Dr. Jonah Tebaa, DBA Founder and Co-Chief Executive, Webspot · PGPro · Artwist
For two centuries the dominant question about technology and work has been which jobs will the machine take. This essay argues that the question is the wrong one, and that asking it has left a generation preparing for the wrong fight. Synthesizing the task-based economics of automation (Autor; Acemoglu and Restrepo), the augmentation versus automation distinction (Brynjolfsson), and the philosophy of human work (Polanyi; Arendt), I advance three claims. First, that artificial intelligence rarely replaces a person outright. It dissolves the function a person performs while leaving, and often increasing, the value of the judgment, accountability, care, and meaning that surround the function. Second, that the defining shift of this decade runs deeper than a move from one job to another. It is the shift from supplying labor to directing it, as the knowledgeable individual becomes the manager of a small, tireless, nonhuman workforce. I propose the term e-mployees for that workforce: artificial agents an individual delegates to, supervises, and remains accountable for, as a founder directs a team. The human on the other side I call an e-mployer. Third, that the capacities which survive and compound in this world are characterological rather than technical, and that they can be deliberately built. I treat this as conceptual theory-building, not forecast, so I also state where the argument could fail, who it does not serve, and the conditions under which it would be wrong.
Keywords: future of work; artificial intelligence; agentic AI; e-mployees; automation; human capital; comparative advantage; career development
A quiet sentence has shaped how most people understand their working lives, and it is false. The sentence is: you are what you do. Not in the sense the poets mean, but in the flat organizational sense, the sense in which a person is the set of tasks they execute, a node in a workflow, a function with a salary attached. Tell me your title and I can list your tasks; automate the tasks and I have automated you. This is the logic on which a generation was told to build a life. It is the logic artificial intelligence is now dismantling, and the dismantling is, in the end, good news wearing a frightening mask.
The fear is real because the lie was believed. If you are your function, and your function can be done by a system that costs a few cents an hour and never tires, then you are in genuine danger and no reassurance should comfort you. But the premise was never true. A human being was never merely a function. The function was the part the organization could see and price, the visible tip of something larger and harder to name. What is happening now is that the visible tip is being commoditized at extraordinary speed, and the submerged mass beneath it, which was always doing the real work, is surfacing into view.
This essay is about that submerged mass. It is written for the person, especially the young person, who was told to pick a skill, become the skill, and ride it for forty years, and who now suspects, correctly, that the advice has expired. The argument is easy to state and hard to live. The route to staying useful is not to become a better function, because the machine will become a better function faster than you can. The route is to become what a function cannot be: a person who directs, judges, and stands behind the work of machines, and who grows more valuable as the machines grow more capable.
Getting there takes three moves. We have to look honestly at what automation has actually done to human work, because the record contradicts the panic and also refuses to flatter it. We have to find what in a human being is genuinely hard to automate, and test that claim rather than assert it. And we have to describe the new role the technology is pressing on every capable person, ready or not: the e-mployer, the director of e-mployees.
Begin with the history, because the history is more interesting than either the optimists or the doom-mongers admit.
In 2013, two Oxford researchers, Carl Frey and Michael Osborne, estimated that 47 percent of United States employment sat at “high risk” of computerization over the following decade or two (Frey and Osborne, 2013/2017). The number circled the world and became the headline under which a decade of anxiety was filed. More than ten years on, the observed pace of automation has run well below that ceiling. Occupations did not vanish at anything like the implied rate, and the jobs most exposed tended to change rather than disappear. Two cautions keep this honest. The estimate measured susceptibility, not a dated prediction of unemployment, a distinction its popularizers dropped; and the longer horizon has not fully elapsed. Even so, the core lesson already holds, and it is not “relax.” It is that the model behind the forecast, a job seen as a fixed bundle of tasks awaiting deletion, does not describe how technology and work actually interact.
David Autor built the better model. His insight is plain: automation acts on tasks, and a job is a shifting bundle of tasks rather than a fixed one (Autor, 2015). Automate one task inside a job and you rarely delete the job. More often you change its mix, raising the weight of the tasks that remain and that the machine cannot do. Autor’s signature example is the automated teller machine. The intuitive prediction in the 1970s was that the ATM would erase the human teller. Instead, as James Bessen documented, teller employment did not collapse. It grew over the following decades, from roughly a quarter-million in the early 1970s toward a peak near 600,000 around 2010, even as ATM units multiplied into the hundreds of thousands (Bessen, 2015). The machine made each branch cheaper to run, so banks opened far more branches, and the teller’s work shifted from counting cash toward selling and solving problems. The honest version of the story includes its losers: the person whose entire skill was counting cash was not automatically the person who moved into relationship work. A role was redefined, and specific people bore the cost of the redefinition. But the aggregate moved the opposite way from the panic.
The pattern recurs. The spreadsheet was supposed to end the accountant. What followed was a fall in routine bookkeeping clerks and a rise in accountants, auditors, and analysts, because the software removed the arithmetic that had been crowding out the judgment (Autor, 2015). Again the same shape, and again with displaced people inside it. The machine took the calculation; the surviving and growing role was the meaning of the calculation.
None of this proves automation is painless. Daron Acemoglu and Pascual Restrepo give the sharpest account of the cost. Every wave produces a displacement effect, in which machines take over tasks and push labor out, and a reinstatement effect, in which new tasks and roles return labor to the center (Acemoglu and Restrepo, 2019). When reinstatement keeps pace, work flourishes and wages rise; when displacement runs ahead, the result is the hollowing-out of middle-wage work we actually lived through: stagnant pay, polarized opportunity, communities stranded. Their empirical decomposition finds exactly that imbalance over recent decades, with accelerating displacement, a weakening reinstatement effect, and too few new human tasks to absorb the people automation freed. So the history does not whisper relax. It says something harder and more usable. The outcome is not written into the technology. It turns on whether enough valuable new human work is created alongside the machines, and, as the next sections argue, on whether individuals position themselves to do that work.
One number is worth carrying out of here. In its Future of Jobs Report 2025, the World Economic Forum projected that by 2030 structural shifts would create about 170 million jobs and displace about 92 million, a net gain near 78 million, atop churn equal to roughly 22 percent of all current jobs, while employers expected close to 39 percent of workers’ core skills to be transformed or outdated over the same window (World Economic Forum, 2025). The direction is corroborated elsewhere: the McKinsey Global Institute earlier estimated that as many as 375 million workers, about 14 percent of the global workforce, could need to change occupational category by 2030 (Manyika et al., 2017). Read those together and the threat sharpens into focus. It is not mass joblessness. It is obsolescence inside continued employment, the experience of keeping a job while the skills that justify it quietly expire underneath you. That is a different danger, and it needs a different defense.
If the machine keeps eating tasks, the only durable strategy is to stand where the machine cannot easily follow. So we must ask, more rigorously than the question usually gets asked, what a human can do that an artificial system cannot, and whether that boundary is real or merely the place the machines have not yet reached.
Start with a paradox named for the chemist and philosopher Michael Polanyi, who observed in 1966 that “we can know more than we can tell” (Polanyi, 1966). A master cannot fully state the rules by which she judges her own work; you recognize a friend in a crowd without specifying how. Autor named this Polanyi’s Paradox and treated it for years as the wall automation could not climb: machines could only do what we could fully specify, and the most valuable human work resisted specification (Autor, 2015). Intellectual honesty requires conceding that modern machine learning has partly scaled that wall. A model now recognizes a face without anyone writing the rule. So if perception, once “tacit,” fell, why should any other human capacity be safe?
Because not every wall is the same wall, and the distinction has to be argued rather than asserted. The machines climbed the wall of perception and pattern, the class of tasks where the answer already exists in the data and the only problem was extracting it. They have not climbed, and may not climb, the wall of consequence, the set of structures that require a bearer with something at stake. Consider three of them.
The first is judgment under genuine novelty. Not pattern-matching against the past, which machines do superbly, but choosing wisely when the situation is new, the data thin, the values in conflict, and someone must simply decide and own the decision. A model can report what usually happened. It cannot want an outcome, and it cannot be wrong in the way that costs, with consequences it must carry.
The second is accountability, and here the strongest objection deserves a real answer. One might say that accountability is a social and legal construct, not a metaphysical gift. We already assign liability to corporations, which are not persons; we insure outcomes and bond contractors. Why will markets not simply route accountability through insurers, operators, and audit layers, leaving the human “voucher” a thin, low-wage notarizing function, commoditized exactly as the teller’s cash-counting was? The answer is that every one of those mechanisms terminates in a human who is the residual claimant of consequence, someone with capital, liberty, or reputation genuinely at risk. Insurance reprices risk; it does not abolish the party who is finally answerable when the model is confidently wrong and the harm is real. As machines do more of the producing, the scarce act becomes the credible human who says I stand behind this and can be made to pay if it fails. I will not pretend this moat is unconditional. If autonomous agents one day attain bonded, legally answerable standing of their own, this wall falls, and the doctrine that follows falls with it. That is the condition. It has not been met, and there are reasons, legal and moral and practical, to doubt it will be met soon.
The third is care and culture, the work whose value is that a human is doing it. Here too, honesty forbids absolutism. AI substitution is already underway at the low end. People use AI tutors, companions, and basic counseling tools today, and at a large enough gap in price and quality, many consumers accept the machine. The human moat in care is therefore shrinking rather than vanishing. It concentrates at the high-stakes, high-trust end, where we still refuse to have our children raised, our gravely ill comforted, or our disputes finally judged by something that cannot care whether we live. And above care sits the broadest human output of all. Hannah Arendt distinguished labor (the repetitive work of survival), work (the making of durable things), and action (the distinctly human business of beginning something new and being seen to do it) (Arendt, 1958). Automation is the conquest of labor and much of work. Action it cannot touch: the founding of a venture, the framing of an idea, the setting of taste, the act that says here is a new thing, and it is mine, and it means something. No quantity of generated tokens adds a line to the record of human action. A machine’s output can be excellent. Only a person’s can be significant.
This is why “you are your function” was always a lie. Your value never lived in the function. It lived in the judgment you brought to it, the responsibility you took for it, the care you gave it, and the meaning you made of it. The machine is stripping away the function and leaving those four standing in the open, provided you built them, rather than letting them wither while you trained to be a faster clerk.
Now the shift everything has been building toward, larger than a change of tools. For the whole history of paid work, most people held one structural position. They were the labor. They sold hours and skill, and someone above them combined that labor with capital and direction to make something of value. To “have a career” was to become a better, more senior instance of labor.
Artificial intelligence inverts the position. For the first time an ordinary individual, not a corporation but a single person with a laptop, can command labor rather than only supply it. The model that drafts your document, the agent that researches your market, the system that writes and tests your code, the tool that designs your images and answers your customers: these are not merely sharper instruments in your hand. They are workers you direct. You set their objectives, delegate the parts, review what returns, correct it, and remain accountable for the result. That is not the relationship of a craftsman to a tool. It is the relationship of a manager to a staff.
I propose a name for that staff: e-mployees.
E-mployees: artificial agents, that is, AI systems and software agents, that an individual directs as a personal workforce, delegating objectives to them, supervising and reviewing their output, and remaining personally accountable for what they produce. To work with e-mployees is to move from being labor to directing it. The person becomes an e-mployer.
The pun carries the argument. An employee is a unit of directed labor inside someone else’s enterprise. An e-mployee is a unit of directed artificial labor inside your own. The same person who, in the old world, was somebody’s employee can, in the new one, become an e-mployer, the holder of a one-person enterprise with a workforce of machines.
This is the point to name the work standing closest to it. Erik Brynjolfsson warned of the “Turing Trap.” When we build AI to imitate humans, machines become substitutes and workers lose bargaining power; when we build it to augment humans, machines become complements and people stay indispensable (Brynjolfsson, 2022). Autor’s recent work pushes the augmentation case further, arguing that AI’s distinctive opportunity is to “extend the relevance, reach, and value of human expertise,” letting more middle-skill workers perform the high-stakes decisions now reserved for elite experts, and so potentially rebuilding the middle class that earlier automation hollowed out (Autor, 2024). Both name the fork between substitution and augmentation. Neither specifies the human role on the augmentation side. That is the gap this essay fills. The e-mployer is augmentation made personal and concrete, a buildable individual competency for living on the complementary side of the line rather than a policy hope that the economy will choose complementarity on one’s behalf.
Make it concrete. Consider a case, composite but unremarkable, of a nineteen-year-old who wants to help a neighborhood restaurant that has no website and no time. She does not write code or design. She directs three e-mployees: one agent to research the local market and competitors, one to build a simple site, one to draft the owner’s outreach to customers. The research agent returns a crisp competitive summary that includes precise rival prices, confidently, and partly invented. She catches it, because she walked two of the restaurants herself and the numbers are wrong, and she sends the agent back with the real figures. The build is neither free nor frictionless. The tools cost her a month’s coffee budget, and the site-builder’s first two drafts come back generic and wrong before her direction gets specific enough to be worth anything. By Sunday, though, the site is live and the owner has emailed forty regulars. She wrote almost none of it. What she supplied was the thing the machines could not: the objective worth pursuing, the local knowledge to catch the confident error, and her name on the result when she handed it to a real person who was counting on it. That is an e-mployer, at the very bottom of the ladder, doing the whole job in miniature.
Notice what the inversion does to the old fear. “Will AI take my job?” is the question of someone who still identifies as labor and sees the machine as a rival for the same task. “How large a workforce of e-mployees can I direct, and toward what?” is the question of someone who has crossed to the other side of the relationship. The first has no good answer. The second has no ceiling. Much of this essay’s practical advice compresses to one instruction: cross over. Stop competing with the machine at the task. Direct the machine at the task, and compete instead on the things that decide whether the direction is any good.
Saying become an e-mployer is useless without saying what the role demands, because a competency you cannot specify is one you cannot deliberately build. I offer the following as the E-mployee Doctrine: five capacities that separate a person who merely uses AI tools from one who genuinely directs an artificial workforce. The first three are managerial virtues sharpened to the specific grain of directing machines; the last two are where the human becomes irreplaceable. Each is stated as a working claim, not a slogan.
1. Delegation, the seeing of the shape of the work. The novice asks a model to do the small thing in front of them. The e-mployer decomposes a large objective into tasks and decides which belong to machines and which to humans. The binding constraint in an AI-rich world is rarely the machine’s capability. It is the human’s ability to frame an objective worth pursuing and to cut it into parts a machine can take. The scarcer the good objectives, the more the value concentrates in whoever can set them.
2. Direction, the specifying of what you actually want. An e-mployee does exactly what it is told, which rewards only the person who knows what they want. Vague instruction yields fluent worthlessness at scale. This is the AI-specific edge of ordinary communication: the machine will not infer your standard, will not ask the clarifying question a junior colleague would, and will fill every ambiguity with confident guesswork. Specifying purpose, constraint, and the standard of “done” becomes an engineering discipline.
3. Discernment, the judging of what comes back. This is the capacity most particular to directing machines, and the most underrated. An e-mployee hands you a confident, fluent, well-formed answer that is sometimes wrong, never flags which times, and never says “I don’t know.” The e-mployer must supply the skepticism and the taste the machine lacks, must know enough, and care enough, to catch the error the model cannot see in itself. This is why deep knowledge does not become obsolete. It becomes the instrument of supervision. You cannot review what you do not understand. A person with no expertise and a thousand e-mployees does not have a workforce. They have a thousand unverified guesses.
4. Accountability, the owning of the result. The e-mployer signs their name to what the machines produce. This is the capacity a machine structurally cannot hold and therefore cannot take: the willingness to be answerable, with something real at stake. In a world flooded with generated output, the scarce and valuable act is the human who says I stand behind this and can be held to it. Reputation becomes the currency, because reputation is the one asset an e-mployee cannot hold on your behalf.
5. Orchestration, the composing of the whole. The highest form of the role is not directing one agent at one task but conducting many toward a single coherent end, the way a founder runs a company or a director runs a film. The value is not in any single output but in the integration, the vision that makes a hundred machine-made parts add up to one human-meant thing. Here the e-mployer becomes, fully, an entrepreneur, a person who assembles labor and direction into something that did not exist before.
The doctrine carries an uncomfortable implication worth stating plainly. Every one of these five is a human capacity, and not one is trained by becoming excellent at a routine task. The professional habits of the last century optimized for the opposite, for becoming a reliable function, and those habits now actively endanger the people who hold them. The clerk who perfected the clerical task trained precisely the thing the machine does best and the human role abandons. The future does not reward the best functions. It rewards the best directors of functions.
A doctrine is only as good as the person running it, and the e-mployer role makes one demand above all. It must be performed by someone who never stops changing. If close to two-fifths of today’s skills will be transformed by 2030 (World Economic Forum, 2025), then any plan resting on a fixed skill carries an expiry date. The durable asset is not a skill but the capacity to acquire skills, and that capacity can itself be built or neglected.
The economics make the case before any psychology does. Skills now reprice on a cycle far shorter than a career, which means the meta-skill of learning new tools quickly dominates the value of any particular tool you currently hold. The practical discipline follows directly. Treat your own skills the way a portfolio manager treats depreciating assets. Assume each is losing value and reinvest continuously. Spend a deliberate fraction of every week working at the edge of the current tools, not because any one of them will last, but because the muscle of learning the next one fast is the single capability that compounds across all of them. To remain “current” is not to have arrived at the present. It is to have built the engine that keeps you arriving.
Psychology supports the economics rather than leading it. The disposition the moment rewards is the one Carol Dweck called a growth mindset, the treating of ability as developed rather than as a ceiling fixed at birth, though its effect is more modest and context-dependent than the popular version claims, so I take it as illustration, not proof (Dweck, 2006). The sturdier finding is structural. David Epstein’s survey of performance in complex, changing fields shows that breadth often beats early narrow specialization, because the capacity to draw analogies across domains and to start over predicts success precisely where the ground keeps moving (Epstein, 2019). The specialist is optimized for a world that holds still. Ours does not. The e-mployer, directing machines across writing and analysis and design and code and customer relations, is by construction a generalist with the judgment to call in depth, their own or another human’s, when it is needed.
An honest doctrine names the conditions under which it fails, the people it does not serve, and the evidence that would refute it. This one has all three, and skipping them would make the essay a comfort rather than a claim.
It could be wrong about the wall. The whole prescription rests on the accountability moat of Section III, the claim that a human must remain the residual bearer of consequence. If autonomous agents acquire bonded, legally answerable standing, and if “vouching” compresses into a thin compliance task that pays accordingly, the moat is shallow and the e-mployer’s edge erodes from underneath. I have argued why I doubt this happens soon. I have not proven it cannot. That is the falsification condition. Demonstrate reliable, accountability-bearing autonomous agents, and the case collapses.
It could be wrong about the macro balance. The optimism here is conditional on reinstatement roughly keeping pace with displacement, and Acemoglu and Restrepo show it has not, lately. Brynjolfsson’s Turing Trap is precisely the scenario in which AI automates the new tasks too, including the orchestrating work the e-mployer depends on, so that augmentation never arrives at scale. If that path wins, “become a director of machines” describes a shrinking island, not a rising tide. The doctrine assumes a tilt toward augmentation in markets and policy that is a genuine choice, not a guarantee.
And it is, by construction, a strategy for a minority. Directing AI rewards those who already hold the inputs: capital, education, networks, judgment, risk tolerance. The nineteen-year-old of Section IV is a real possibility, but she is also the survivor case, and the same logic that lets her command cheap labor lets it compete against everyone who cannot. Honestly read, the e-mployer path concentrates gains among the already-advantaged. I do not claim it nets out positively for everyone, or that “take responsibility for a real outcome” is equally available to all. It is not. What I claim is narrower and, I think, defensible: that for an individual deciding what to do with the next decade, becoming a director of artificial labor is the best available hedge against obsolescence, and that the distributional question of what happens to those who cannot cross over is real, unsolved by this essay, and owed a different and more political answer than the one offered here.
Stating these does not weaken the prescription. It earns it. The advice that follows is for raising the odds of the person who can act on it.
To the person at the start of a career, the abstractions resolve into a few concrete choices, and they are not the choices most have been advised to make.
Stop optimizing to be hired as a function. “Pick a marketable skill and master it” was sound when a skill lasted a working life. It now means building, at great cost, the very thing the machine is cheapest at. Acquire skills, because you cannot direct or verify what you do not understand, but acquire them as equipment for judgment, not as the product you intend to sell. Learn to code so you can direct and check machines that code, not so you can race them at typing.
Run a real e-mployee team this month, on a real problem. Do not wait for a job or a title. Pick one actual problem you have, a club that needs a site, a dataset you want answered, a small thing you could sell, and direct three agents at it: one to plan, one to build, one to check the other two. Ship it to one real person by the end of the week. You will hit the friction immediately, and you should, because good tools cost real money each month, the learning curve is real, and bad direction wastes both. You will also watch an agent hand you a confident, wrong answer, and catching it will teach you more about your own value than a semester of theory. That catch, the moment your knowledge supervises the machine, is the whole job in miniature. Do it five times and you will have trained every capacity in the Doctrine.
Build the four irreducibles on purpose. Put yourself, early and often, where you must exercise judgment under real ambiguity, be accountable for outcomes you cannot fully control, actually care, and make something that means something. These are not soft skills to add after the technical ones. They are the foundation, and they are built only by doing, usually uncomfortably. A young person who has owned a real failure, run the project, faced the customer, has trained something no curriculum and no model supplies.
Choose to be a beginning, not a means. The machine is superb at being a means. Resolve, in Arendt’s sense, to be an actor: someone who starts things and stands behind them. A career spent as a high-grade instrument is a career spent racing better instruments. A life spent as an author of outcomes is one the machine makes more powerful, not less.
The anxiety of this moment is not really about income. It is about worth, the fear that if a machine can do what I do, then what I am is redundant. The answer is the essay’s thesis turned to face the person rather than the labor market. You were never your function. The organization paid you for the function because that was the part it could measure, but it was always the smallest part of you, and it is now being handed to machines so the rest of you has room to work: the judgment, the responsibility, the care, the meaning that the daily grind kept crowding out.
This is not guaranteed and it is not equally available, and the earlier sections said so. But for those who can take it, the offer is not a smaller future than their parents had. It is a far larger one. The machines will do the labor. The action, the beginning of new things and the standing behind them, remains ours, and there is more of it to do now than there has ever been. The only people the change truly threatens are those who insist on staying a function. The invitation to everyone else is to stop being employed by the world and to start e-mploying it.
Tebaa, J. (2026). The E-mployee: Reclaiming Human Value in an Age of Artificial Labor. https://doi.org/10.5281/zenodo.20492469
DOI: 10.5281/zenodo.20492469 · ORCID: 0009-0007-5046-3790
An e-mployee is an artificial agent, an AI system or software agent, that an individual directs like a member of a personal workforce: delegating objectives to it, supervising and reviewing its output, and remaining personally accountable for what it produces. The term was introduced by Dr. Jonah Tebaa, DBA.
The terms e-mployees and e-mployer were introduced by Dr. Jonah Tebaa, DBA, in the essay “The E-mployee: Reclaiming Human Value in an Age of Artificial Labor” (2026).
An e-mployer is a person who directs a workforce of e-mployees rather than only supplying labor themselves. The e-mployer sets objectives, delegates tasks to agents, reviews and corrects their output, and remains accountable for the result, the way a founder directs a team.
The E-mployee Doctrine is a framework of five capacities that distinguish a person who merely uses AI tools from one who genuinely directs an artificial workforce: delegation, direction, discernment, accountability, and orchestration.
Acemoglu, D., and Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30.
Arendt, H. (1958). The human condition. University of Chicago Press.
Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30.
Autor, D. H. (2024). Applying AI to rebuild middle class jobs (NBER Working Paper No. 32140). National Bureau of Economic Research.
Bessen, J. (2015). Learning by doing: The real connection between innovation, wages, and wealth. Yale University Press.
Brynjolfsson, E. (2022). The Turing trap: The promise and peril of human-like artificial intelligence. Daedalus, 151(2), 272–287.
Dweck, C. S. (2006). Mindset: The new psychology of success. Random House.
Epstein, D. (2019). Range: Why generalists triumph in a specialized world. Riverhead Books.
Frey, C. B., and Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254–280. (Original work circulated 2013)
Manyika, J., Lund, S., Chui, M., Bughin, J., Woetzel, J., Batra, P., Ko, R., and Sanghvi, S. (2017). Jobs lost, jobs gained: Workforce transitions in a time of automation. McKinsey Global Institute.
Polanyi, M. (1966). The tacit dimension. University of Chicago Press.
World Economic Forum. (2025). The future of jobs report 2025. World Economic Forum.