There is a conversation happening in every boardroom, every Slack channel, every VC partner meeting, about what AI agents can and cannot do. Can they write code? Yes, mostly. Can they handle customer support? Getting there. Can they reason about complex problems? Depends on the problem. Can they replace a senior engineer? Not yet, probably, maybe, it's complicated.
This is the wrong conversation.
The right conversation is not about what agents can do. It is about what humans require. Not their skills, not their output, not their creative capacity. The other stuff. The stuff that never appears in a job description but consumes half of every manager's week and all of HR's reason for existing.
I.
You have a team of six engineers. One of them, let's call him Marcus, has decided to leave. Marcus is good. Marcus has institutional knowledge. Marcus knows why the authentication service is written the way it is, because Marcus wrote it, two years ago, during a weekend when everything was on fire and the previous authentication service had been leaking session tokens.
Marcus has given two weeks' notice, which is polite of him.
Here is what happens next.
His manager has a meeting with HR to discuss the departure. HR initiates the offboarding checklist, which is fourteen items long and includes revoking access to nine different systems, retrieving a laptop, canceling a parking pass, and scheduling an exit interview. The exit interview is conducted by someone from People Operations who will ask Marcus how he felt about the company culture, whether he felt supported in his role, and whether there is anything the company could have done differently. Marcus will say something diplomatic. The People Operations person will write it down. Nobody will read it.
Meanwhile, the manager posts the job listing. The job listing takes a day to write, because it has to be reviewed by Legal for compliance with local employment law, and by Recruiting for keyword optimization, and by the hiring manager's manager to make sure the level and compensation band are correct. The listing goes live. Three hundred applications arrive in the first week. A recruiting coordinator screens them down to forty. The hiring manager reviews forty resumes and selects twelve for phone screens. The phone screens are thirty minutes each. Six hours of phone screens produce six candidates for technical interviews. The technical interviews are four hours each, involving three engineers who are now not writing code. Twenty-four person-hours of interviews produce two finalists. The finalists do a final round with the VP. One of them accepts. The other one was the better candidate but took an offer at Stripe.
The new engineer starts. She needs to be onboarded. Onboarding takes three weeks, during which a senior engineer is partially allocated to answer her questions, review her code, and explain why the authentication service is written the way it is, which is harder now because Marcus is gone and the documentation is a Confluence page last updated eighteen months ago.
Six weeks have passed. The team has produced meaningfully less during this period. The total cost—recruiter fees, lost productivity, interview time, onboarding—is somewhere between $50,000 and $150,000, depending on how you count it and whether you're being honest.
An agent does not give two weeks' notice. An agent does not require an exit interview. An agent does not take institutional knowledge with it when it leaves, because it does not leave, and if you shut it down, you can start an identical one tomorrow and it will not need three weeks to figure out where the bathroom is.
II.
But the departure is the easy part. Departures are clean. They have a beginning and an end and a checklist. The hard part is the ongoing, daily, invisible cost of humans being human while they are employed.
Consider performance reviews.
Twice a year, every manager in the company stops building things and starts writing things. Specifically, they start writing narratives about each person on their team: what they did, how they did it, where they exceeded expectations, where they fell short, what their development areas are, what their career trajectory looks like. Each narrative takes between two and four hours to write, because it has to be specific enough to be useful, positive enough to not demoralize, honest enough to not be dishonest, and calibrated to the company's rating system, which has either four levels or five levels or a 1-10 scale, depending on which framework the Head of People adopted last year.
Then there are the calibration meetings. Eight managers in a room for three hours, arguing about whether a particular engineer is a 3 or a 4, whether "meets expectations" is really fair given the circumstances, and whether the curve should be enforced strictly or whether this quarter is special because of the reorg.
Then there are the delivery meetings. Each manager sits down with each direct report for an hour. Some of these go well. Some of these do not go well. The ones that do not go well require follow-up meetings, sometimes involving HR, sometimes involving a Performance Improvement Plan, which is a document that takes a week to write and satisfies no one.
An agent does not need a performance review. An agent does not have a career trajectory. You do not need to tell an agent that it is meeting expectations while gently suggesting it could show more initiative in cross-functional collaboration. You do not need to manage an agent's feelings about being rated a 3.
III.
People get sick. This is not a complaint; it is a fact. People get the flu and miss a week. People get COVID and miss two weeks. People have chronic conditions that require appointments. People have children who get sick and need to be picked up from daycare. People have mental health days, which are either a sign of a healthy culture or a sign of an unhealthy one, depending on who you ask.
People go on parental leave. This is good and correct and necessary, and it also means that for three to six months, a role is either unfilled or filled by someone temporary who does not know the codebase and will leave just as they're starting to be useful.
People have bad days. Bad weeks. Bad quarters. People go through divorces that halve their concentration. People have parents who are dying. People have anxiety that makes 2 PM meetings feel like walking into traffic. People are not machines, which is both the best and the most expensive thing about them.
An agent does not call in sick. An agent does not need a mental health day. An agent's parents are not dying. This is not a point in the agent's favor, exactly. It is a point about what you are actually managing when you manage people, which is not just their work output but their entire human situation, all the time, whether you signed up for it or not.
IV.
Here is a partial list of things that humans require and agents do not:
Health insurance. Dental insurance. Vision insurance. A 401(k) match. An HSA. An FSA. The difference between an HSA and an FSA, explained, again, during open enrollment, by a benefits coordinator who is also a human being who requires health insurance.
A desk. A chair. An ergonomic assessment. A monitor. A keyboard. A mouse. A laptop. A laptop bag. A second monitor because the first one is too small. A standing desk because sitting is killing them. A chair mat because the carpet is too thick for the chair wheels. An office, or a cubicle, or a seat in the open floor plan, which they hate, which everyone hates, which has been proven by seventeen studies to reduce productivity, and which the company chose anyway because the CEO saw the Airbnb office in a magazine in 2014.
A parking space. Or a transit subsidy. Or a bike room. Or remote work, which requires a VPN, which requires IT support, which requires an IT support person, who also requires health insurance.
Snacks. The snacks are important. The snacks are more important than you think. A company that removes the free snacks will lose more goodwill than a company that cancels the annual bonus, because the snacks are visible and daily and the bonus is invisible and annual. There have been Glassdoor reviews—real ones—where employees gave one star because the company switched from name-brand to generic granola bars.
An agent does not have opinions about granola bars.
V.
The thing nobody says out loud is that a significant portion of what we call "management" is not actually management. It is emotional labor.
You are managing someone who is upset about a promotion they didn't get. You are managing someone who feels undervalued because their work was presented by their manager's manager without attribution. You are managing someone who is in a conflict with another team member about code review comments that were "too harsh," where "too harsh" means technically correct but emotionally unpleasant. You are managing someone who wants to transfer to a different team, not because the work is wrong but because they don't like their current manager, who is you.
You are managing egos and insecurities and ambitions and resentments and politics and alliances and the specific, irreducible fact that every person on your team has a complex inner life that affects their output in ways that are impossible to predict and exhausting to accommodate.
A VP at a company I know spent an entire quarter managing a situation where two senior engineers refused to work together because one of them had made a comment about the other's code architecture at a team offsite. The comment was "I think we should rethink the caching layer." That was the comment. That was the inciting incident. Three months of mediated meetings and HR involvement and ultimately a team restructuring, because one engineer interpreted "I think we should rethink the caching layer" as a personal attack on their competence.
You do not need to rethink the caching layer of an agent's ego.
VI.
And then there is the part that nobody wants to talk about, which is the end.
Not the voluntary departure, where Marcus gives his two weeks and everyone signs a card. The other kind. The kind where you have decided that someone's role is no longer needed, or that someone is not performing, or that the company is contracting and twelve people have to go by Friday.
Layoffs are a specific kind of organizational trauma. They require lawyers. They require severance calculations that account for tenure, local employment law, outstanding equity vesting, accrued PTO, and whether the employee is in a protected class. They require a script, written by Legal and rehearsed by the manager, that uses words like "elimination of position" and "business decision" and does not, under any circumstances, use words like "fault" or "performance" even when performance is the reason, because using those words creates legal liability.
They require a room. They require a box of tissues in the room. They require a security escort for the walk back to the desk, which is humiliating for everyone involved but required by company policy because someone, somewhere, once did something regrettable after being fired. They require a follow-up email to the team, written in a tone that is simultaneously compassionate and matter-of-fact and that absolutely does not disclose the reason for the departure.
They require the surviving employees to process what happened, which they will do for approximately two weeks, during which productivity drops by a measurable amount and the Slack channels fill with anxious speculation about who's next.
You do not lay off an agent. You stop running it. There is no severance. There is no exit interview. There is no box of tissues. There are no surviving agents in the Slack channel wondering if they're next, because agents do not have Slack channels and agents do not wonder.
VII.
I am not arguing that agents are better than humans. That is a category error, like asking whether a hammer is better than a hand. They are different things that do different things and require different things.
But I am arguing that the conversation about AI replacing human workers has been almost entirely about capability. Can the agent do the work? How well? How reliably? What are the failure modes?
These are the right questions for a technology assessment. They are the wrong questions for a business assessment. The business assessment includes everything else. The health insurance and the performance reviews and the severance packages and the exit interviews and the emotional labor and the granola bars and the three months spent mediating a conflict about the caching layer.
When a CEO says "we're going to replace some of our workforce with AI agents," the analyst hears "AI agents can now do what those workers did." But that is only half the statement. The other half, the one that goes unsaid because it would sound callous, is: "and we will shed the entire human infrastructure that surrounded those workers."
The infrastructure is enormous. It is ancient. It is so deeply embedded in how organizations operate that most people cannot see it, the way fish cannot see water. HR departments, benefits administration, office management, employee relations, recruiting, onboarding, offboarding, legal compliance, payroll processing, workers' compensation, unemployment insurance, OSHA compliance, anti-discrimination training, sexual harassment training, diversity initiatives, employee engagement surveys, team-building events, holiday parties, birthday celebrations, the person whose job it is to order the cake.
All of this exists because humans are humans. It is the overhead of being a species that has feelings and bodies and complex social dynamics and a legal system designed to protect them. It is expensive and necessary and, in many individual moments, genuinely good.
And none of it applies to an agent.
VIII.
There is a particular cruelty in this observation, and I want to name it directly.
The things that make humans expensive to employ are the same things that make them human. The sick days are because they have bodies. The emotional labor is because they have feelings. The severance is because they have dignity. The health insurance is because they can suffer. The performance reviews are because they need to grow.
To point out that agents don't require any of this is to point out, by implication, that these human needs are costs. And they are costs. That is the uncomfortable truth. Every humane policy a company implements—generous parental leave, mental health support, reasonable working hours—is a cost that a competitor using agents does not bear.
This does not mean we should stop implementing humane policies. It means we should stop pretending that the competition between human workers and AI agents is happening only on the axis of capability. It is happening on every axis simultaneously. Capability, cost, speed, availability, and the total overhead of maintaining a human being in a professional context.
The agent doesn't need to be better than the human at the work. It just needs to be good enough at the work while also not requiring everything else.
And "everything else" turns out to be a lot.
