Anthropic's AI Jobs Paper Misses the Real Risk for Operators

Anthropic Published a Paper on AI and Jobs. Read It Like a Scout, Not a Soldier.

One disclosure before we start. I run a manpower services company. People are my primary cost line and my primary product. My financial interest in this question is the mirror image of Anthropic's. Read this knowing that. Apply the same test to me that I am about to apply to them.

Anthropic released a research paper this week claiming AI has not yet caused significant unemployment. The headline finding: no systematic rise in joblessness among workers in the most AI-exposed roles since ChatGPT launched in late 2022. If you run a business where people are your biggest cost line, you probably felt a small wave of relief reading that.

That relief is worth examining.

The paper is not neutral. No paper from this source could be.

Start with the contradiction you need to hold in your head while reading this paper. The same week it ran its "no systematic unemployment yet" headline, Anthropic's own CEO made a very different claim. Dario Amodei told Axios that AI could wipe out half of all entry-level white-collar jobs.[⁴] Unemployment could spike to 10 to 20% within one to five years. That is not a rounding difference. Two messages, two audiences. Research paper for procurement teams and regulators. CEO interview for everyone else.

That gap is the most important fact in this entire discussion.

Now the context behind it. Anthropic builds and sells AI. Their customers are companies. Those companies carry two budgets, an IT budget and a labour budget. The numbers are not close. Across Fortune 500 companies, labour runs roughly 12% of revenue. IT sits at 2 to 4%.[¹] Take Indian IT services. TCS and Infosys run labour at 50 to 60% of revenue. IT sits at 3 to 5%.[²] In most industries, labour is 10 to 50 times larger than IT spend. Labour is the bigger prize by a large margin.

If Anthropic wants to grow, it is reasonable to suspect they benefit when companies redirect even a fraction of that labour spend toward AI. A paper that says "AI is causing mass unemployment" creates regulatory risk, public backlash, and procurement hesitation. A paper that says "no harm yet, but we are watching carefully" signals responsibility without admitting damage.

That may be the most commercially convenient conclusion possible. And it happens to be the one they reached.

CNN's review of Amodei's public statements, citing tech futurist Tracey Follows, characterised the messaging as "part truth-telling, part reputation management, part market positioning, and part policy influence."[³] The implication: by predicting mass unemployment, Anthropic cannot be blamed when it arrives. They already warned you.

This is not an accusation of fraud. The incentive structure makes this conclusion attractive to a vendor. That is worth knowing when you read their findings, not as proof of design, but as a reason to look carefully at what they chose to measure and what they chose not to.

The basic question every operator should ask before acting on any research: what does the author gain if this finding is true?

They measured the wrong thing and called it a result.

The paper tracks unemployment as its primary outcome. The logic appears clean: a displaced worker shows up in unemployment data. But your entry-level hire who was never made does not show up anywhere.

The data that matters is buried below the headline. Recent labour market studies drawing on Stanford and ADP payroll data find a roughly 13 to 14% decline in employment for workers aged 22 to 25 in the most AI-exposed occupations since late 2022.[⁵] Workers older than 25 saw no such drop. Anthropic's own paper echoes this suggestively, flagging a slowdown in hiring of young workers into exposed roles, but does not quantify it at this level and moves on.

I saw this in my own business. In January 2026, an employee left. I did not replace them immediately. My existing team absorbed the workload using the systems we had built over years. We managed the delays and kept the output stable. The headline unemployment number stayed flat. But for a period of time, that entry-level job simply ceased to exist.

That pattern, multiplied across thousands of companies, is what Anthropic's paper cannot see. A plausible interpretation: companies are using existing mid-level staff, augmented by AI, to absorb work that entry-level hires used to do. Unemployment stays flat. The headline stays clean. But the pipeline below mid-level has quietly stopped filling.

Those mid-level people will move up in three to five years. When they do, who has been trained to replace them at the level they are vacating? If this pattern holds, a plausible outcome is a capability gap around 2029 to 2031, not because unemployment spiked, but because the pipeline quietly stalled. Companies will have spent three years patching entry-level gaps with AI instead of people. The bill comes due when today's mid-level staff move up and nobody is ready to follow.

This is a scenario, not a forecast. But it is a scenario the paper has no mechanism to detect. It only measures current unemployment. The job never created is invisible to it.

They called a governance problem an adoption problem.

The paper identifies a large gap between what AI can theoretically do and what it is actually being used for. It lists the reasons: model limitations, legal constraints, human verification steps. Then it frames this gap as temporary friction, i.e., a measure of how far AI still has to travel before it reaches its potential.

Here is the problem with that framing. Legal constraints and human verification steps are not friction. They are rational behaviour in an environment where nobody has clearly answered who is liable when AI fails.

I learned this the hard way. I built an automated system to send clients a photo after every site visit, proving our guards were awake and on duty. One night at 3 AM, the system delivered a high-resolution photo of our failure: a guard snoring.

The client sent my manager the screenshot and asked: are you showing how bad the guard is, or how honest you are?

My supervisor had found the sleeping guard. He did not wake him. He just took the photo and hit Submit. My system rewarded compliance over judgment. He thought the mission was to finish the form. I saw the chat the next morning.

My system had turned a smart employee into a mindless tripod. The automation was perfect. That was the problem. We killed the real-time alerts immediately. Now we use a weekly batch report that a human reviews before it goes to the client.

If you are the human who signed off on deploying the system, you carry the full downside. Your rational response is to give AI only the lowest-risk tasks. Not because you doubt its capability. Because the liability is yours and the accountability chain is unclear.

My model: treat the adoption gap not as temporary friction but as a rational response to a liability vacuum. Those are different problems. Adoption lag resolves as technology improves. A liability vacuum requires governance infrastructure, i.e., clear law, clear accountability chains, clear insurance mechanisms, that does not yet exist and that no AI lab can build alone.

The paper reads governance hesitation as a technical barrier to overcome. It may actually be a signal that governance infrastructure has not caught up to capability. Calling them the same thing produces completely different solutions.

The historical argument has a structural flaw.

The standard rebuttal to any AI displacement concern goes like this: every major wave of automation destroyed jobs and created new ones. The loom replaced weavers and built the textile industry. The calculator replaced arithmetic clerks and created financial analysis. People moved up the value chain.

This argument has always held because the escape route was consistent. Humans moved into roles requiring judgment, creativity, and reasoning that machines could not replicate.

The historical narrative may not hold this time. Technology used to automate lower-skilled jobs. Displaced workers retrained upward. The assumption was that judgment and reasoning were safe. AI is targeting judgment and reasoning directly. Workers who spent years training into specialised roles may not find an equivalent rung above.[⁷]

The paper's own data shows computer programmers at roughly 75% task coverage.[⁶] Three quarters of their core tasks can theoretically be sped up significantly by an AI tool. Programming was supposed to be the safe rung. The rung people moved to after automation took the layer below.

If we are going to borrow comfort from history, I want to see a concrete story for what a 28-year-old displaced programmer does in 2030 that AI cannot. The paper does not provide it. It measures exposure. It does not catalogue escape routes.

What this paper needed to answer. And did not.

Two questions. Neither is answered.

First: what are the new jobs? Name them. Not in principle. Specifically. Until that question has a concrete answer, the historical reassurance is a feeling dressed as an argument.

Second: who is liable when AI fails? Not in theory. In practice, in your industry, under current law. If the honest answer is "the human who deployed it," then your rational AI adoption ceiling is not set by capability. It is set by how much accountability you are willing to absorb. Calling that a temporary barrier misreads the situation entirely.

The paper is not useless. The exposure metrics are genuinely interesting. The methodology is transparent enough to critique, which is more than most industry research offers. But it answers the question "has AI hurt employment yet" while leaving untouched the question that actually matters for your business: are you building on solid ground, or are you building on a lava floor?

Anthropic is mapping the lava floor as it exists today. You are building on it. Those are different problems.

Think of AI capabilities as a lava floor. Constantly shifting. Can break anywhere, at any moment. It looks solid until it does not. And unlike a human employee who resigns with notice, the lava floor gives none.

We have already seen this move. I built a custom compliance summary report using Claude Haiku 4. It was poor initially. I had to feed it highly specific examples to make it work. When Claude Haiku 4.5 launched, the model was smarter. It no longer needed my old examples. But it started making entirely new errors.

I had to spend hours reading the new output, diagnosing the failures, and rewriting the prompts from scratch. No warning. No notice. Same API, different floor.[⁸]

This is the real decision Anthropic's paper cannot make for you. Not whether AI is dangerous. But whether you know what you are standing on, and whether the part of your business that matters most is sitting on the lava or above it.

The four questions the paper never thought to ask. But you need to.

One: What is the real business underneath the business you think you are running?

Most operators answer this with their invoice. Manpower supply. Security services. Logistics. That is what they bill for. It is not what their best clients are actually paying for.

A useful test: if AI could do everything your people currently do, every task, every process, what would your client still need from you that they could not get from a machine? That answer is your actual business. Everything else is a delivery mechanism.

The way I found my own answer was simple. I imagined a client call going badly. Not a smooth one. A furious client, a compliance cost increase, a meeting where the room was tense and the numbers were unwelcome. I asked: what part of that moment could not be handed to a machine? That is how I found it.

I use AI to draft contracts, write code, generate policy documents. But when a client is furious, I cannot email them a formula. I have to sit across from them, read the room, and explain the risk in their language. That real-time human friction is what the client is paying for. Not because the output is wrong. Anyone with the same API access could generate it. That is the trust problem.

Find that thing in your own business. It is the only part that sits above the lava.

Two: What would an outsider do, and how do you get one?

Andy Grove and Gordon Moore built Intel on memory chips. When the memory business started dying, Grove asked Moore a sharp question: if the board brought in a new CEO tomorrow, what would they do? Moore said immediately: get out of memory. Grove said: then why don't we walk out that door, come back in, and do it ourselves?[⁹]

The outsider's advantage is not superior intelligence. It is no emotional attachment to the existing answer.

AI can play this role, but only if you force it to. Open a fresh conversation. Provide full context. Then give it one instruction: treat everything as sunk cost. No loyalty to any existing decision. Tell me what a new owner walking in tomorrow would change first.

Do not ask it to evaluate your current direction. Ask it to ignore your current direction entirely.

The discipline this requires is the meta-prompt. Without it, you will unconsciously feed AI the context that justifies decisions you have already made. You will end up with a more sophisticated version of confirmation bias dressed as independent analysis. The question that exposes this: does my critical thinking have an agenda? Apply it to your AI usage first. Then apply it to the research you are reading.

Three: Are you making today's decisions as the person you are, or the person you are building toward?

The pipeline collapse argument cuts both ways. Companies that stopped hiring entry-level workers in 2023 are making a 2029 decision right now. They just do not know it yet. You are doing the same thing every time you choose between retraining an existing employee and replacing them with an AI agent.

To illustrate, these numbers are not from my specific operations. Treat them as a thinking frame rather than a benchmark. Suppose an employee costs Rs 30,000 per month. An AI agent handling similar tasks costs Rs 7,000. The reactive calculation looks obvious: replace the employee, save Rs 23,000.

The evolutionary calculation looks different. That same employee, retrained to supervise AI output rather than produce it, can absorb work that previously required three people. Their pay goes up because the savings are real. Their incentive now points in the same direction as yours, i.e., make the AI perform well, not fail quietly.

You hire a new person only when the work genuinely requires more people than your current team with AI can handle. The numbers in your business will be different. The ratio of people to AI will be different. The cost of retraining is real. These are questions you have to answer with your actual figures and your actual error rates. The point is not the calculation. The point is which question you were solving for when you made the decision. Reactive: protect this month's margin. Evolutionary: build the capability that is defensible in 2029.

Four: What is your reliability workflow, and what are its real incentives?

AI is not a new employee. The analogy is useful until it breaks. It breaks in three specific places.

When a human employee makes an error, someone usually notices because it looks wrong to a trained eye. When you explain the mistake, the employee builds a mental model that prevents the same error next time. When an error compounds over weeks, there is a human decision trail you can reconstruct.

Current AI systems fail on all three. The error arrives with total confidence so it does not look wrong. The explanation disappears at the end of the session because the next conversation starts fresh. A compounding error can run for weeks before anyone catches it. Nobody was watching for the pattern.

Small errors compound. A tiny mistake at the start of a long calculation produces a massive wrong answer at the end. Past decisions constrain current options even when the original reason is gone. Humans used to catch these quietly through instinct built over years in the same role. AI removes that instinct from the process.

So your reliability workflow is not just "check AI's work." It is: who checks it, using what trained judgment, at what point in the process, with what authority to stop and override.

At Knighthood, the first role I looked at was the person who prepares client-facing compliance summaries. That output now goes through a human review before it leaves the building. Not because I distrust the tool. Because I identified that this is the role where an error would reach a client before anyone caught it. Starting there was the minimum viable reliability step.

What I have learned is that the incentive question matters as much as the process question. The person best positioned to catch AI errors in a given role is usually the person whose job AI is replacing. Their incentive to build a great reliability workflow is not obvious. Solve that first. Pay them more when AI performs well. Make their growth contingent on the system's reliability. Make what is good for them the same as what is good for you.

The liability question and the incentive question are the same question asked from different directions.

The practical rule I now apply before automating any client-facing process: assume the AI will silently compound an error for six weeks. Identify the exact human who will sit across from the furious client to explain why it happened. If you cannot name that person, or if that person does not have authority to override the AI today, do not automate the workflow. Insert a human review step instead.

What to do with this as an operator.

Read Anthropic's paper. The exposure metrics are useful. Knowing which roles sit at high task coverage tells you where to look first in your own business.

But read it the way a scout reads a map drawn by someone who owns the territory.

The paper tells you where AI is being used today. It cannot tell you whether the disruption in your specific business is over or just quiet. Quiet and over look identical until the moment they do not.

The real decision is not whether AI replaces your people. It is whether you control the sequence or react to someone else's. A business that waits for certainty will find the decision has already been made for it. A business that moves without answering the liability question first will find the decision has been made badly.

It was never a yes or no question. It was always what order.

References

[1] Fortune 500 labour vs IT spend ratios https://financialliteracyclub.org/the-warrant/the-fortune-behind-the-500-labor-and-revenue

[2] Indian IT services labour costs, TCS and Infosys FY24 https://hr.economictimes.indiatimes.com/news/industry/it-giants-skyrocket-revenue-but-soaring-employee-costs-spell-trouble/117778632

[3] Tracey Follows, tech futurist and Futuremade CEO, cited in CNN Business review of Amodei's public statements https://www.cnn.com/2025/05/29/tech/ai-anthropic-ceo-dario-amodei-unemployment

[4] Dario Amodei warning on 50% entry-level job loss and 10 to 20% unemployment https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic

[5] 13 to 14% employment decline for workers aged 22 to 25 in AI-exposed roles, Stanford and ADP study https://www.cnbc.com/2025/08/28/generative-ai-reshapes-us-job-market-stanford-study-shows-entry-level-young-workers.html

[6] Computer programmers at roughly 75% AI task coverage, Anthropic Labour Market Impacts paper https://www.anthropic.com/research/labor-market-impacts

[7] This is the author's inference from the structure of the problem. No study yet quantifies the retraining gap for specialised white-collar roles at scale. CNN Business reporting on Amodei's statements raises the same concern. https://www.cnn.com/2025/05/29/tech/ai-anthropic-ceo-dario-amodei-unemployment

[8] Author's direct experience with model version update affecting automated workflow output quality, 2025

[9] Grove and Moore, Intel memory business exit decision, Andy Grove, Only the Paranoid Survive, 1996