To better understand these changes, I spoke to over 20 current and former employees of Amazon, Atlassian, Block, Google, Meta, Oracle, and several startups.

The consensus is that this is not a solved problem. From a leader at a large technology company: “I think your expectation on how ‘AI-first’ we are is not yet aligned with reality. This is all evolving.”

It may be evolving but it is evolving rapidly. Extreme urgency is driving adoption of AI, experimentation in how to accelerate the work of individuals, how to make existing processes more efficient, and how to rewire work itself.

Being first to figure it out is existential - the difference between pulling ahead and being left behind.

Accelerating Individuals

Standard Rigs. Many companies are still at the bolt-on stage and have just given people lots of tools to play with. Block gives employees “literally every tool you could want” and is building out agents across many of those tools creating “overwhelming choice.” Another startup I spoke with is leaning into standardization, with the CTO aiming to get every single engineer to operate at the same level, executing the same way by creating a standard Claude Code “rig” that includes markdown files containing repository overviews, style guidelines, skills for repetitive tasks.

Slop Fatigue. One side effect of leaning into AI is the slop fatigue of senior staff who have to review the code and docs being fired at them by the cannon. A frustrated startup leader shared that most of the docs he’s getting from his team are AI generated (fine) but haven’t been iterated on to refine the content (not fine). The review experience is like reading a paragraph of text where letters have been transposed: when you skim it’s readable but under closer inspection it falls apart.

Making processes more efficient

Bottleneck shift aka The Slop Cannon Corollary. “The Theory of Constraints” focuses on identifying and removing bottlenecks in order to increase overall system output.

The bottleneck in product development used to be creating code: There were never enough engineers on a team to build everything the company wanted. Now with AI, everyone is a builder and the bottleneck has shifted from creating code to reviewing code before it goes into production. Companies are attacking this new bottleneck with very different approaches:

• YOLO. If a human sanity checks an AI code review and it looks good, it goes into production.

• Using a different AI to review the other AI. For example, if Claude was used to generate the code, then GPT is used to review the code. Rationale is that hopefully have different blind spots. A fine idea but feels a bit like hiring a horse to watch your dog.

• Risk stratified. Atlassian has two different processes: one for “infra” and one for “product.” On the infra side, the expectation is that every change is manually reviewed because the blast radius is large whereas on the product side, it’s closer to YOLO.

Automation of expertise. When I worked at Google, researchers with Computer Science PhD’s were responsible for tracking the state-of-the-art in machine learning, make high judgement decisions around what new techniques might improve model performance, and work with teams to run experiments.

Today at Meta, there is an agent pipeline that scrapes arXiv (website where academic papers are shared rapidly before peer review), identifies relevant papers, ranks them, writes new code in a sandbox, and tests new methods to see which have the most impact. Fully automated, scalable, and several times faster.

This automation of business processes will become more common outside of tech with companies hiring for “AI Transformation” roles and bringing in consultancies to help them retool (Financial Times).

Rewiring work with extreme urgency

There is a gnawing sense from executives to individual contributors that their orgs are too big and not organized correctly, driving extreme urgency for leaders to figure out how to rewire work. As one CRO put it, “Any company started more than two years ago is already behind.”

5x too many people. When speaking to people at Google and Amazon, they think orgs have roughly 5x more people than they need to keep the lights on. Obviously a company has to do more than keep the lights on to grow but that is an enormous amount of extra cost and complexity to carry vs. an AI-first competitor. Individuals are making alternative career plans, and proactively leaving. Leaders are trying to redesign the org.

New orgs - Smaller pods of builders (Google, Meta). One Google org is breaking itself up into full stack pods of about 30 people. Meta is on the same path, breaking down its 1,000 person Reality Labs group into “AI Pods” (Business Insider). The pendulum is swinging from functional orgs to much smaller product / business focused teams.

Same org - Far fewer people (Block). Block had grown considerably before it cut 40% of staff in February. It added headcount organically with the Cash App run as its own BU, inorganically through the acquisition of Afterpay, and across the board during COVID with management projecting a new normal rather than a one-time boost. The result was an organization that felt bloated. Too many people held too narrow scopes leading to “20 people being involved to get one thing done.” Notably there was and still is no real plan for how to organize: “I don’t know how we are structuring teams. I don’t know if anyone knows.” It appears Jack believed two things: AI would enable running the company more efficiently and that the cuts would create the conditions for that rewiring to occur. (More on that in the next section).

Observations

Very little structured thinking. Individuals and companies are doing a lot but I haven’t seen a clear articulation of how they are approaching AI adoption, rewiring the business, and measuring success. I plan to spend more time on this in a followup post.

Opportunities for a complete new ways of working. Early in March I shared an AI Chief of Staff that I’d developed (link). The core ideas were a) AI should not be a reactive tool the same as a spreadsheet or a word processor, but a proactive partner in work and b) AI should not be siloed in one app but review emails, chats, meetings, and docs to help identify and prep the most important work. This should change how work is done, how teams communicate, and how information is shared across an enterprise. Last week Google announced Workspace Intelligence which addresses (b) but not (a). A step in the right direction and hopefully more to come.

Process power will be a source of differentiation. Because AI adoption is not yet a solved problem, companies that figure it out will pull ahead of competitors. In the late 1990s NVIDIA figured out how to design chips using hardware emulation (Acquired, Sequoia). This allowed them to skip fabrication steps and ship twice as fast as their competitors. Those same opportunities exist today. Take the highly contested market for customer support AI agents. The underlying models and frameworks are the same so expect to see greater focus on process innovation around customer onboarding and continuous improvement and less transparency on the specific methods.

Creative destruction and renewed culture. The experience for employees still at Block is worth highlighting. Because there are fewer people, there is both the need and opportunity to change how work is done. There are fewer places to hide and poor performance is easier to spot and address. The reduction has also helped break down some of the old Block/CashApp/Afterpay cliques. That coupled with retention bonuses, salary bumps, and stock refreshes adding 30-50% to people’s total grant have helped renew the culture.

Be sure to reach out and share how AI is changing how your company works.

The steamroller is coming and we have a chance to get out of the way. Whether Block is really rewiring with AI or is just AI washing is irrelevant: the direction of travel is clear. We’re lucky if it’s AI washing because then we’ve got more time to take action to avoid the worst.

Pausing is not action

Calling for a pause in datacenter development limits US competitiveness and a pause in model development is a collective action problem that’s not enforceable.

Even if we did pause, we need to know what we’re giving ourselves time to implement. At best a pause buys time to do something and we need to be clear about what that something is.

Instead of dreaming of a pause, we need to develop a positive agenda that will maintain US competitiveness while addressing the downsides for workers. I’ve organized policy proposals into three buckets: No Regret, Worth Exploring, and Nope.

No Regret

These are proposals that have been around for a long time and that we can pursue with more urgency in this moment.

Stronger, More Uniform Unemployment Insurance. This is supported by economists because programs exist across all 50 states and support would go directly to those affected by any displacement, direct or indirect (i.e. it doesn’t matter if someone works in tech or downstream services they’re still eligible). While the federal government provides funding for administration, the programs are designed and run by the states which results in wide variation of benefits. For example, the maximum Weekly Benefit Amount (WBA) ranges from above $1,000 in Washington State to $275 in Florida (link). States in partnership with the federal government have an opportunity to strengthen these programs proactively.

Portable Benefits. It’s a quirk of history that our health benefits are tied to our employer: During WWII, the government instituted wage controls but companies still needed a way to attract employees without increasing wages. They started offering health benefits which employees valued and - because they weren’t taxed as income - were a cheap way to provide more compensation (link).

Today, the US is the only high income country with this setup which has all sorts of unintended consequences including compounding a person’s financial risk by taking away their health insurance at the same time they lose their job.

Making benefits portable would reduce disruption of workers’ coverage, reduce administrative overhead, and enable companies to contribute to the benefits of full-time/part-time employees as well as contractors (link).

Worth exploring

Movable Mortgages. During economic dislocations jobs and the people that need them may not be in the same region. The biggest barrier to relocation is often housing related: workers can’t afford to move because it would require getting a new mortgage at a higher rate. It’s possible to solve this by making mortgages portable and assumable (link). Portable means you bring your mortgage with you and apply it to the purchase of the new home, requiring only a smaller supplemental mortgage if needed. Assumable means you can assume the mortgage of the seller who may have a better rate, adjusting as needed for any differences in credit that would impact the rate (diagram).

Skills Development. Singapore has been running the SkillsFuture program for 10 years which provides every Singaporean with a credit they can use toward approved courses. There is a larger mid-career refresh to the credit to ensure life-long investment in workers in the right areas. ~70% of participants said it helped them advance and do better work. Making this work in the US could be challenging - it’s a national pastime to game government programs (see e.g. 324 people charged with $14.6 Billion in Alleged Fraud in 2025) - but given the need for people to continually reskill it’s worth a look.

US Government AI Fund. The first oil field in Norway was discovered in 1969. Today, Norway exports ~$150B/yr in oil and gas - quite a lot when you consider there are less than 6M people living there. Initially, Norway spent its oil revenue but this led to inflation, credit bubbles, and currency depreciation (IMF). In 1990 it established the Norwegian Government Petroleum Fund that turned a volatile income stream into a perpetual source of national wealth by only spending the real returns of the fund. This allowed the principal to grow and today the fund is worth $2 trillion or $360k per person.

Data centers like oil wells require significant capital investment and generate revenue streams that are significant but also unpredictable. The argument against a similar program funded by taxing chips, data centers, or tokens is that AI and associated revenues are the product of private companies rather than a found national resource. However, as AI displacement takes place the industry creates a significant negative externality that the government has an obligation to address because the market won’t.

This would be the hardest proposal to get right - you don’t want to crush a critical industry as it’s developing. But an industry that is growing rapidly and displacing more jobs than it creates needs to contribute to minimizing the downside.

Nope

Variable taxation as a function of labor intensity. There are some proposals that would lead to higher taxes on companies that lean on AI more than humans. Aside from the obvious measurement and reporting problems, this creates perverse incentives. There is an apocryphal story about economist Milton Friedman traveling to China in the 70’s when he saw a large construction crew using shovels to dig a canal. He asked “Why aren’t you using excavators which would be much more efficient?” “You don’t understand. This is a jobs program.” To which Milton responded, “Oh, I thought you were trying to build a canal. If it’s jobs you want, you should give these workers spoons, not shovels!”

Universal Basic Income (UBI). At some point this may make sense but we are so far from figuring out how to fund UBI that it’s not worth talking about. $1k per month for every person would cost over $4 trillion dollars a year. We don’t have that money. And any restrictions you put on distributions make it less “universal” and then it just looks like public assistance which becomes a political lightning rod.

Act!

Call your representatives! Their staff are required to listen to you and categorize your complaints feedback. They are looking for stories and soundbites. Find the contact info for your senators and reps.

Speak up! The next time you hear someone say “we just need to stop everything” ask them “what are we stopping for?” Let’s strengthen the programs we’ve got and take advantage of the crisis to make changes we should have made decades ago.

I’m not a doomer: As I’ve said before, I believe in the potential of AI to transform the human experience for the better and think it will make work better as well. But I’m also worried about what happens to the Bay Area and Seattle when the jobs start going away before new ones get created.

Because of the high concentration of tech jobs in the Bay Area and Seattle the unemployment rate could increase 2-10x.

That’s pretty dark which is why I wrote Part 2 to try and understand what economists and policy makers think we can do to avoid that future.

Concentrating the mind

There’s been a lot of country level stats about unemployment but those are averages which smooth over painful spikes. Two reasons to pick the Bay Area and Seattle:

• Concentration of tech jobs. On average only 5.8% people in the US are employed in tech. It’s 30% in the San Jose metro area. Given software engineering has the highest rate of AI agent adoption, these areas are ground zero for AI driven displacement.

• Feelings. My family and I have spent 10 years living and working in the Bay Area and now Seattle so this is home, these are our friends and colleagues.

Scenarios

I’ve modeled three job displacement scenarios:

• Goldman: 6% based on the Goldman Sachs report. This was their total displacement number over 10 years but I think it’s fair to use it as a shorter term estimate since we’re just focusing on tech, not a national average across multiple industries.

• Dot Com Part II: 17% which was the peak unemployment rate in tech during the Dot Com Bust of the early 2000s.

• Block: 40% assuming tech companies follow Block’s lead. (I’ll be following up in another article on AI washing and what’s actually happening at Block, Atlassian, and Oracle. For now I’m leaving this as the high end of the range.)

While I’m using these scenarios as “low, medium, high” these are all terrible outcomes.

Economic impact

We’ll look at the economic impact along three dimensions: Unemployment, Commercial Vacancies, and Housing Prices.

Unemployment

Using data on the number of tech jobs in each of the metro areas and the three scenarios above we can estimate the number of lost jobs in tech. But that’s just the start.

UC Berkeley economist Enrico Moretti estimates that each high tech job creates 5 indirect jobs in services (e.g. layers, teachers, nurses, waiters, hairdressers, carpenters).

When the high tech jobs go away, these jobs are impacted too. Not all 5 of the services jobs will be displaced as some jobs like teachers and nurses are not immediately affected by a contraction in discretionary spending (though they will be in the long term as population shifts). My estimate is that 2.5 of the 5 of the downstream jobs will be impacted.

A simple model tying all the numbers together is here and chart below. The result is that unemployment could surge 2-10x depending on the scenario. The most conservative would match peak unemployment during the Dot Com Bust.

Commercial Space

During the Dot Com Bust office rents fell as vacancies soared to 25% in SF.

After recovering, vacancies increased again post COVID and sit today at 30% in the SF and 23% in Seattle (link, link) - that’s higher than the peak of the Dot Com Bust! Rates of 40-50% would almost certainly lead to delinquency and default.

Housing prices

The story on housing in the Bay Area during the Dot Com Bust was more nuanced. From Zillow:

“Home values in the top third of the market fell 9.5 percent in San Jose and by 3.8 percent in San Francisco between May 2001 and January 2002. Home values in the bottom third fell 3.8 percent in San Jose, and actually increased by 4.8 percent in San Francisco.” (Zillow)

It’s not clear the we would see the same pattern today - interest rates are higher, inflation is still a concern, and the total unemployment rate would be higher in any scenario more aggressive than Goldman’s.

What this might look like

It is tempting to compare this to the Dot Com Bust which I’ve referenced several times. The Goldman scenario leads to a similar net unemployment rate and we’ve seen how the industry can recover and grow again. The difference is that in 2001 there was no structural change. Yes there was a bubble, exuberance, and jobs disappeared but the Bay Area remained a dynamic place to start and build businesses. Today, jobs are going away because not as many people are needed. The Bay Area may still be a good place to start and build businesses but fewer people will be needed to run them.

A better comparison would be what happened to manufacturing in the Rust Belt starting in the 1980s. Structural changes - increased foreign competition, factories closing here and opening in Asia - meant that not as many people were needed at manufacturers. Today, there are half as many jobs in manufacturing and many of the communities with the highest concentration of manufacturing jobs were devastated (link).

To avoid a repeat in different regions and a different sector, we’ve got work to do. On to Part 2.

When electrical motors were first developed, factory operators made the very logical decision to keep their factories largely the same but just replace their old engines (water, steam, diesel) with a big electrical motor. This was a simple substitution that gave them some benefits in terms of predictability and total available power.

But it didn’t change the complexity of how their factories were organized:

And zooming out a bit more:

Powered axels ran the length of the factory, driving a nest of belts and wheels that in turn ran the machines. Clutches would move the belt to an idler if the machine was not needed or more power was needed elsewhere.

Of course, factories look nothing like this today because eventually individual motors were placed in each one of the machines. This innovation allowed companies to eliminate the complex nest of axels and belts, reorganize their factories, and optimize material flow.

Post-AI organizational design

Today, companies look a lot like those old factories except that instead of the axles and belts we have a nest of teams, meetings, chat messages, emails, and other lossy connections that are required to build and ship products.

We’ve bolted on AI by giving ChatGPT/Claude/Gemini to each employee and telling them to figure it out. (And by the way if you don’t figure it out, you’ll get dinged on your perf.).

This limits the power and potential of AI because we haven’t changed how we organize work. All of the old axels and belts are in place and an an employee who uses Claude to draft an email, a piece of content, or new piece of code still has to push it through an organization that looks a lot like that old factory:

Why this means less jobs

What we are seeing is that fewer people will be required to design, ship, sell, and support products because with AI one person is doing not just more work, but work that was historically done by different functions. (See my previous article)

For a product manager that understands architecture, what is the benefit of extended back and forth with an engineer to build a product? For an engineer with strong product sense, what is the benefit of working with a PM to build a product?

A VP of Engineering I spoke with described the change as “PMs shifting right while engineers are shifting left”. A CTO described a new role that we’re just going to call “Product Builder.”

The same will happen for specialist roles like legal: Anthropic already uses a system they built in-house to automate legal review (link). The same will happen for roles that require taste: Google has shipped Stitch to enable people to “vibe design” (link).

If you can reduce the number of unnecessary nodes in a system you reduce its complexity and increase its velocity. Reducing nodes is a brute force approach and still doesn’t get at the reorgnaization of the “factory floor” for knowledge work that is needed.

What it will look like

The feeling in the pit of our stomach when we see photos of old factories is the same feeling we will get when we look back at how we work today.

That nest of lossy connections of chats/emails/docs/meetings will be replaced with AI systems that enable efficient coordination across larger groups. Every employee will have their own AI Chief of Staff organizing their work, fielding the stream of pings asking for status updates, asking how something works, where that file is, freeing them for the work that actually matters. Meetings will get smaller, more productive, and worth showing up to.

I’ve been spending a lot of time talking to people that are trying to figure out what that new factory floor looks like at big companies and startups. It’s still very much a work in progress with lots of pain points but there are some early learnings which I’ll share in a follow-up post.

If we know the steamroller is coming, let’s move our feet. Or we look like this guy:

My priors

This article will spend time talking about the downsides because those are problems we need to focus on and solve, but I believe that AI will transform humanity for the better.

It’s easy to see why: if you have 10x, 100x, 1000x more intelligence focused on solving humanity’s greatest challenges - health, energy, climate - we will make more progress faster. For those still skeptical that this technology represents intelligence that can make those sorts of contributions, you’ve got to challenge your assumptions. Dario’s piece does a good job laying out what that world could look like.

But we’ve got to struggle through to get there.

Fewer jobs (in the short term)

This is be self evident to people that have worked with the tech but a few data points:

• Layoffs across medium to large technology companies. Block at 40%, Atlassian 10%, Meta reportedly getting ready for 20%.

• Philosophy of “talent dense” teams at startups. Translation: “We have been Claude pilled and need fewer people to get things done.”

• Slower / no hiring for specific roles. “We are increasing the scope of responsibilities for existing people because it’s possible for them to get more done.” “There is no need to ever hire someone with the words ‘Content’ in their title ever again.” This is what I hear from startup execs across verticals.

Tech is first and the rest of knowledge work will follow. This chart from Anthropic was getting shared around as inspiration to entrepreneurs to find the next great vertical SaaS AI opportunity. But it’s also the roadmap of which tasks and industries that will be hit after software engineering and tech.

Economic hardship

I’m going to do a deep dive on economic hardship in a later post so here I’ll focus just on lost wages. Goldman Sachs seems to land in between “We’re Doomed Citrini” and “Nothing to See Here Citadel.” Goldman says 6-7% of jobs will be displaced over a decade leading to an unemployment rate that is 0.6% higher than today. “But if it’s more frontloaded, the impacts on the economy are much larger.” Notably Goldman did not provide a high end number if it was frontloaded. That’s ominous.

Let’s say unemployment goes up between 0.6% and 1.8% on a working population of 162.9M (BLS) which puts between 997k and 2.9M people out of work. The initial impact will be highest on white collar workers with average income of $68k ($32.69/hr * 2080 hours per year).

Total lost wages then will be between $66B and $200B.

That’s just lost wages, it doesn’t take into account the downstream effects of lost wages on services businesses, housing, cars, travel, etc.

Current funding to mitigate job displacement is a rounding error

In 2025 there was ~$4B in funding for jobs training and displacement support through the WIOA programs (link). It’s not enough and that number is set to shrink:

• That’s 2% of the $200B in lost wages

• That’s 0.5% of the planned $750B in data center capex this year (link)

• The Administration actually proposed reducing the WIOA budget from $4B to $3B (link)

As a country we are investing at an unbelievable rate to win the AI race (which we should continue to do) but are failing to invest in what may be the largest worker dislocation in human history.

Activity over accomplishment

Our elected officials are not working together to address the problem. Since 2023, over 250 federal bills related to AI have been introduced. Only 2 have become law (link):

• Take it Down Act. Criminalizes non-consensual intimate imagery including those from deepfakes. This is good, common sense legislation and focuses correctly on harms. (And should have been a law a long time ago)

• One Big Beautiful Bill Act. Funding for the government to adopt AI across Department of War, Department of Commerce, etc. No focus on mitigating impacts of job displacement.

Chat GPT was released in November 2022. More than 3 years there is zero federal response to address jobs displacement.

What to do

It is critical that the US stay ahead in the AI race. But right now we're making the steamroller bigger and cutting the budget for the people in its path. Let’s figure out how to get them out of its way.

Share