The Fully Automated Company: 7 Real Case Studies From 2026

Everyone talks about the fully automated company. Almost nobody can point to one. The phrase floats around keynotes and LinkedIn posts the way "the paperless office" did in 1995 - aspirational, slightly suspicious, and never quite the thing in front of you. So we went looking. Not for press releases, not for demos, but for organizations that have actually pushed automation past the point where the org chart still looks normal.

We found seven. None of them is fully automated in the literal sense. Each one has redrawn the line between what a machine does and what a human decides, and each one has paid a price for either pushing the line too far or stopping just short. Together they show what the fully automated company actually looks like in 2026: not a building with the lights off, but a small group of people sitting on top of a stack of agents, watching the dashboards, and occasionally pulling a lever.

1. Medvi: The Two-Person, Billion-Dollar Company

Matthew Gallagher built Medvi as a telehealth platform for weight-loss prescriptions. By spring 2026, the company was on track to clear 401 million dollars in revenue with two employees on payroll. Gallagher himself, and a co-founder who handles operations. That is the entire headcount.

The trick is not that Medvi avoided regulation. It is that almost every function a normal telehealth company would staff - intake screening, doctor matching, prescription routing, refill reminders, payment reconciliation, customer support - runs through a stack of contracted clinicians and AI agents wired together with a thin layer of orchestration code. The doctors are not employees; they accept cases through a marketplace. The agents do the rest. Gallagher's job, by his own description, is to watch the dashboards and decide which knobs to turn.

The model is not without consequences. The FTC has opened a preliminary inquiry into whether the lack of human oversight at the intake layer is creating safety risks, and several state pharmacy boards have started asking questions about who, exactly, is responsible when a script gets written badly. Gallagher's answer - that the AI is auditable and the contracted doctor signs every prescription - is technically correct and politically fragile.

What is automated: intake, triage, doctor matching, billing, support, refill flows.
What is still human: the prescribing physician, the founder watching the dashboard, and the lawyers.
The question it raises: if a two-person company can do 400 million in revenue in a regulated industry, what does it mean for the thousand-person companies trying to compete with it?

2. Base44: An 80 Million Dollar Exit in Six Months

Maor Shlomo shipped Base44 in late 2024 as a vibe-coding platform - a place where non-developers could describe software in plain language and watch it appear. He had no employees, no funding, and no go-to-market team. By June 2025, the platform had 250,000 users. Wix bought the company for 80 million dollars in cash, with a performance-based earn-out that could push the total close to 170 million.

What makes Base44 interesting is not the exit. Exits happen. It is that the entire product was built and operated by one person using AI tooling, and that the acquirer's first instinct after the deal was not to integrate the team into Wix's engineering org but to leave the architecture alone. Base44 has since crossed 100 million dollars in annual recurring revenue inside Wix, and the small group now running it operates more like a portfolio company than a department.

The Base44 story is the cleanest version of a thesis that keeps showing up in 2026: when the cost of building software collapses, the constraint moves to taste and distribution. Shlomo had both. The team he did not hire would have slowed him down.

What is automated: the entire product surface, including the code-generation pipeline users interact with.
What is still human: product direction, pricing, the decision to sell.
The question it raises: how many companies in the Wix portfolio - or any portfolio - could have been built the same way?

3. Cognition (Devin): The AI Employee With a Salary

Cognition's Devin is sold as the first autonomous AI software engineer. Customers do not license a copilot; they pay a monthly fee per Devin instance, the way they would pay a contractor. By early 2026, the company was reporting 73 million dollars in annual recurring revenue and was in talks at a 25 billion dollar valuation. Dell, Cisco, and Goldman Sachs are among the named clients.

The framing matters. Devin is not a tool inside an IDE. It is a named worker that picks up tickets, opens pull requests, runs tests, and asks for human review when it is uncertain. Inside the customers we spoke to, Devin instances are assigned to product teams the way a junior engineer would be. They have Slack handles. They show up on sprint boards. The humans on those teams describe the work the same way they would describe working with a remote contractor in another time zone - sometimes great, sometimes frustrating, occasionally surprising.

What is being automated here is not coding. It is the role. The unit of hiring has changed from a person to a subscription, and the management overhead that used to come with adding a junior engineer - onboarding, mentoring, performance reviews - has been replaced with a billing relationship.

What is automated: ticket triage, implementation, testing, and a meaningful share of code review.
What is still human: the senior engineer who decides what gets built and approves the merge.
The question it raises: when an AI agent has a salary line in the budget, what is the difference between a contractor and an employee?

4. Walmart: The Self-Healing Supply Chain

Walmart is not a startup story, which is why it matters. The company has spent the last eighteen months wiring AI agents into the spine of its supply chain, and the results are now showing up in the financials. The internal Wally agent, which sits on top of inventory data and store-level demand signals, has been credited with 55 million dollars in waste reduction over the last year by catching slow-moving stock before it spoils and rerouting it.

The more ambitious piece is what Walmart calls Trend-to-Product, a system that ingests social signals, search data, and store-level sell-through, and uses them to compress the cycle from "we should make this" to "it is on the shelf" from months to weeks. Roughly 60 percent of Walmart's stores now receive automated distribution recommendations that go straight from agent to truck without a human signing off on the line item.

This is a different shape of automation than the Medvi or Base44 stories. Nobody at Walmart has been replaced by an agent in a clean one-for-one way. Instead, the work of thousands of merchandisers, planners, and logistics coordinators has been quietly rerouted. They now spend their time reviewing what the system suggests, overriding it when local knowledge matters, and feeding back the cases where it got things wrong. The headcount has not collapsed. The job has.

What is automated: inventory rebalancing, demand sensing, distribution routing, waste detection.
What is still human: merchandising taste, vendor relationships, the override layer.
The question it raises: what happens to a 1.6 million person workforce when most of the decisions it used to make are pre-made?

5. Goldman Sachs: The Hybrid Workforce Experiment

Goldman Sachs spent most of 2025 running what it now calls a hybrid workforce experiment. The bank embedded Anthropic engineers inside its own technology teams and built internal agents on top of Claude to handle developer onboarding, code review, and a growing share of internal documentation work. The bank manages roughly 2.5 trillion dollars in assets; the experiment was small relative to the firm but large relative to anything Wall Street had previously tried.

The published results are striking. New developer onboarding is roughly 30 percent faster. Developer productivity across the participating teams is up around 20 percent on internal metrics, with most of the gain coming from the elimination of context-switching tasks - hunting through documentation, writing boilerplate, summarizing meeting threads - rather than from the AI writing production code unsupervised.

What is interesting is what Goldman did not do. The bank did not announce headcount cuts in the affected teams. Instead, it used the productivity gain to absorb a backlog of internal projects that had been sitting in the queue for years. The political read inside the firm is that this was deliberate: prove the model works, build trust, and let the org chart catch up later. The financial read is the same, expressed differently.

What is automated: onboarding flows, code review, internal documentation, meeting summaries.
What is still human: every decision that touches a client or a regulator.
The question it raises: how long can a firm absorb productivity gains into backlog before the headcount question becomes unavoidable?

6. Klarna: The Cautionary Tale

Klarna is the case study that gets cited in every boardroom that wants to slow down. In 2024 the company built an AI customer service agent on top of OpenAI's models, deployed it across its full support surface, and reported that it was handling 2.3 million conversations - the equivalent workload of 700 human agents - within months of launch. CEO Sebastian Siemiatkowski went on a press tour. The stock got a bump.

By late 2025, Klarna had quietly reversed course. Customer satisfaction scores had dropped meaningfully across the channels the agent was handling. An internal audit put the hallucination rate on complex cases at around 5 percent - low in absolute terms, catastrophic in a financial services context where a wrong answer about a payment dispute can trigger a regulatory complaint. The company started rehiring human agents and rebuilt the support stack around a hybrid model: AI handles intake, simple lookups, and resolution of well-understood cases; humans handle anything involving money owed, disputes, or emotional escalation.

The lesson Klarna draws publicly is not that the AI failed. It is that the company misjudged which conversations were actually simple. The lesson the rest of the industry has drawn is more cautious: the easy wins are real, and the last 10 percent is where the franchise lives.

What is automated: intake, FAQ resolution, status lookups, refund eligibility checks.
What is still human: disputes, complaints, anything where the customer is angry or confused.
The question it raises: how do you measure the value of the human in the loop before you remove them?

7. Duolingo: When AI-First Content Backfired

Duolingo announced an AI-first content strategy in early 2025. The plan was to use generative models to produce lesson content at a pace no human curriculum team could match, then layer human review on top. By mid-2025, the pipeline was running at scale. By the end of the year, it was being walked back.

The numbers, eventually reported by the company, were rough. Around 40 percent of AI-generated lesson units contained errors significant enough to require correction - a wrong translation, a clumsy phrasing, a culturally tone-deaf example. User retention on the affected language tracks dropped by roughly 18 percent. The internal post-mortem concluded that the cost of fixing the content after the fact was higher than the cost of producing it more slowly in the first place.

Duolingo has since rebuilt the pipeline as a hybrid: AI drafts, humans edit, and a smaller curriculum team operates as the quality layer. The company is still using AI heavily; the difference is that the model is no longer treated as the writer. It is treated as a faster, less reliable junior.

What is automated: first drafts of lesson content, exercise generation, voice synthesis.
What is still human: curriculum design, cultural review, the final say on what ships.
The question it raises: when AI output is cheap and review is expensive, what is the right ratio of writers to editors?

What the Seven Tell Us

The pattern across these seven is not that automation is winning, and not that it is losing. It is that the companies pushing hardest are recalibrating in public. Medvi keeps pushing. Klarna and Duolingo are pulling back. Walmart and Goldman are holding a careful middle. Cognition and Base44 are betting that the new shape of the company is permanent.

Gartner published a finding in May 2026 that captured the mood. Across the enterprise programs it tracked, more than half of the projects framed as "AI replaces human role" had failed to hit their original targets within twelve months. A separate MIT study put the success rate of fully autonomous agent deployments at somewhere between 5 and 7 percent. The headline number is grim. The underlying number is more interesting: the projects that succeeded almost all did so by redefining the role rather than removing the human, and by treating the agent as a new kind of worker rather than a replacement for an old one.

The fully automated company is not a finish line. It is a moving line, drawn and redrawn every quarter by the people who are willing to take the risk of getting it wrong in public. The seven companies here are not the end state. They are the working draft. Some of them will look prescient in five years. Some will look like Klarna in 2025 - confident, photogenic, and quietly reversing.

What they share is something more useful than a template. They have all stopped pretending that the question is whether to automate. They are now arguing, in front of customers and regulators and their own employees, about what should be automated and what should not. That argument is the actual work. That gap is what Without Human is here to watch.