№ 028Deep DiveMay 29 202613 min read

AIAI Models

When 95% Is Zero

AI is a 95% solution. In high-stakes work, that is not enough.

By Akram Ayyash

The closer of my last piece was that the two most important things to watch out for are always inflation and unemployment, and that AI will disrupt both. That piece covered inflation. This one covers unemployment.

The consensus view is that AI is about to drive a mass-displacement event in white-collar work. The Anthropic CEO has put a number on it. Major investment banks have put numbers on it. The IMF has put a number on it. The headline-level Challenger reports point the same way. Put end to end, the forecasts run from a few million jobs to a fifth of the global workforce.

I do not think that is what happens. AI is real, the productivity gains are real, and there will be real layoffs at the margin. But the bloodbath forecast assumes AI is going to replace human knowledge workers in droves. My thesis is that it won't.

The Forecast, and the Business Model Behind It

Anthropic CEO Dario Amodei said last May that AI could eliminate half of all entry-level white-collar jobs in five years and push US unemployment to 10–20%. Goldman Sachs estimates 300 million jobs globally are exposed to automation and that 25% of US work hours are automatable. The IMF puts global employment exposure at 40%, and 60% in advanced economies. Challenger Gray's April 2026 report shows AI-attributed cuts at 49,135 year-to-date, around 16% of all 2026 cut plans. In all of 2025, AI was cited in 5% of cut announcements. Now it leads the list.

The numbers, and the anxiety attached to them, could be real. That is not the point I am making. The point I am making is who is loudest about it. The 50%-of-jobs forecasts originated from the companies selling the picks and shovels. The narrative that humans had become useless was the marketing campaign that funded the buildout. It is worth noticing that, as Anthropic and OpenAI now position for IPOs, Amodei has begun walking the forecast back, pivoting to the Jevons Paradox: automation expands demand for the thing it touches rather than contracts it. Sam Altman has done the same dance. The doomers have become the optimists, on a timeline that maps cleanly to their funding round.

That is the first signal. The companies most exposed to the labor-displacement narrative do not believe their own forecast strongly enough to keep saying it.

The 95-vs-99 Gap

I have spent my career inside high-stakes, mission-critical financial environments. BNP Paribas trading credit and derivatives. Citadel funding the credit business across bonds, stocks, and ETFs. And now running treasury at Ava Labs. The throughline of those roles is that the software you ship has to work. Not 95% has-to-work. Has to work. Rain or shine. 100% of the time. Otherwise, it's useless. Consistency and reliability is why we pay for software solutions.

The gap between purpose-built software and vibe-coded AI-generated tools is small but very meaningful. The difference between 95% and 99% is game-changing in high-stakes corporate environments. In high-performing organizations, 95% is as good as zero.

That is why AI is deceiving. It looks amazing on the surface because it is 95% of the way there. You can demo it. You can put it in a deck. The first hour with any frontier model is dazzling. But for high-stakes, important, and mission-critical applications, it is not there yet. The last few percent is everything: the edge cases, the integration, the auditability, the accountability when something breaks at 3am.

Box CEO Aaron Levie makes the same observation in different language: "When AI completes the first 80% of a job, the extra 20% is all the value creation of that profession, with all the expertise and domain knowledge in that last 20%." Levie also has the sharper version of the point. CEOs, he wrote, are "uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work that still has to happen to generate most value with AI."

The CEOs writing 50%-displacement forecasts and the operators trying to ship production AI systems are looking at the same technology and reaching opposite conclusions. The operators have the better view.

The 95% That Never Ships

If you want to see the gap quantified, look at how much enterprise AI actually reaches production.

Start with the pilots. MIT's 2025 State of AI in Business found that 95% of generative AI pilots fail to scale to production. RAND found that more than 80% of AI projects fail to reach meaningful production, roughly twice the failure rate of IT projects without AI. IDC's figure is starker: 88% of AI pilots never graduate from proof-of-concept to production. Klarna walked back its all-AI customer service push in mid-2025 after customer satisfaction dropped, rehiring human staff. The company that, 18 months earlier, claimed its AI assistant was doing the work of 700 agents now operates a human-AI hybrid model. The headline was the easy part. The last mile was the rest of the story.

Uber's own COO admitted the same thing on the Q1 2026 earnings call. AI costs, he said, are getting "harder to justify" because there is no clean line between AI spend and meaningful new features. "That link is not there yet, right? I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and 'Okay, we're actually producing 25% more useful consumer features.'" This is one of the largest, most operationally sophisticated technology companies in the world. If anyone should be able to draw the AI-to-output line at scale, it is Uber. They cannot.

The 5% gap, Levie's 20%, and Uber's missing link are three formulations of the same observation. AI does the easy part well. It does the hard part badly, or not at all.

Compute Is Not Getting Cheaper

The cost-vs-employee math behind the displacement forecast assumes AI gets cheaper. It is not getting cheaper.

OpenAI raised enterprise pricing by roughly 120% year-over-year on aggregated customer data. GPT-5.5, released April 24, 2026, is priced at $5 per million input tokens and $30 per million output tokens, roughly twice the rate of GPT-5.4 at comparable usage. The Pro variant runs $30/$180 per million. Anthropic's Claude Opus 4.7, released April 16, kept the headline rate of $5/$25 per million but moved to a new tokenizer that uses up to 35% more tokens for the same text, which works out to an effective cost increase of 12-27%. Both companies have moved to lock enterprise customers at the new prices rather than the introductory discounts.

This is the post-subsidy regime. Investor capital paid for the first three years of below-cost inference. The bill is now in the customer's hands.

I remember when I used to Uber all over New York City for $6. That was the subsidized pricing Uber had to live with to penetrate that market. Today, an Uber from Manhattan to LaGuardia can easily exceed $100. The same thing is happening with AI. These model companies have yet to turn a profit, and once they have penetrated the market sufficiently, they switch to unsubsidized rates.

The "AI replaces an employee at one-tenth the cost" math runs on the assumption that the cost line stays flat or falls. When the cost line is up 120% year-over-year and rising, the math gets more interesting. CFOs are noticing. Salesforce's CIO research shows enterprises now spending four times more on data infrastructure than on AI itself, and 98% of CIOs report board pressure to prove ROI, with more than 60% unable to tie current deployments to hard commercial value.

The pricing trajectory is not a permanent feature. Hardware costs may eventually fall again. But for now, the cost-vs-employee delta is tightening, not widening.

The Probability Problem

The deeper question is whether the gap can ever close.

AI is a probabilistic tool, not a deterministic one. It generates outputs by statistical sampling, not by following a logical rule that always produces the same answer to the same input. That is what makes it powerful at writing, summarization, coding assistance, and pattern recognition. It is also what makes it structurally unfit for the last 5% in mission-critical environments.

Mission-critical software runs on determinism. A trade either books or it does not. A wire either settles or it does not. A regulatory filing either passes validation or it does not. The probabilistic foundation of LLM-based AI cuts directly against this requirement. It is not a question of more training data or more compute. It is a question of paradigm.

Gary Marcus has been making this point for years. Yann LeCun, Meta's chief AI scientist, said in May that LLMs are pattern-matching systems "without genuine comprehension" and predicted they would be "largely obsolete within five years, except for narrower purposes." Even Anthropic's own engineering disclosures suggest the company has moved to wrap probabilistic models inside deterministic, symbolic guardrails for the highest-stakes use cases.

The serious engineering view is that the path from 95% to 99% may require a different architecture entirely. That does not mean it will not be built. It means the timeline for closing the gap is not a function of next quarter's model release.

This is the structural reason the displacement forecast is wrong on the timeline. It is not just that the last 5% is hard to engineer. It is that the last 5% may not be reachable inside the current architecture.

The Data CEOs Will Not Hand Over

The other constraint sitting on top of the 95% problem is access.

To deploy AI on any non-trivial enterprise workload, the company has to feed it proprietary data: customer records, trade tickets, financials, legal documents, source code, employee data. That data is the firm's competitive advantage. Handing it to a third-party model provider is not a cost decision. It is a strategic decision.

The CIO survey data is consistent. Eighty-two percent of EMEA enterprises plan on-premise or edge AI workloads. Sixty-one percent of IT leaders say AI is increasing cybersecurity risk; only 31% are confident in their ability to address it. Salesforce's data, again: CIOs are spending four times more on data infrastructure than on AI, because the data-readiness problem is the actual problem.

Anyone who has sat through a bank or hedge fund vendor-onboarding process knows what this looks like. The legal, compliance, and information-security review is not a formality. It can take quarters. The trade-off is real: the productivity gain has to clear a high bar before the firm is willing to expose the data to clear it. Many CEOs will not make that trade for anything beyond well-bounded, low-risk workflows.

On-premise AI, open-weight models, fine-tuning on private data, and no-training contractual clauses are all real and growing. The barrier is friction, not a wall. But friction is enough to slow the adoption curve materially.

The Honest Counterweight

None of the above means nothing is happening.

Entry-level tech hiring has fallen 25–30% from its 2022 peak. Fresh-graduate hiring at the top-15 tech firms is down more than 50% over three years. The Dallas Fed found that workers aged 22–25 in the most AI-exposed occupations have seen a 13% decline in employment since 2022. Recent-graduate unemployment hit 5.6% in March 2026, one of the highest readings in a decade outside the pandemic spike. Customer service, paralegal, and junior copywriter are visibly thinner. The entry-level rung is genuinely being squeezed.

The mechanism, however, is not the bloodbath. It is what Uber confirmed on its earnings call: hiring slowdowns, not layoffs. The economy is substituting AI for new hires at the margin. It is not displacing existing workers en masse. Anthropic's own labor-market research, published while its CEO was forecasting 10-20% unemployment, found "no systematic increase in unemployment for highly exposed workers since late 2022." Anthropic did identify a 14% slowdown in hiring for the 22-25 cohort in exposed occupations. Slowdown, not displacement. The same picture from the doom company's own data.

Aggregate US unemployment in April 2026 was 4.3%, and labor force participation 61.8%, both little changed on the month. By the standards of any prior labor-market downturn driven by technological change, those are not crisis numbers.

The cohort that should worry is the one at the entry-level rung. The cohort that should not is the prime-age incumbent. The piece the headlines miss is that those are two different stories.

What This Means For Macro

If unemployment does not spike the way the consensus expects, the medium-term disinflation thesis from my last piece has to be re-examined.

The disinflation case ran through consumer behavior. Households watching the job-displacement headlines were supposed to tighten discretionary spending, drive down inflation, and force the Fed off its tightening bias. If the displacement turns out to be a slower-moving, entry-level-concentrated phenomenon, the disinflation it produces is shallower and slower than the bears expect. Consumer sentiment has fallen, but spending has held up better than sentiment would suggest. That gap may persist longer than the model thinks.

The Fed's job gets harder, not easier. Near-term inflation from the AI buildout stays sticky, the disinflation that was supposed to follow arrives slower, and headline unemployment stays low enough to keep the Fed comfortable while the entry-level damage builds underneath it, out of sight in the aggregate. Every signal the Fed reads is muddier, and the risk of a policy mistake rises in both directions.

The same muddiness sits under asset markets. The market keeps reaching for a clean resolution: a fast AI productivity boom, or a labor-driven slowdown that forces the Fed to cut. My read is that the data hands it neither. Crypto has spent 2026 grinding inside a tight range across BTC and ETH, which is what a market pricing "nothing resolves, nothing breaks" looks like.

Inflation and unemployment are the two forces that matter most, and the Fed sits at their intersection. My last piece argued the consensus was wrong on inflation. This one argues it is wrong on unemployment, in the same direction and for a related reason: the economy AI is actually reshaping is not the one in the Fed's data yet, and not the one the doomers have already priced.

In Short

Strip away the forecasts and the argument is simple. The people who put the scariest numbers on AI displacement, Anthropic and OpenAI, were also selling the product, and they have quietly walked those numbers back on a timeline that maps to their funding rounds.

AI does the first 95% of knowledge work well and the last 5% badly or not at all, and in high-stakes environments that last 5% is the entire job. The cost of closing the gap is rising, not falling, now that the labs have ended the subsidy era.

The probabilistic architecture underneath today's models may not be able to close it at all without a different paradigm. And the proprietary data required to deploy AI on serious workloads is the data most CEOs will not hand to a third party. That is not the recipe for a bloodbath. It is the recipe for a hiring slowdown, concentrated at the entry level, with prime-age incumbents largely untouched.

So watch three things.

The capability gap. Enterprise pilot-to-production conversion, shipped-product percentages, and whether more CFOs follow Uber in refusing to draw a line between AI spend and output. Uber said it first. Others will follow.

The cost line. Frontier-lab enterprise pricing and the slope of inference cost per token, now that the subsidy era has ended. The next move sets the regime.

The cohort split. Recent-graduate unemployment, the 22-25 employment ratio in exposed occupations, and JOLTS hires rather than separations. Aggregate unemployment will keep the Fed comfortable. The signal is the cohort underneath it.

AI will take some jobs. The entry-level rung is genuinely thinning, and that is a real problem for the people standing on it. But the mass displacement the consensus has priced does not arrive. Real change does, slowly, in the places the aggregate numbers will not show for a while.

It is going to be a long, strange decade for labor.

Drop me a line:

What do you think? Do you like this? Do you not like this? I would love to hear your thoughts, so please reach me at akram@span.blog

Enjoyed this? Subscribe for the weekly deep dive, or drop me a line at akram@span.blog.