Everyone had a reaction when Elon Musk cut half of Twitter.
He was a genius. He was reckless. The site would be offline by Thanksgiving. Twitter had been hiding thousands of useless jobs. Musk had destroyed the company in a week. For a while, the internet managed to hold every possible opinion at the same time, which may have been the strongest evidence that Twitter was still working.
But almost nobody stopped on the detail that should bother every corporate leader planning an AI transformation.
He started before ChatGPT.
Let that sink in.
Musk completed the acquisition on October twenty-seventh, twenty twenty-two. About half of Twitter's seven thousand five hundred employees were gone by the end of his first week. ChatGPT did not launch publicly until November thirtieth.
There was no coding agent rebuilding services while the engineers packed their laptops. No general-purpose model reconciling advertiser accounts, drafting policy, clearing support queues, or preparing a neat summary for the person who no longer had fourteen managers between them and a decision. The first fifty percent cut was not an AI workforce transformation. It was concentrated authority, severe financial pressure, and a willingness to absorb risk that most hired executives would not be allowed to propose in PowerPoint.
By April twenty twenty-three, Musk said Twitter employed about one thousand five hundred people, down from nearly eight thousand. That larger reduction did not happen in two weeks. It unfolded over roughly six months, and pretending otherwise turns an already extreme story into mythology. We do not need the mythology. Half the company was gone in a week, before the technology now dominating every workforce plan was available to the public.
Now do the quieter math. Start with eight thousand employees and reduce the workforce by fifteen percent each year. After one year, you have six thousand eight hundred. After five, about three thousand five hundred fifty. After ten, about one thousand five hundred seventy-five.
A conventional company can arrive at almost the same number. It will just take a decade, with today's AI, to make the operating change Twitter started without it.
I have a theory about why. Most hired leaders are not paid to absorb that much risk, and they are not rewarded for compressing ten years of improvement into one spectacular year. They are paid to produce another strong year next year.
The fifteen percent example is a model, not a recommendation. AI-created capacity can become lower cost, more output, faster growth, better service, or some combination of the four. You do not have to turn every hour saved into a person removed. The point is the size and pace of the operating change, not a quota for layoffs.
The technology may be ready today. The incentive system would like it to wait for the next bonus cycle.
The principal lever was not AI. It was control backed by ownership. Musk said the company was losing more than four million dollars a day, and he was willing to let the organization absorb an extraordinary amount of risk to stop it.
His incentive was also wonderfully uncomplicated. Musk led the ownership group that had just paid roughly forty-four billion dollars for Twitter. At the loss rate he described, waiting one year to make the hard changes meant watching about one point four six billion dollars disappear. He was not trying to earn a maximum annual bonus by beating an operating target five points at a time. He needed a return on the purchase, and every day of delay made that return less likely.
There was no reason for him to save part of the improvement for next year's plan. He owned the upside if the company became more valuable, and he owned a meaningful part of the loss if it did not. You can argue with what he changed, how he treated people, and how much risk he created. You cannot argue that the incentive was vague.
If you want to see what that control looked like, look at the Sacramento data center. Twitter's infrastructure team had planned an orderly move that would take months. Two days before Christmas, Musk went to the facility, started disconnecting server racks, and brought in moving trucks. More than seven hundred racks were moved within days. Walter Isaacson's account of how it happened is worth reading, including the part where Musk later acknowledged that the abrupt shutdown had been a mistake.
He did not optimize a data-center migration. He dropped a data center and accepted what broke. Twitter suffered outages, infrastructure engineers inherited the cleanup, and the business kept moving. That is not a recommendation for anyone responsible for customer data, production resilience, or employees hoping to spend Christmas Eve somewhere other than a server room. It is an example of an owner taking risk that a conventional executive would need six approvals to discuss.
That should make the comparison more uncomfortable, not less. X had fewer tools and one owner willing to move. Your company has better tools and a leadership team paid to make sure nobody moves quite that fast.
Most large companies now have tools that can remove work and multiply capacity in ways Twitter did not have during that first cut. An agent can draft the analysis, modify the code, generate the tests, reconcile the records, write the customer response, and prepare the decision for a human. It cannot safely replace every person involved, and anyone selling that fantasy should be required to operate the system they propose. But it can remove a meaningful amount of work that once justified roles, teams, vendors, and layers of coordination.
So the question is no longer whether a large organization can survive a dramatic operating-model change without modern AI. We watched one do it in public, however messily. The question is why organizations with much better automation will still choose to move much more slowly.
People lost jobs, income, healthcare, colleagues, and a sense of security with almost no time to prepare. The people who remained inherited the work, the operational risk, and the consequences of decisions they did not make. X also absorbed outages, advertiser flight, regulatory scrutiny, legal disputes, and reputational damage. Anyone presenting the story as a clean productivity case study has removed most of the case.
AI does not prove that eighty percent of your people are unnecessary. A social network surviving after a massive reduction does not prove that your bank, hospital, manufacturer, insurer, or software company can do the same thing safely. Running is not the same as healthy. Existing is not the same as creating durable value.
There is also an enormous difference between an owner and a professional executive.
Musk bought the company and became its sole director. He could make the decision, accept the consequences, and keep control if the decision became unpopular. His financial exposure was extraordinary, but so was his authority. He did not need to preserve a resume that another Fortune Five Hundred nominating committee would find reassuring.
Your divisional president does. Your chief technology officer may believe AI can remove three layers of coordination, combine four teams, retire half the project portfolio, and run the remaining work with a much smaller organization. They also know that if they are wrong, the outage has their name on it. The customer departures have their name on them. The board memo has their name in the first sentence. The people who approved the strategy will remember that they approved a carefully governed transformation while the chief technology officer apparently chose to fire half the company before lunch.
If they are right, the reward is stranger than it looks.
They get one exceptional year. Then the savings become the new baseline, the board resets the target, and somebody asks what they plan to do for an encore. If the change becomes a cautionary case study, who hires them next? If it works, does the next company hire the executive known for making most of their own organization unnecessary, or the one promising dependable improvement without frightening the board?
That is not cowardice. That is a rational actor reading the compensation plan and the labor market.
Imagine you run a business unit with five hundred million dollars in annual controllable cost. You use AI to redesign the work, remove waiting, automate routine production, reduce outside services, and eliminate roles through attrition and selective restructuring. The cost base falls ten percent.
You just created fifty million dollars in recurring benefit.
At fifteen percent, you created seventy-five million dollars. In most companies, that is not a timid result. That is the lead slide at the earnings call, the top line in the board packet, and the reason the compensation committee discovers adjectives it had been saving for a special occasion.
Public proxy statements make the game fairly plain. Annual executive awards are commonly tied to revenue, earnings, margin, cash flow, or operating targets. In one recent plan, reaching one hundred percent of the revenue and billings target funded a one hundred percent payout. Reaching one hundred five percent funded two hundred percent.
Five points of performance can double the annual bonus in a plan like that. Long-term equity complicates the picture, and every compensation contract is different, but nobody needs to remove eighty percent of the cost base to have a very good year.
Now suppose the same executive sees a credible path to two hundred million dollars in savings. They have three choices.
They can attempt the full change now, take the operational and human risk in one dose, and create a baseline they have to beat next year. They can capture fifty million dollars this year, another fifty million dollars next year, and continue producing record improvements through the useful life of their equity grants. Or they can propose the full change, watch three committees turn it into a twelve-quarter program, and receive roughly the same outcome with much better political cover.
Which choice does the system reward?
The executive is not trying to win one race by five laps. They are trying to collect enough points across the season to finish comfortably in the top half of their segment and trigger a great payout. Revenue, margin, cash flow, delivery, retention, and whatever strategic modifier the compensation committee added after the last offsite all put points on the board. A twelve percent improvement can produce a very good score without requiring the executive to bet their career on one turn.
Winning the race is awesome. It also creates the nineteen nineties Chicago Bulls problem.
The Bulls won six championships in eight seasons. After a while, winning was no longer extraordinary. It was the job description. Anything short of another championship looked like decline, because a dynasty teaches everyone to treat exceptional performance as the baseline.
The same thing happens when an executive produces forty percent in one year. The board does not preserve the old target and send a thank-you card. It builds the next plan on the new cost base, raises the expectation, and asks why the miracle cannot happen again. Most executives are not trying to become the nineteen nineties Bulls. They are trying to put enough points on the board to maximize the payout, remain in the top half of the peer group, and still have a believable plan for next season.
Boards say they want transformation. Compensation plans often ask for a dependable staircase. The executive who provides twelve percent this year and eleven percent next year is a disciplined operator. The executive who provides forty percent this year and cannot repeat it may be described as having benefited from a one-time action, right before the baseline is reset and the long-term plan discovers that last year's miracle is now the minimum.
Nobody puts heroically made next year's target impossible in the succession plan.
This is why I think many large organizations will distribute an AI-sized change across ten annual planning cycles. Each year will look successful. Each leader will be congratulated. Each target will move just enough to make the next round of savings valuable.
The company may even set records all the way to irrelevance.
Most executive incentive plans contain a clever asymmetry.
The upside is usually capped. A leader might earn one hundred fifty percent or two hundred percent of a target bonus for a remarkable year. The downside is not. A failed restructuring can cost the bonus, unvested equity, the job, the next job, and years of professional reputation.
If you are already on track for the maximum payout with a twelve percent improvement, taking enough additional risk to produce thirty-five percent is economically irrational unless you own a meaningful share of the upside. The company gets most of the benefit. You absorb a disproportionate share of the career risk.
The org chart adds another incentive that rarely appears in the formula. Executive status still tracks scope: budget, headcount, reporting layers, and the size of the thing you control. A leader with two thousand people is described differently from a leader with four hundred, even if they were the person smart enough to prove that four hundred could outperform two thousand. They can create enormous value for the company while making their own role look smaller to the next compensation consultant or recruiter.
We are asking executives to automate away part of their authority, lower the visible scale of their jobs, accept uncapped career risk, and celebrate a capped bonus. Then we commission a change-readiness survey when they move carefully.
This is the part missing from most AI transformation decks. The slide shows technical feasibility. It shows labor capacity, model cost, process automation, cycle-time improvement, and a large green number labeled "opportunity." It does not show the personal risk-adjusted return for the person being asked to approve the change.
The consultant calls the gap resistance.
The executive calls it Tuesday.
You can dislike that behavior, but moralizing about courage will not change it. If the board wants an owner-sized decision from a hired operator, the board has to create owner-sized upside, multi-year protection, and room to survive a responsible miss. Otherwise the rational move is to harvest the opportunity in annual pieces.
This is change management before the town hall, the training plan, and the slide with the cheerful arrows. Change management starts with deciding who carries the risk, who receives the gain, and what happens to each person if the theory is wrong.
Once you see the incentives, recent corporate cuts become easier to read. Not every workforce announcement with AI in the first paragraph was caused by AI. Sometimes the tools crossed a real threshold. Sometimes a company already correcting its size found a better operating model. Sometimes AI turned a difficult course correction into a story the market wanted to buy.
More than one of those things can be true.
Amazon makes the distinction wonderfully clear because it put a target on management itself. In September twenty twenty-four, Andy Jassy told each senior organization to increase its ratio of individual contributors to managers by at least fifteen percent. He described pre-meetings for the pre-meetings, long lines of managers reviewing decisions, and people closest to the work waiting for someone farther away to approve it.
AI did not create those layers. People built them over years, one reasonable promotion, reorganization, dotted line, and director at a time.
This is not a new disease with a fashionable diagnosis. BCG documented traditional-company leaders describing organizations with as many as fourteen layers a decade ago. The technology is new. The hierarchy is not.
In October twenty twenty-five, Amazon announced an overall reduction of roughly fourteen thousand corporate roles. The explanation included AI, but it also named the old problems directly: bureaucracy, too many layers, diluted ownership, and resources sitting away from the company's biggest bets. In January twenty twenty-six, Amazon announced another sixteen thousand affected roles while using the same language about layers, ownership, and bureaucracy.
That is thirty thousand roles across two rounds. AI may make the smaller organization more capable. It may accelerate decisions that would otherwise take years. It did not travel back in time and schedule the pre-meeting for the pre-meeting. Some of this is an AI operating model. Some of it is a very large company admitting that it became a very large company.
Box offers the useful counterargument. Its chief executive officer, Aaron Levie, has been publicly skeptical of the assumption that every AI gain should turn into a layoff announcement. Box's operating argument is that when AI lowers the cost of work, companies usually demand more of that work and reshape the workforce around the new capacity. That is how Box describes the change, and it is a very different theory from cutting the company in half.
Box may be right. Box may discover that some roles still disappear faster than new demand arrives. The useful part is that the company names the second half of the equation. Productivity is not only a cost event. It can become more products, more customers, better service, and whole categories of work the old economics could not support.
That is why AI is becoming both an accelerant and an alibi. It can make a smaller organization genuinely more capable. It can also provide modern language for an old-fashioned correction to overhiring, weak margins, duplicated work, and layers of people forwarding a decision to one another.
Most companies will still do neither version quickly. They will add AI to existing work, reshape roles carefully, remove a layer here, consolidate a vendor there, decline to refill some open positions, and ask every remaining team to find ten percent more capacity. Slow change does not produce a dramatic announcement. It produces ten annual operating plans.
You may not control the board's compensation philosophy. You do control whether the next ten years happen to you or produce something for you.
Start with a boundary that sounds cynical only if you have never spent a few years without one: you cannot care more than the executive team does.
That does not mean doing bad work, hiding an opportunity, or making your colleagues carry you. It means you stop donating founder-level anxiety to leaders who are being paid for a controlled twelve percent improvement. If adequate performance against the stated goals produces the maximum bonus, hit the goals, take the bonus, and go home with enough energy to remain a person. The company designed that game. You are not morally required to play a more punishing version in private.
Do not confuse self-destruction with commitment. You cannot privately out-transform a leadership team that has publicly chosen the staircase. If the chief executive officer accepts ten years, your unpaid weekends will not turn it into ten weeks. They will make one year look slightly better while teaching the system that your extra effort is free.
First, stop treating "learning AI" as learning a chat interface. Learn how work moves through the business. Find the approval queue, the manual reconciliation, the weekly status ritual, the brittle handoff, the outsourced report, and the three people translating data between systems that should already agree. Then learn enough AI, software, data, testing, security, and product judgment to remove one of those constraints safely.
The valuable skill is not prompting. It is turning machine capability into accepted business value without creating a larger mess downstream.
Second, keep your own economic record. If you redesign a claims workflow and remove two thousand hours of annual handling, write it down. If release lead time falls from eighteen days to six, record the baseline and the result. If an agent reduces a vendor engagement from six hundred thousand dollars to one hundred eighty thousand dollars, keep the numbers, the controls, and the names of the people who will confirm them.
Do not build a resume around tool names. Tools change. Build it around outcomes you can defend.
Third, negotiate for a share of the value before the value becomes normal. Ask how your scorecard changes if the work produces more than the original target. That may mean a larger bonus opportunity, a promotion, retention equity, a funded role leading the next stage, or explicit gain-sharing for the team. This conversation is not greedy. The company is already doing the same math. You are simply asking whether the people taking on the new work, learning the new system, and making the old structure unnecessary participate in the return.
Managers should do the same for their teams. If five people automate the work of twenty, do not reward them with the abandoned calendars of the other fifteen. Give them money, scope, time to learn, and a credible path to the next job inside the company. Otherwise your best AI operators will learn that every productivity gain is converted into more work at the same pay.
They will respond rationally too.
Finally, move toward the work that receives the gains. If your workflow saves ten million dollars, find the product, market, or operating problem that the company can now fund. Do not spend five years defending the queue your own automation is emptying.
Build portable judgment while you do it. Learn to measure model quality, design evals, map a value stream, automate a workflow, read a profit and loss, explain risk to legal, and tell whether generated work is correct. The safest career is no longer the one attached to a stable box on the org chart. It is the one that can walk into a new system, understand how value is created, and make that system materially better.
If you want a different job, let the current one pay for the education. Use real workflows, real constraints, and real consequences to learn the skills the next company will hire. Keep confidences and leave the proprietary material where it belongs. Take the judgment, the scar tissue, and the documented outcomes that belong on your resume.
If your company chooses the ten-year path, you should not finish the decade with the same skills and a commemorative mug.
The gradual strategy works beautifully until a competitor refuses to participate.
A startup does not have to protect ten years of executive targets, seven layers of management, a shared-service allocation model, or a workforce plan negotiated before the current models existed. It can design the company around today's cost and capability from the first employee. It can price against your future cost base while you are still celebrating this year's twelve percent improvement.
That startup may not need to beat the incumbent everywhere. It needs one customer segment, one product line, or one ugly workflow that the incumbent cannot serve without protecting its existing economics.
Large companies have an answer for this too. They can use the savings from gradual change to buy the challenger.
Return to the five hundred million dollars cost base. A ten percent improvement creates fifty million dollars a year. Three years of that benefit can support a meaningful acquisition, several minority investments, or a portfolio of partnerships. The executive gets credit for margin expansion and then gets credit for buying growth. The startup's investors get an exit. The board gets a strategy slide. Everyone receives something recognizable.
Sometimes that is a smart outcome. Buying proven capability can be faster and safer than recreating it. The people who built the new thing may deserve the capital and distribution an incumbent can provide.
Sometimes the incumbent buys the startup, hands it an enterprise architecture review, a procurement queue, and twelve mandatory steering committees, then wonders where the speed went.
Cash is a defense against disruption. It is not immunity from it.
The acquisition strategy also fails when the challenger becomes too valuable to buy, refuses to sell, or changes customer expectations faster than the incumbent can integrate anything. A company can use annual savings to purchase optionality. It cannot purchase back every year it spent protecting the old model.
A company that genuinely wants the faster path has to make the faster path rational. The executive would need a multi-year scorecard, meaningful participation in the value created, room to survive a responsible miss, and permission to reinvest savings before finance quietly absorbs them into next year's baseline. The people carrying the transition would need to share in the gain too.
That is change management before the town hall. It is compensation design.
But this is not a charge to today's leaders. They can read their own bonus plans. If the company announces a revolution while paying everyone to produce an escalator, the result is not mysterious.
The more useful charge is to the rest of us.
Learn the game while you are still playing from the middle of the field. Learn how the profit and loss works. Learn which risks get discussed in the boardroom and which ones get buried in a project plan. Learn how a fifty-million-dollar improvement changes an executive payout, how a failed restructuring changes a career, and why the person praising transformation in the town hall may still choose the slower operating plan.
Take your bonus. Build the skills. Keep the economic record. Move toward work where you can participate in the upside. Do not spend ten years donating extra anxiety to a company that has already decided to use all ten.
And do not become too comfortable judging the person in the CxO seat. The first time you become the CxO, your entire outlook will change.
Yesterday, five hundred people were colleagues, managers, and names you knew. Today, they are also a line in the operating plan that you signed. The outage is no longer something leadership will explain. It is your name in the board memo. The customer loss, regulatory call, missed quarter, and restructuring decision all arrive at your desk. Your upside is still defined in the compensation agreement. Your downside has become much easier to imagine.
The executive you thought was timid may still have been timid. They may also have understood something you could not see from three levels down. Once the decision and the consequence belong to you, courage and recklessness stop looking like opposites. They start looking like two possible descriptions of the same Tuesday.
That does not excuse bad leadership. It gives you time to prepare for leadership without lying to yourself about what the job will do to your incentives.
I may be wrong about this. It is a theory built from the way corporate targets, careers, and change programs behave when they meet a large pool of technically available savings. I would genuinely like to hear from people working their way up, first-time C-suite executives, and operators who have already crossed that line. Did your outlook change when the risk became yours? Did the compensation plan change what you were willing to do?
And if you are a board member or executive with a different view, I would love to learn where this theory is wrong. What incentive am I missing? What have you seen make it rational to deliver a decade of responsible change now?
AI may be able to compress ten years of work into one difficult year. Your company may still choose the decade.
You do not have to waste yours.