Amplify, Don't Replace: Why the AI Layoff Trap Is a Strategy Mistake, Not an Inevitability
- Nikola Njegovan
- Apr 30
- 11 min read
The trap, in plain language
In March 2026, two economists, Brett Hemenway Falk of the University of Pennsylvania and Gerry Tsoukalas of Boston University, published a paper titled The AI Layoff Trap. It is the most rigorous economic argument I have read this year, and every executive thinking about AI strategy should sit with it.
Their core finding is uncomfortable: when firms automate jobs faster than the economy can create new ones, they erode consumer demand. Workers are also customers. The wages a company "saves" by replacing people with AI come out of the spending pool that all companies (including the one doing the cutting) eventually depend on.
The trap is that no individual firm can stop. Even when leaders see the cliff clearly, competitive pressure forces them to keep accelerating. Falk and Tsoukalas show that this is a prisoner's dilemma at industrial scale. The only intervention their model finds effective is a Pigouvian tax on automated tasks which is a structural fix, not a voluntary one.
That part is for policymakers. What I want to talk about is the part for operators.
What the layoffs are actually saying
The receipts are no longer abstract:
Block cut nearly half of its 10,000-person workforce in February 2026, with leadership attributing the cuts to AI.
Salesforce replaced approximately 4,000 customer support agents with agentic AI systems.
Over 92,000 tech workers have been laid off across 98 companies in 2026 alone, following more than 100,000 in 2025, with AI cited as the primary driver in over half of cases.
These are not careless companies. They are well-capitalized, sophisticated, and run by people with access to better data than most. So why does the strategy still feel wrong?
Because every one of these decisions rests on an assumption that is too coarse:
That AI replaces function.
It does, but the trouble is the word "function" is being made to do too much work. To get the strategy right, we have to be more precise about what AI is actually good and not good at.
Three categories, not two
Most of the public conversation about AI and work splits the world into two buckets: deterministic work that AI handles, and non-deterministic work that humans handle. That framing is wrong, and it is wrong in a way that produces bad decisions.
The honest map has three categories.
1. Deterministic work
This is rules-based, repeatable, and bounded. Given the same inputs, the same outputs are correct. Invoice processing, ticket routing, document classification, structured data extraction, standard reporting cycles. Automation has been eating this work for forty years, and AI accelerates the trend. If your team is doing deterministic work by hand in 2026, that is a problem worth solving and you should solve it without sentimentality.
2. Non-deterministic work that AI is genuinely good at
This is where the public debate goes off the rails, because most arguments against AI-driven layoffs pretend this category does not exist. It does. AI is meaningfully strong at:
Synthesizing across large bodies of material. Reading a hundred research papers, a thousand support tickets, or a year of Slack history and surfacing the patterns. Faster than a human, and often more comprehensively.
Generating drafts, options, and variants. First-pass copy, alternative approaches, exploratory designs, candidate plans. The blank page is no longer a bottleneck.
Pattern-matching across precedent. "Have we seen something like this before?" is now a question with a useful answer in seconds, not days.
Surfacing non-obvious connections. Across disciplines, across data sets, across time. This is the work that used to require senior generalists who had been everywhere; AI does a real share of it now.
First-pass analysis and classification. Reading the inbox, triaging the tickets, summarizing the call, flagging the anomaly. Not perfectly. But well enough that a human can supervise rather than perform.
Stress-testing arguments and exploring "what-ifs". Before a human commits, AI can pressure-test the reasoning, surface counterarguments, and walk through scenarios.
This is real, valuable, exploratory, non-deterministic work. Operators who refuse to use AI here are leaving meaningful productivity on the table, and any pitch that pretends otherwise will not survive contact with a CFO who has tried it.
3. Non-deterministic work that AI is genuinely bad at
This is the category that gets collapsed into the previous one and when that collapse happens at scale, the result is the layoff cycle we are watching. AI is meaningfully weak at:
Owning a decision when there is real downside risk. A model has no persistent identity, no career, no relationships to protect, and no ability to be held to account when something goes wrong. Accountability is a feature of humans, not a feature you can bolt onto an LLM.
Truly novel situations. AI is essentially a high-resolution rearview mirror. When the situation is genuinely new, no precedent, no analogous data, its outputs become confident pattern-matches that can be plausibly wrong in dangerous ways.
Reading the room. Political, cultural, emotional context. The customer who is technically asking one thing but actually means another. The team member whose silence in the meeting is the most important data in the room.
Knowing when it is confidently wrong. This is a meta-skill humans develop through experience and consequence. Models do not have it; they generate fluent text whether they are right or wrong, and the fluency itself is the failure mode.
Carrying long-term trust. The customer relationship that took eight years to build. The supplier who picks up the phone at 11 p.m. because of who is on the other end. The regulator who gives you the benefit of the doubt because of your track record.
Mentoring, developing, and judging humans. Performance, growth, fit, culture. The work that builds the next generation of operators in your business.
Final judgment in high-consequence decisions. Pricing strategy, hiring, M&A, crisis response, ethical calls. AI can inform every one of these and should. AI cannot make any of them.
This is also non-deterministic work. It is, in fact, the non-deterministic work that defines whether your company is a good business or just a busy one. And it is the work that gets ceded when leaders treat all "judgment work" as a single bucket.
The accountability disclaimer — a small symptom of a large problem
Before we go further, I want to flag a pattern I have started to see everywhere, because it is the clearest possible illustration of what category-collapse looks like when it shows up in a single email.
You have probably seen it yourself. A colleague sends a document, a proposal, a market summary, a candidate list, an analysis and somewhere in the cover note is some version of:
"FYI, AI helped me put this together — take it with a grain of salt."
"I had ChatGPT draft this, so let me know if anything looks off."
"This was AI-generated, just flagging in case there are issues."
I want to be careful here, because this is almost never malicious or lazy. The people writing those notes are usually thoughtful, well-intentioned, and trying to be transparent. They are new to working with AI, they don't yet have a clean mental model for what they should be checking, and the disclaimer feels like the honest move. In a sense, it is.
But look closely at what is actually happening in that sentence.
The sender has produced a piece of work and is, very politely, declining to be accountable for it. The accountability is being handed back to the reader: if there's something wrong here, that's something for you to catch, not me, the model wrote it. The work is going out into the world without a human standing fully behind it.
Multiply that pattern across a thousand emails a day at a large company and you can feel the strategic problem in your bones. Outputs are flowing. Volume is up. But the chain of accountability has weakened at every node, because every individual node is making a small, polite, completely understandable disclaimer. Nobody is being reckless. The system is becoming reckless without anyone meaning for it to.
This is the failure mode that scales when companies treat category 2 and category 3 as the same. Category 2 work, the synthesis, the draft, the first pass, was correctly delegated to AI. But the accountability for what gets sent, signed, decided, or shipped is category 3 work, and it cannot be delegated. The disclaimer is the moment the handoff fails.
The fix is not to ban AI assistance, and it is not to police people's email signatures. The fix is a different operating standard, taught and reinforced inside the company:
If you put your name on it, you own it. AI helping you produce it does not change that. If you are not in a position to vouch for what you're sending, then the work isn't done yet and you should go finish it.
That standard does something important. It forces the human back into the loop at exactly the place where their judgment is required. It changes the question from "did the AI write this correctly?" to "do I, as a professional, stand behind what is going out under my name?" Those are different questions, and they produce different work.
This also raises the bar on the humans you employ. To stand behind AI-assisted work, your people have to be sharper, more attuned to the material, more skeptical of fluent-sounding output, and more deliberate about what they are tuning before it goes out the door. They have to be able to read a generated paragraph and know — quickly — whether it actually represents the company's position, the customer's situation, or the truth of the matter. This is a real skill. It is teachable. It is also exactly the kind of human capability that compounds, and exactly the kind that gets eroded if you cut the people who would have developed it.
The accountability disclaimer is small. The pattern it represents is enormous. Every leader thinking about AI strategy should be watching for it, and every leader using AI personally should be holding themselves to the same standard they want from their team.
The strategic mistake hiding inside the layoff cycle
Once you have the three-category map in your head, what is happening in the headlines becomes clearer.
The companies cutting hardest are not making a category-1 mistake. They have correctly identified that some deterministic work can now be automated more aggressively than before. Good.
They are making a category collapse mistake, treating category 2 and category 3 as if they were the same. Because AI is genuinely useful in category 2, the assumption gets extended that it is also sufficient for category 3. It is not. And by the time the consequences show up in eroded customer trust, brittle decision-making, novel situations handled badly, and a thinning bench of humans who would have caught the problem, the institutional knowledge that would have prevented it has already been laid off.
When a company hands category-3 work to AI without humans accountable for it, it gets confident, scalable mistakes at customer-experience scale, at operational scale, at reputational scale. And those mistakes compound, because the model is sure of itself and the humans who would have caught them are no longer in the room.
The amplification thesis
At ODNOS, the way we think about this is straightforward.
You have two strategies available to you. They look similar from the outside. They diverge sharply over five years.
Strategy A: Replace. Identify roles that AI can plausibly cover. Cut them. Bank the savings. Hope that the remaining humans can manage what's left, that the customer doesn't notice, and that the model holds up as the operating environment shifts.
Strategy B: Amplify. Take your existing team, the people who carry your context, your standards, your customer relationships, your hard-won institutional knowledge, and equip each of them with AI tools that take the deterministic load off their plate and serve as a tireless category-2 partner: synthesizing, drafting, surfacing, exploring. Free their attention for the category-3 work that actually defines the business. Watch their output expand without diluting their quality, and watch their judgment get sharper over time because they're spending more of their day on the work that develops judgment.
The companies running Strategy A are betting that the model's ceiling is high enough to support their business. As the operating context shifts, and it will, they have no humans left to adapt the system. They are locked in to whatever the vendor ships. Welcome to being led against your will.
The companies running Strategy B are compounding. Their humans get sharper because they're spending time on the work that develops sharpness. Their systems get smarter because thoughtful operators are continuously improving them. Their judgment quality rises while their cost-per-output falls.
One of these is a strategy. The other is a cost-cutting exercise wearing strategy's clothes.
What this means for how you operate
If you are a founder or CEO making AI decisions this quarter, here is the frame I would offer:
Map your work into the three categories. Be specific. Most companies underestimate how much of their value sits in category 3, and overestimate how much of category 3 can be safely delegated to a model.
Automate the deterministic work aggressively. Without sentimentality. The goal is to remove drudgery from your people, not to remove your people.
Deploy AI as a partner in category 2 and measure it. Drafting, synthesis, exploration, pattern-matching, first-pass analysis. Equip every knowledge worker with these tools. Track time saved and quality lifted. This is where the real productivity gain lives, and most companies are leaving it on the table while they argue about layoffs.
Set an accountability standard and teach it. "If your name is on it, you own it." No disclaimers, no grain-of-salt caveats, no handoffs of responsibility to the model. Train your people to read AI output critically, tune it deliberately, and stand behind what they send. This is a real skill, and the companies that build it will move faster and better than the ones that don't.
Around category 3, build amplification systems, not replacement systems. AI as copilot, second opinion, draft generator, scenario explorer. Always with a human accountable for the decision. Always with a clear escalation path when the model encounters something outside its training patterns.
Invest in the humans who do the category-3 work. Pay them well. Train them. Give them better tools. They are the asset that compounds. Cutting them is the modern equivalent of selling the printing press to save on ink.
Be specific about where AI belongs in your systems. Generic "AI transformation" initiatives produce generic results. Surgical deployment, designed by operators who understand both the work and the technology, is what produces the real returns.
The bigger picture
Falk and Tsoukalas's paper is, at heart, a warning that markets cannot solve this problem on their own. They are right about that. The competitive structure they describe is real, and a Pigouvian tax may eventually be part of the answer.
But individual companies still get to choose how they show up inside that structure. You can be part of the wave that hollows out the economy on the way down and discovers, three years from now, that your customer base has thinned, your remaining team is brittle, and a culture of "the AI did it" disclaimers has eroded the credibility you spent decades building. Or you can be part of the wave that proves there is a better way to build a business in this era: with sharper people, clearer judgment, real accountability for what goes out the door, and AI applied where it genuinely belongs, including the parts of judgment work where it is genuinely useful.
The first path is short-sighted. The second is the actual work.
At ODNOS, we know which side we are on. We help companies amplify their people, their teams, and their operations to build greater value, not strip them down in the hope that a model can carry the weight. The future belongs to the operators who are thoughtful enough to know the difference between the AI work that compounds human judgment and the AI work that replaces it, and disciplined enough to keep humans accountable for what gets shipped.
If you are weighing this decision in your own company right now, I would invite you to reach out. The most consequential AI strategy decisions of the next decade are happening in boardrooms this quarter. Make the one you will still respect in five years.
