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

Author:Lex Sokolin

Other Organiser

THIS PAPER EXPLORES HOW AI CAN RESHAPE THE STRUCTURE ITSELF. THE COMPANY IS FROM THE AMAZON'S DOUBLE PIZZA TEAMA team of about 6-10 people, maintaining an agile organizational structureTURNING TO THE "AI RAW" GROUP OF THREE TO FIVE PEOPLE, WITH A HUGE PRODUCTIVITY LEAP。

We compared two paths:

Klarna's AI alternative strategy to fail. The number of employees was cut from 5,500 to 3,400, and the quality of services eventually forced them to recruit again。

Coinbase and Ramp choose to reorganize their business around AI enhancements and organization. Coinbase downsized 700 people while turning to single-person product teams and AI code generation。

Ramp has developed an internal AI steering framework (harness), which 99.5 per cent of employees use daily, covering over 350 business skills。

In addition, we have analyzed why firms like Box and Plaid are repricing capital markets as AI infrastructure, at the core of which is that they control the authorized enterprise-level data necessary to operate AI intelligence。

Third evolution of organizational patterns

A few months ago, we discussed Zero Human Companies and AI's Economic Autonomy Curve:

Although there are forces to promote the establishment of organizations that are completely free of human intervention, the current economic body remains our human race。

THE MOST DIFFICULT TASK AT THIS TIME IS TO TRANSFORM EXISTING TRADITIONAL COMPANIES INTO AI-PRIORITIZED FORMS。

This is a huge opportunity for Anthropic to work with the entire private equity industry to move this forward。

IN ADDITION TO THOSE AMAZING FINANCIAL DATA, WE'RE BEGINNING TO NOTICE CLEARLY ANOTHER ENTRY POINT FOR AI'S IMPACT: HOW PEOPLE BUILD AND ORGANIZE THEIR COMPANIES。

The organizational structure is in itself a technology。

Waterfall breeds a great alligator who reigned in the early age of technology, with a very sophisticated software development。

The industry then turned to a lean team using the agile method, which then evolved into the original "two pizza team" of the Amazon. It is this operating structure that has shaped every modern financial technology company today。

But the tide has changed again。

Martin Harrysson of McKenzie and Natasha Maniar gave a prognosis for the next version in late 2025:

"AI'S ORIGINAL ROLE IN ESSENCE MEANS THAT WE'RE MOVING FROM 'TWO PIZZA STRUCTURES' TO 'ONE PIZZA TEAM' COMPOSED OF THREE TO FIVE PEOPLE."

Halve human beings and do business。

On May 5th, 2026, Brian Armstrong added a strong note to this assertion by downsizing 700 people。

Coinbase did what

Coinbase reduced the number of employees by 14 per cent out of 4,951。

THIS IS DUE IN PART TO THE FACT THAT THIS IS STILL THE NORMAL MARKET CYCLE OPERATION OF A COMPANY WITH A HIGHLY TIED BUSINESS AND VOLUME OF TRANSACTIONS, WHICH IS EXPECTED TO RECEIVE $1.7 BILLION IN THE FIRST QUARTER (A DECREASE OF 26 PER CENT OVER THE SAME PERIOD) AND A SHARP DECLINE OF 86 PER CENT PER SHARE OF PROCEEDS (EPS)。

HOWEVER, IT IS OF GREAT CONCERN HOW ITS MANAGEMENT PLANS FOR AI LANDING PATHS IN MODERN FINANCIAL TECHNOLOGY/ENCRYPTING COMPANIES, AND THEIR EXPECTATIONS OF FUTURE PER CAPITA PRODUCTIVITY。

Coinbase's engineers now have only a few days to release the products that used to take weeks to get online, and this is accelerating。

Armstrong is reshaping its business lines to ensure that only five levels of management exist under CEO and COO。

Purely "managers" will no longer exist -- every leader must be a personal contributor and a "player and coach" with modern tools that can both lead the team and go down in person。

The cross-functional "AI Original Team" completely replaced the traditional team. Coinbase even piloted internally the integration of engineering, design and product functions into a single-person team。

Coinbase, a $7 billion listed giant, is running a single product team。

In September 2025, Armstrong publicly stated that 40% of Coinbase's code is generated by AI and plans to increase it to 50% in October。

Among the podcasts of Cheeky Pint, co-founder of Stripe, John Collison, he confessed to dismissing the engineers who had refused to use Cursor and GitHub Copilot within a week of the issuance of the enterprise version:

"Some people just don't, so they're fired."

V1 VERSION IS A DIRECT REPLACEMENT, BUT IT FAILED

Coinbase, however, is not the first financial technology firm to reduce its staff on AI。

Remember Klarna's "AI Dropbook" experiment in 2024? At that time, it seemed to show an alarming productivity outbreak in the future。

But we thought it was more like a credit-cycle tightening than a real innovation。

CEO Sebastian Siemiatkowski had announced that the OpenAI-driven AI assistant had handled 2.3 million conversations in the first month, representing two thirds of all clients ' conversations, and had completed a workload equivalent to 700 full-time visitors。

• Total number of employees dropped from 5,500 to 3,400
• Increase in expected profits: $40 million
• Client problem resolution time reduced from 11 minutes to 2 minutes

However, all of this quickly collapsed after touching reality。

CLIENT SATISFACTION WITH COMPLEX WORK ORDERS (CSAT) FELL SHARPLY AND THE RATE OF DUPLICATE CONTACT SURGED。

By May 2025, Siemiatkowski had told Bloomberg that the company was going too far. Klarna had to start re-engineered with a remote model similar to Uber — hiring time-flexible students, full-time parents and workers in remote areas。

The Federal Bank of Australia stopped 45 voice robot replacement projects in a few days. Taco Bell also withdrew voice from 500 car shuttle restaurants。

Gartner predicts that by 2027, half of the companies that had a "comprehensive replacement plan" will have abandoned the programme。

Klarna ' s IPO continued to grow by 30 per cent to reach a $20 billion valuation on the first day, reflecting, in part, the fact that the open market is fairly tolerant as long as companies correct it in a timely manner。

BUT THIS SIMPLE AND BRUTAL "SUBSTITUTION" LOGIC, NAMELY, THE DIRECT LOSS OF A HUMAN POST AND THE INSERTION OF A LARGE LANGUAGE MODEL (LLM). IT MAY WORK ON "QUANTITATIVE" INDICATORS, BUT IT WILL COLLAPSE ON "QUALITATIVE" INDICATORS。

THE COST OF RE-RECRUITMENT IS FAR HIGHER THAN THE COST SAVINGS ORIGINALLY MADE. IT IS CLEAR THAT FOR THE FIRST TIME IN THE FIELD OF FINANCIAL TECHNOLOGY, THE AI DIGITAL TRANSFORMATION ATTEMPTED TO DELIVER A MIXED RESPONSE。

But this will never be the last attempt。

V2 VERSION IS ENHANCED BY Hi don't think soFor the moat

Ramp officially released "Glass" in early April 2026。

An internal AI expert, Seb Goddijn, who worked with five colleagues to build the tool, published a long paper. On the same day, the CEO Eric Glyman of Ramp was tweeting. Within hours, the text is on the front page of Hacker News。

Goddijn points out with blood:

"THE PRIMARY BARRIER TO THE SPREAD OF AI IS NOT THE MODEL ITSELF, BUT THE EXTREME COMPLEXITY OF THE AI OPERATING ENVIRONMENT."

Glass was created to break this barrier:

First of all,Automatic access configuration— Every authorized internal tool (Salesforce, Goong, Notion, Linear, Snowflake, Slack, Zendesk, and Ramp own internal tool) is accessible on the ground floor, just through Okta SSO。

Second, set upDojo- A market with more than 350 AI skills, each of which is a Markdown file, which is responsible for teaching intelligence to do a job. They are all stored in Git, code reviewed and version controlled。

A nameSenseiOn the first day of a new employee's employment, the intelligent pushes five of his most relevant skills。

Third, buildSustainable Memory Library— Automatically generated on the basis of an authentication connection and continuously refreshed through a 24-hour integrated treatment pipeline. As a result, when involved in each of the dialogues, the intelligent have a full grasp of the staff member ' s team, participation in the project, active work list and ongoing communication。

Today, 99.5% of Ramp employees are using AI every day。

Ramp half the code is written by AI and is moving towards 80%. Its Chief Product Officer, Geoff Charles, has introduced a L0-L3 maturity framework in which L3 represents the direct distribution of production-level functions through AI intelligence。

ANY EMPLOYEE WHO REMAINS AT THE L0 LEVEL IS IN ESSENCE CONSIDERED TO BE IDLE。

Ramp is currently valued at $32 billion, ARR (annual recurrent income) at $1 billion, with Fast Company 2026 being the leading financial innovation company。

Klarna tried to lower the manpower threshold by automating it, while Ramp was struggling to lift the output line for each employee. Coinbase is in between。

AI Harness

It's all at the heart of this"AI Harness."The concept。

Companies like Manus have created structures that compress the original AI intelligence and transform it into a repeatable business stream, while frameworks such as OpenClaw push it to the general public。

A set of Harnesses is a combination of identification, system integration, memory banks, skill catalogues for team deposition, night runs, and multi-window interfaces that allow analysts to run in parallel with multiple lines。

And those front-line large-language models are just parts of the Harness that can be replaced -- when OpenAI rolls out GPT-5.5, or Anthropic releases Opus 5, Ramp simply replaces the model, and everything around it works。

Anthropic's own Cowork product, which was formally commercial (GA) in the first quarter of 2026, contained 11 job-specific plugins across sales, finance, legal affairs, marketing, HR, R & D, design and operation - a job classification logic similar to that of Dojo in Glass。

ONCE YOU ACCEPT THAT "AI'S PRODUCTIVITY IS SHAPED BY BUSINESS STREAMS RATHER THAN CHAT FRAMES,",Position RoleIT IS LOGICAL TO BECOME THE SMALLEST NATURAL UNIT OF THE AI ORGANIZATION。

And that's the bottom logic of thinking about how to build an AI priority organization. Details are given below in Polsia and later in the fast-disaggregated industry。

Capital markets are catching up

WHILE MANY TRADITIONAL SOFTWARE COMPANIES ARE STRUGGLING WITH THE DE-BROKERING OF AI, ONE TYPE OF PLAYER IS RAGING BACKWARDS。

THESE COMPANIES DUG DEEP INTO THEIR OWN DATA MOORINGS, AND NOW HAVE ROOM FOR A ONE-OFF FOLDING OF AI SOFTWARE。

In the case of Box, an enterprise-level document storage company, the equity price increased by 10 per cent following the release of the fourth quarter of fiscal year 2026. Aaron Levie said at the newspaper conference:

"THE FILE, AFTER ALL, IS THE NATURAL WORK UNIT OF THE AI SMART BODY."

Enterprise Advanced, a high-level subscription level for Box main-playing AI and the workflow, is priced 30 to 40 per cent more directly than the traditional flagship version of Enterprise Plus。

billings for the fourth quarter amounted to $420 million, an increase of 5 per cent over the same period。

• Box ExpressCapable of extracting structured data from contracts
• Box Shield ProDirect introduction of intelligent AI into access control
• Box AI StudioProfessional mode and extension mode to allow smarts to handle multistep loads in larger context windows

Levie, in an interview with GeekWire, deplored:

Except for the first 12 months, Box has never felt as if it were a start-up company

YOU KNOW, UP TO 95% OF BUSINESS DATA IS UNSTRUCTURED. AI INTELLIGENCE IS EXTREMELY THIRSTY FOR THESE DATA AND MUST BE CALLED WITH THE PERMISSION BOUNDARIES REMAINING INTACT。

Whoever controls the authorized data bank can get rid of the "cheap storage" label and be revalued by capital markets as "smart infrastructure"。

Once upon a time, the market saw Box as the slightly awkward brother of Dropbox whose stock price was hovering at $26. Today, the Wall Street target has been set at $35.63, with a 35 per cent premium against current prices。

Another example is Plaid, a financial data collector who was almost acquired by Visa and was hoping to become a direct payment network。

But for some time, Praid was in a rather awkward position: Web3 later went up and replaced Web2 as a new financial infrastructure pet。

From the peak of the valuation of $13.4 billion in 2021, Plaid went down to $6.1 billion in April of 2025, and then recovered to $8 billion in February 2026, when a secondary market offer to provide staff mobility was made。

It has to evolve。

About 20 percent of the newest clients of Plaid are AI Natives - they are building smarts that require access to financial data and depend on a credible identity base。

During tests in early 2026, the anti-fraud platform of Plaid Protec detected 50 per cent more fraudulent attempts than similar identification tools。

Plaid Bank Inteligence, with Retention Score and upcoming Primacy Indicators, sells the loss of predictive capabilities to banks in reverse。

Plaid is being re-pricing toThe largest worldwide source of authorized data on financial transactions LibraryI don't know。

IT IS NOT A DATA CONDUIT — IT IS ALWAYS CHEAP. THE TRUE ASSET IS THE INTELLIGENCE THAT IS BUILT ON IT, AND THE PROPORTION OF AI'S ORIGINAL CLIENTS IS THE STRONGEST SUPPORT FOR THIS ARGUMENT。

A typical case is that it's integrated with Perplexity -- building a fully integrated personal financial management computer. How do we miss Mint.com

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

Both families were priced according to the logic of the "SaaS Hegemonic" in the ZIRP era, and are now re-admitted to a completely new logic — the unstructured content vault, and the authorized data network — as the bottom base that V2-era firms can read with intelligence。

THE V3 VERSION IS ORGANIZED -- SINGLE COMPANY WAS BORN

Sam Altman and other technology CEOs have a bet on when the first billion-dollar single-person company will be born。

Dario Amodei set its probability of appearing in 2026 at 70 to 80 per cent and named three areas: self-employment, developer tools, automated client services。

Sequoia is adjusting its investment underwriting model to include "intellectual leverage" as the primary signal of per capita income. Y Combinator's early batch of incubators, 95% of the codes have been generated by AI。

IN FACT, COMPANIES HAVE BEEN ABLE TO CREATE AMAZING ECONOMIC LEVERAGE THROUGH AI。

In this kind of company, CEOs become "agent orcherator" and dispatch numerous AI intelligents in a giant cockpit。

The organizational chart became a business flow map that could be outsourced to the machine for execution. The labour budget has become a capital budget。

The first-generation form of such companies would be in narrow areas — self-trade, developer tools, sub-consumption software with network effects. These scenarios are fully digital, with light regulation and low trust costs。

They are vulnerable because all single point failure systems are fragile。

They also have difficulty accessing the regulated business market, where the name and face of the contract are structural。

But this kind of company, it's already here。

every technological change that takes place destroys the role that the previous paradigm considers to be key -- - "computer " , production line foreman, project manager, middle manager。

And companies that take the lead in identifying "new forms of economic organization" tend to get huge returns by moving ahead。

By way of example, the Amazon's double pizza rule and its ability to remain innovative at the size of a million employees is itself a moat。

It's not really a question of whether we end up in a single company or a zero human company。

Now, we're still in the process of digital transformation, and we're in the same vein. Delivery of value on an economy-wide basis would yield hundreds of billions of dollars in returns。

The real question is:Whoever can own or build the right AI Harness today can design the right organizational chart for the 2026 company。

This means upgrading the corporate super-organism so that it can continue fighting and live another day。

I hope that we, the human race, will also benefit from it。