Former Consensys CMO: The Evolution of Corporate Forms in the AI Era

Author: Lex Sokolin

Compiled by: Jiahui, ChainCatcher

This article explores how AI is reshaping organizational structures themselves. Companies are moving from Amazon-style “two-pizza teams” (a team of about 6–10 people, keeping the organization agile) to “AI-native” mini-teams of 3 to 5 people, with a significant leap in productivity.

We compared two paths:

Klarna’s AI replacement strategy ended in failure. The number of employees was cut from 5,500 to 3,400, and service quality issues ultimately forced the company to rehire.

Coinbase and Ramp, meanwhile, chose to reorganize their businesses around AI enhancement and orchestration. Coinbase laid off 700 people and shifted to single-product teams and AI code generation.

Ramp built an internal AI harness framework, used daily by 99.5% of employees, covering more than 350 business skills.

In addition, we also analyzed why companies such as Box and Plaid are being repriced by capital markets as AI infrastructure—the key reason is that they control enterprise-grade, permissioned data that is essential for running AI agents.

The Third Evolution of Organizational Forms

A few months ago, we discussed the “Zero Human Companies” and the AI economic autonomy curve:

Although there are forces pushing toward establishing organizations with no human intervention, the current economic entities are still us humans.

The most pressing work right now is transforming existing traditional companies into AI-first forms.

It is such a massive opportunity that Anthropic is teaming up with the entire private equity industry to push this forward.

Beyond those astonishing financial figures, we also began to clearly perceive another entry point for AI’s impact: the way people build and organize companies.

Organizational structure itself is a technology.

Waterfall development nurtured the software giants of the early tech era—hierarchical and rigid.

Then the industry shifted to lean teams using agile methodologies, and later agile evolved into Amazon’s pioneering “two-pizza teams.” It is exactly this operational structure that underpins every modern fintech company today.

But the tide is turning again.

At the end of 2025, Martin Harrysson and Natasha Maniar from McKinsey provided the next forecast:

“AI-native roles essentially mean that we are shifting from ‘two-pizza structures’ to ‘single-pizza teams’ of 3 to 5 people.”

Halve the people, keep the work going.

On May 5, 2026, Brian Armstrong added strong evidence to this claim by laying off 700 people.

What did Coinbase do?

Coinbase cut 14% of its 4,951 employees.

Part of the reason is simply the normal market-cycle operation of a business tightly tied to trading volume—its Q1 revenue is expected to be $1.7 billion (down 26% year over year), and earnings per share (EPS) plunged 86%.

But what deserves close attention is how its management planned the path for AI rollout in modern fintech/crypto companies, and their expectations for future productivity per person.

Coinbase engineers can now ship products in a few days that previously took weeks—and that efficiency is accelerating.

Armstrong is restructuring business lines to ensure that under the CEO and COO there are at most five management layers.

Pure “managers” will no longer exist—every leader must also be an individual contributor, proficient with modern tools, able to both lead and personally jump in as a “player-coach.”

Cross-functional “AI-native teams” fully replace traditional teams. Coinbase even pilots internally integrating engineering, design, and product functions into one-person teams.

Coinbase, a listed giant with $7 billion in revenue, is operating with single-product teams.

In September 2025, Armstrong publicly stated that 40% of Coinbase’s code is generated by AI every day, and plans to raise that ratio to 50% in October.

In Stripe co-founder John Collison’s Cheeky Pint podcast, he admitted he fired engineers who still refused to use Cursor and GitHub Copilot even within a week of the enterprise license being issued:

“Some people just don’t use them, so they were fired.”

The V1 version was a direct replacement, but it failed

However, Coinbase was not the first fintech to lay off staff under the banner of AI.

Remember Klarna’s textbook “AI cost-cutting” experiment in 2024? At the time, it seemed to foreshadow an astonishing productivity boom in the future.

But we believed then that it was more like credit-cycle tightening than true innovation.

CEO Sebastian Siemiatkowski loudly announced that the OpenAI-powered AI assistant handled 2.3 million conversations in its first month—two-thirds of all customer chats—completing work equivalent to 700 full-time customer service agents.

• Employee count dropped from 5,500 to 3,400
• Expected profit increase: $40 million
• Customer issue resolution time shortened from 11 minutes to 2 minutes

Yet all of it collapsed quickly once it met reality.

Customer satisfaction (CSAT) for complex tickets plummeted, and the rate of repeat contact surged.

By May 2025, Siemiatkowski admitted to Bloomberg that the company “took too big a step.” Klarna had to start rehiring using a remote model similar to Uber—hiring flexible students, full-time parents, and workers in remote areas.

In just a few days, the Commonwealth Bank of Australia rapidly halted 45 voice-bot replacement projects. Taco Bell also pulled voice AI from 500 drive-thru restaurants.

Gartner predicts that by 2027, half of the companies that had drawn up “full replacement plans” will abandon those plans.

Klarna’s IPO still jumped 30% on its first day, reaching a valuation of $20 billion—reflecting, to some extent, that if companies correct course in time, public markets can be fairly forgiving.

But this blunt “replacement” logic—cutting a single human role and inserting a large language model (LLM)—may work for “quantity” metrics, but it will inevitably break down for “quality” metrics.

The cost of rehiring far exceeds the initial savings. Clearly, the fintech industry’s first attempt at AI digital transformation delivered a mixed result of wins and losses.

But it will absolutely not be the last attempt.

The V2 version is capability-enhanced, with Harness as the moat

Ramp officially released “Glass” in early April 2026.

Seb Goddijn, an internal AI expert who co-built the tool with five colleagues, published a long-form article. That same day, Ramp’s CEO Eric Glyman retweeted it. Within hours, the article topped Hacker News.

On why the V1 version failed, Goddijn was sharp and to the point:

“The biggest barrier blocking AI adoption is not the models themselves, but the extreme complexity of configuring the environment in which AI runs.”

Glass is exactly what Ramp created to smash this barrier:

First, automated access configuration—after logging in via Okta SSO, every authorized internal tool (Salesforce, Gong, Notion, Linear, Snowflake, Slack, Zendesk, and Ramp’s own internal tools) is integrated at the underlying level.

Second, set up Dojo—a marketplace containing more than 350 AI skills. Each skill is a Markdown file that teaches the agent how to complete a task. All of them are stored in Git, undergo code review, and are under version control.

An agent named Sensei pushes the five most relevant skills to new employees on their first day.

Third, build a persistent memory bank—generated automatically based on identity verification, and continuously refreshed through a 24-hour end-to-end processing pipeline. As a result, whenever the agent steps into a conversation, it already fully understands the employee’s team, projects, active tickets, and the communication context across ongoing threads.

Today, 99.5% of Ramp employees use AI every day.

Half of Ramp’s code is written by AI, and it is moving toward 80%. Its Chief Product Officer Geoff Charles has rolled out an L0–L3 maturity framework, where L3 means deploying production-level features directly through AI agents.

Any employees still stuck at the L0 level are essentially treated as idle.

Ramp’s valuation is currently as high as $32 billion, with ARR (annual recurring revenue) of $1 billion, ranking at the top of Fast Company’s 2026 list of the most innovative companies in the financial industry.

Klarna tries to lower the staffing threshold through automation, while Ramp is pushing hard to raise the productivity floor for each employee. Coinbase sits somewhere in between.

AI Harness

At the core of everything is the concept of “AI Harness.”

Companies like Manus pioneered architectures that compress raw AI intelligence into repeatable business flows, while orchestration frameworks like OpenClaw have popularized them for a wider audience.

A Harness is a comprehensive system that perfectly integrates identity verification, system integration, a memory bank, a skills directory accumulated by teams, nightly batch scheduling procedures, and multi-pane interaction interfaces that allow analysts to run multiple workflows in parallel.

And those cutting-edge large language models are simply plug-and-play components within this Harness. When OpenAI releases GPT-5.5, or when Anthropic releases Opus 5, Ramp only needs to swap out the model—everything else continues running as usual.

Anthropic’s Cowork product officially went into general availability (GA) in Q1 2026. It includes 11 plugins tailored to specific job roles, spanning sales, finance, legal, marketing, HR, R&D, design, and operations. This job-classification logic is the same as Glass’s Dojo.

Once you accept that AI productivity is shaped by business flows rather than chat windows, job roles naturally become the smallest natural units of an AI organization.

This is also the underlying logic behind the tools that aim to build “Zero Human Companies” when they think about how to construct AI-first organizations. See the Polsia described below, as well as the subsequent industry’s rapid segmentation map.

Capital markets are catching up

While many traditional software companies are struggling because of AI’s disintermediation, a certain type of player is accelerating against the trend.

These companies dug deep into their own data moats early, and now they can comfortably and seamlessly overlay one-off AI software on top.

Take the enterprise file storage company Box as an example: after it released its FY2026 Q4 earnings report, its stock surged 10%. Aaron Levie revealed the key in the earnings call:

“Files, at their core, are the natural work units for AI agents.”

Enterprise Advanced—the premium subscription tier for Box’s AI and workflows—costs 30% to 40% more than the traditional flagship Enterprise Plus.

Q4 billings reached $420 million, up 5% year over year.

• Box Extract can precisely extract structured data from contracts
• Box Shield Pro brings agentic AI directly into the access control system
• Box AI Studio’s professional mode and extended mode allow agents to handle multi-step workloads within larger context windows

In a GeekWire interview, Levie commented:

“Apart from the first 12 months, Box has never felt more like a startup than it does today.”

Remember that as much as 95% of enterprise data is unstructured. AI agents crave this data, and they must be able to access it while fully preserving permission boundaries.

Whoever controls this permissioned data vault can shed the “cheap storage” label and be revalued by capital markets as “agent infrastructure.”

In the past, the market treated Box as the slightly awkward sibling of Dropbox, and its stock price hovered around $26 for a long time. Now, Wall Street’s consensus price target is $35.63, leaving about 35% upside versus the current price.

Another example is Plaid—a financial data aggregator that nearly got acquired by Visa and had hoped to become a direct payments network.

But for a period of time, Plaid’s situation was awkward: Web3 overtook Web2 and replaced it as the new darling of financial infrastructure.

From a peak valuation of $13.4 billion in 2021, Plaid declined to $6.1 billion by April 2025 in the primary market round, and then rebounded to $8 billion in February 2026 after a secondary market tender offer that provided liquidity for employees.

It has to evolve.

About 20% of Plaid’s latest customers are AI-native companies—they are building agents that require permissioned access to financial data and rely on trusted identity infrastructure.

In early 2026 testing, Plaid Protect’s anti-fraud platform detected 50% more fraud attempts than comparable identity-verification tools.

Plaid Bank Intelligence, together with Retention Score and the upcoming Primacy Indicators, sells back customer churn prediction capabilities to banks.

Plaid is being repriced as the world’s largest permissioned financial transaction data corpus.

It’s not just a data pipeline—data pipelines have always been cheap. The true asset is the intelligence built on top of it, and the share of AI-native customers powerfully supports this argument.

A typical case is its integration with Perplexity—together building a complete personal finance “computer.” How we miss Mint.com! (the US national personal finance app launched in 2006)

Box and Plaid are on the same side of the same track.

Both were priced during the zero-interest-rate (ZIRP) era under the “SaaS king” logic; they watched valuations get cut in half. Now they are being re-underwritten under a new logic: an unstructured content vault plus a permissioned data network—these are the foundational substrates that AI can read in the V2 era.

The V3 version is orchestration—“the single-person company” is born

Sam Altman and other tech CEOs are making a bet on what year the first “billion-dollar-scale single-person company” will emerge.

Dario Amodei set the probability of it happening within 2026 at 70% to 80%, and named three areas: proprietary trading, developer tools, and automated customer service.

Sequoia is adjusting its investment and underwriting model, treating “agentic leverage”—revenue per capita—as the top signal. For companies from Y Combinator’s early batches, 95% of the code is generated by AI.

In fact, there are already companies that have created astonishing economic leverage with AI.

In these companies, the CEO becomes an “agent orchestrator,” dispatching countless AI agents from a giant command cockpit.

Organizational charts turn into business-flow diagrams that can be outsourced to machines. Workforce budgets turn into compute budgets.

The initial form of these companies will live in narrow niches—proprietary trading, developer tools, and consumer software with network effects. In these scenarios, work is fully digital, regulation is light, and trust costs are low.

They will be fragile, because systems with single points of failure are inherently fragile.

They will also struggle to penetrate regulated enterprise markets, because there, the name on the contract and the face behind it are structural.

But these companies already exist.

Every technological revolution destroys the roles that the previous paradigm considered critical—“computers” (early human calculators), assembly line foremen, project managers, and middle managers.

Companies that figure out the “new economic organization form” early often reap huge rewards for moving first.

For example, Amazon’s “two-pizza rule,” and its ability to maintain innovation at a scale of millions of employees, are themselves a moat.

Whether we ultimately land on “single-person companies” or “Zero Human Companies” is not the real question.

We are still in the middle of a digital transformation process, and delivering value across the entire economy along this path will generate returns in the trillions of dollars.

The real question is: who can own or build the right AI Harness today, and therefore design the correct organizational chart for companies in 2026?

This means upgrading this enterprise super-organism so it can keep fighting—alive for another day.

Hopefully, humans can also get what they want from it.