People Are Not Exception Queues

The new AI workplace has found a very tidy job for the human being. We set the intent, define the quality bar, approve the plan, handle the escalation, review the exception, and own the outcome. It sounds almost noble when written that way, as if the drudgery has been politely swept away and the person has been left with only the rich, adult work of judgment.

This is certainly better than the older automation fantasy, where people were treated as temporary obstructions on the road to a cleaner spreadsheet. The current language is warmer. Enterprise AI vendors talk about agency, orchestration, oversight, trust, and human judgment. Microsoft describes a progression from author to editor to director to orchestrator. OpenAI's Frontier platform talks about AI coworkers that operate inside business systems, complete real work, and escalate when needed. Anthropic's economic research keeps showing that much AI use still looks like augmentation rather than clean replacement. The story has matured. It has put on a blazer. It knows not to say the quiet part quite so loudly.

But I worry that this more respectable story may be hiding a new kind of bad job.

When software absorbs routine execution, what remains for the person is not automatically creativity, strategy, or calm discernment. Often it is the residue. The ambiguous case. The failed handoff. The angry customer who has already explained the problem three times. The missing context. The compliance boundary. The weird invoice. The strange symptom. The policy that almost applies. The output that looks plausible but makes the back of your neck tighten because something in it is wrong, even if you cannot immediately prove why.

That is not a small category of work. It is the bucket where systems put everything they cannot comfortably digest, and then we ask a person to sit beside it with a headset, a dashboard, a queue, and a professional expression. The person becomes less like a craftsperson with a continuous relationship to the work and more like an emergency service for uncertainty. They are not making the thing so much as appearing at the moment the thing has gone sideways.

There is nothing wrong with an exception queue in itself. Every real system needs a place where uncertainty goes to be looked at by somebody with a pulse and a conscience. The problem begins when the exception queue becomes the primary theory of human value. If our best answer to "what are people for?" is "handling what the agents cannot," then we should at least be honest about the psychological shape of that work. It is fragmented, interruptive, and often high-stakes. It arrives stripped of the ordinary context that used to come from doing the ordinary parts first. It can make a person responsible for outcomes they did not shape, only rescue.

This is the human factor that a lot of agentic workplace design still manages to make strangely abstract. In the product demo, the workflow moves beautifully until the machine asks for approval. The human clicks, corrects, redirects, or escalates. The chart improves. Cycle time falls. Everyone agrees that judgment stayed with the person, which is a comforting thing to agree about while watching a slide deck.

But judgment is not a button. It is not a thumbs-up icon with a compliance log attached. Judgment is formed through contact: with ordinary cases, repeated friction, boring patterns, minor mistakes, familiar constraints, and the slow accumulation of taste. It comes from seeing enough normal invoices to know when one smells odd, enough routine support requests to hear when a customer is saying the polite version of something worse, enough mediocre drafts to recognize the one that is quietly dangerous. A person who only sees exceptions may become more important and less grounded at the same time. They are asked to make sharper calls with less lived familiarity. They are told their agency has expanded while their relationship to the work has narrowed.

Exception work also has a way of flattering organizations. It looks advanced. It gives managers a clean story: agents do the repetitive work, people do the meaningful work. But meaning does not automatically live in the leftover category. Sometimes the leftover category is just stress with a better title. Sometimes it is the work nobody has designed well, wrapped in the language of empowerment because "senior escalation specialist" sounds better than "person we summon when the system has run out of confidence."

Microsoft Research's work on multi-horizon task environments points at the problem from the machine side. Real work is not one task at a time, patiently waiting its turn like a well-behaved tutorial. It is dozens of interdependent tasks, partial progress, context switches, delays, dependencies, stale assumptions, and tiny bits of knowledge living in somebody's memory because no database has ever been quite humble enough to contain them. Under that load, agent performance drops. The obvious response is to make agents better, and we should. The humane response is also to ask what happens to the people placed around those drops.

If every degraded edge becomes a human escalation, the workplace does not get calmer. It gets reorganized around interruption. The quiet violence of that arrangement is that it can look efficient from far away. A dashboard can show fewer people touching more work. A manager can see faster cycle times. A vendor can point to the human still "in the loop." Meanwhile, the person in the loop may be living in a day made of alarms, contextless decisions, emotional residue, and the constant feeling of being late to a story that already started without them.

This is where interface design becomes organizational design. The issue is no longer whether a human remains in the loop. That phrase has become too vague to carry much moral weight. A loop can be a healthy rhythm of intention, review, learning, correction, and return. It can also be a tripwire tied to someone who is already tired. We should be much more specific about which one we are building.

Good human-agent systems will need more than approval screens and audit logs. They will need caseload design, not just task routing. They will need ways to preserve context before the escalation, not merely summarize it afterward in six bullet points with the confidence of a consultant. They will need rotations so the same people are not always absorbing the ugliest edge cases. They will need clear ownership for agent-shaped decisions, because "the human approved it" should not become a magic spell that transfers institutional risk onto the nearest employee. They will need recovery time after dense judgment work. They will need feedback paths where human corrections change the system, rather than becoming an endless tax paid by the most competent people in the room.

They will also need to protect some ordinary work, which may sound inefficient if you have spent too much time around people who confuse repetition with waste. Ordinary work is not always noble, and I am not interested in preserving drudgery in amber so everyone can keep suffering with historical accuracy. A lot of routine work is genuinely deadening, and tools that reduce it can be humane. But ordinary work is also how people stay calibrated. It is how a designer keeps feeling the material resistance of layout instead of only reviewing polished approximations. It is how a manager remembers the texture of the tradeoffs before approving an agent-written plan. It is how a clinician holds the full arc of care instead of inheriting only algorithmic uncertainty. It is how a support lead avoids spending the whole day inside other people's worst moments.

Work is not only a production environment. It is a learning environment, a social environment, and a place where identity is built. Strip out the boring middle too aggressively and people may not be freed into meaning. They may be stranded in the jagged parts, expected to be wise on demand, patient on demand, accountable on demand, and grateful that the repetitive work has been taken from them.

The next serious workplace question is not whether agents can execute more tasks. They can, and they will. The more interesting question is what kind of human presence we design around them. If we are careless, we will call people orchestrators while treating them as exception queues. We will celebrate agency while routing every unresolved mess to a blinking notification. We will remove the visible labor and leave behind the invisible burden.

A better version is possible, but it asks for a less glamorous goal than total automation. It asks us to design systems where people can still understand the work, influence the work, recover from the work, and remain more than the place where the machine sends its uncertainty. That is a narrower promise than the future of work usually wants to make. It is also a more humane one.