You can find out why. That is not a smaller version of an AI worker. It is the only version a serious business can actually put to work.
