S02E07: The 100x employee and the AI-native organization
2026 predictions, smallest value-adding unit, agentic companies, shadow orgs, introducing the trillion dollar question
You’re in a board room. It’s December 2026.
The board room belongs to a company called Cynalco Medics. They make a wearable patch called Meridian. It sits on your upper arm and reads glucose, cortisol, and heart rate variability continuously. The hardware is a marvel: clinical-grade sensors, three-week battery, medical adhesive that doesn’t irritate skin. It predicts metabolic crashes and stress events before they happen, then coaches you through them in real time.
The quarterly review is underway. Around the table: CEO, CFO, VP of Product, VP of Hardware, the clinical lead. In the corner of the conference display, a small waveform pulses gently. Everyone calls her Audrai. She’s the company’s agent layer, connected to the operating model, all systems of record, 12 months of meeting transcripts, chat history, the clinical trial database, even customer support tickets.
The CEO asks about Meridian 2.0. The launch slipped by five weeks.
Before anyone speaks, Audrai’s voice comes through the room’s speakers. “The delay wasn’t in the firmware or the companion app. Those were ready by October 15th. The timeline change can be traced to the insight layer. The original recommendation engine was flagging cortisol spikes too aggressively. Users got twelve alerts a day. In early feedback, sixty-three percent said they were ‘overwhelming’ or ‘not actionable.’”
No one asked Audrai to research this. She’d already pulled the context before the meeting started.
The VP of Product nods. “We rebuilt the personalization model. Took three weeks to get the false positive rate down.”
Audrai adds: “And another two weeks for the clinical team to validate the new thresholds. I have the exact handoff logs if you want them.”
In 2024, this scene would have gone differently. The product lead would mention “unexpected complexity.” The hardware VP would note that the sensors were ready on time. Someone would invoke “cross-departmental dependencies” as a catchall. Thirty minutes of corporate choreography, shuffling blame without assigning it.
But it’s the end of 2026, so the facts are on the table.
The CFO leans forward. “Audrai, what’s our current position on unit economics?”
“Better than projected. The extra five weeks let us ship with the refined model. Early cohort NPS is 71, versus 54 for the original version. Support tickets per user are down 40 percent. If those numbers hold, customer acquisition cost recovers in four months instead of seven.”
The room is quiet for a moment.
“So the delay was the right call,” the CEO says.
“The delay wasn’t a call,” Audrai replies. “It was a discovery. You didn’t know users needed a calmer AI until the first build proved the opposite. The time wasn’t lost. It was spent learning what the product actually needed to be.”
The meeting ends eighteen minutes early. There’s nothing left to debate.
That scene isn’t real. Not yet.
But it’s late January 2026, and I’m placing this prediction on record. Apologies for the timing. “Predictions for 2026” posts are supposed to land in December. I’ve been busy building the systems I’m about to describe. More on that later.
Marc Andreessen was right, but early. Software is eating the world, but we haven’t seen nothing yet. We are in the middle of the Youtube moment for software. The amateurs are coming and they are vibecoding patches for every software-shaped hole they can find. This also means a lot of smart coders will find new hard problems to work on, from hardware to ASICs.
The Cynalco boardroom is a projection, maybe twelve months out, maybe eighteen. The specific details will be wrong but I have unwavering conviction that the trajectory is correct. Every component already exists: voice interfaces that feel natural, agent layers that can query operating models, context architectures smart enough to hold months of company history.
Companies that treat this as a 2028 problem may not have a boardroom to sit in by 2027. The gap between “experimenting with AI” and “running on AI” is closing quickly, and the organizations on the wrong side of that gap will find themselves outmaneuvered by leaner, faster competitors who made the switch earlier.
So what has to change between now and the boardroom scene? A few things, in rough order.
First: the smallest value-adding unit shifts from the team to the individual.
The smallest unit
For decades, teams were the smallest value-adding unit of any organization. Specialized contributors created value through coordination. Methodologies mattered because they governed how teams moved together. Agile, Scrum, SAFe: operating systems for groups of humans trying to build software without stepping on each other’s toes.
These methodologies are obsolete, but the consultants don’t know it yet.
Individual contributors who have learned to direct swarms of agents will have a disproportionate effect on an organization’s success. One person with judgment and a well-tuned fleet of AI tools can outpace a ten-person team operating the old way. We’ll see a rise in solopreneurs and it’s going to be a tough time to be an incumbent.
For most of industrial history, individual contributor performance was normally distributed. Most people clustered around the middle. The difference between a good employee and a great one was meaningful but bounded. You could build organizations around averages because averages were predictive.
With AI, that distribution is becoming a power law (there is emerging science to back up this claim).
In a power law, a small number of individuals account for most of the output. The gap between the median and the top widens dramatically. A single person with the right capabilities can now do what used to require a team, a department, a small company. The “10x engineer” was already a cliche. We’re entering the era of the 100x employee.
This isn’t about replacing employees with AI (not yet, at least). I believe we are about to see a new category of employee: the hired gun who shows up with a personal software stack that compounds every task they touch. They fix friction points; automate workflows, and accumulate leverage the way investors accumulate returns. These people become irreplaceable. And the organizations that enable them will pull ahead of those that don’t.
The other side of this distribution is harsher. In a power law, long tails are long. People who refuse to adopt and treat AI as a threat rather than a tool, will find themselves competing for a shrinking pool of roles where the old rules still apply.
The ancestral lizard brain recoils from change by default. Most people will resist AI, and their resistance will cost them. Change management has always been the bottleneck in organizational transformation. Human rewiring is often harder than technical change. But the math is shifting: when your AI-native competitor can ship in a week what takes you a quarter, “we’ll get there eventually” stops being a viable strategy.
Teams without the burden of change management—whether because they’re new or because leadership forced the transformation—gain the most advantage during generational platform shifts. You’re all going to have to ADKAR faster.
What else will change
Pay variance is about to explode. When a single employee can do what used to require a team, compensation models built for normal distributions stop making sense. Companies that figure this out early will restructure equity and ESOP programs to attract and retain high-agency individuals. Those that don’t will watch their best people leave.
Hiring changes. The traits that mattered in a team-centric world (collaboration, consensus-building, “culture fit”) still matter, but they’re insufficient. Agency becomes the filter. Independent thinking. The willingness to build rather than wait for permission. Corporate theater won’t survive against individuals who can demonstrate outcomes in a live context window.
Methodologies won’t disappear, but they’ll change in character. The question is no longer how to coordinate humans. We’ll see new methodologies that govern how humans and agents can productively collaborate in a sociotechnical system. The rituals of Agile were designed for a world where the bottleneck was code. Now that code is abundant, the bottleneck moves to intent, to verification, to discovery.
Explore and exploit
I’ve written before about the twin engines of high-performing companies: optionality and focus. The best organizations don’t alternate between exploration and exploitation. They run both simultaneously; ringfenced bets alongside a tight, disciplined core. Optionality without focus leads to wasted potential. Focus without optionality leads to death by disruption.
AI-native velocity shifts this balance.
When execution speed increases tenfold, the cost of exploration drops. You can run experiments that used to take a quarter in a week. You can kill bad ideas faster and double down on good ones sooner. This means not taking bets becomes riskier than it was before. The companies that double down on their existing playbook without experimenting, will find themselves outflanked by competitors who are testing three new approaches for every one they’re defending.
The shadow org loses power
Every company runs on two systems. The visible one shows up in org charts, roadmaps, and OKRs. The invisible one lives in calendars, hallway conversations, and decisions that are technically reversible but socially settled.
For decades, the invisible system determined how things got done. You learned to read calendars like org charts. You figured out which documents were ceremonial and which ones mattered. You discovered who would block a decision even while nodding in the meeting.
AI changes this. Except maybe in governments and institutions but that’s another story.
When the operating model is expressed as markdown files, you can talk to it directly. You don’t need to “know the right person to invite to the pre-meeting.” When decisions are captured with reasoning attached, institutional knowledge isn’t locked in individual heads. When discussions happen with an AI in the chat, they become more factual and grounded, because the shared context is explicit.
I didn’t expect this when I started experimenting with an operating model in markdown. The speed of iteration increased, yes. But the second-order effect surprised me: company politics started melting away. The AI played a role as neutral arbiter, as long as we agreed on the shared context. Discussions that used to devolve into status games became tractable.
SOPs that existed only on paper started to matter. Not because we enforced them, but because they actively made work easier while making output more consistent. The gap between documented process and real process is shrinking.
The awkward parts
Not everyone benefits from this shift.
The employees who built careers on navigating invisible systems will find their skills devalued. The managers whose authority came from gatekeeping information will discover that agents bypass gatekeepers. The companies that confused activity with outcome will be exposed when competitors demonstrate what lean, high-agency teams can achieve.
Most organizations will do this wrong. They’ll buy AI tools without changing how decisions get made. They’ll train people on prompts without addressing the organizational drag that makes prompts necessary. They’ll automate dysfunction and wonder why they’re not seeing results.
The gap that opens in 2026 runs between companies that adopted AI properly (building discipline as a byproduct) and those that adopted superficially (accelerating waste at scale).
The trillion-dollar question
The blogosphere (well, Substack and X) are abuzz with an interesting debate: how do you build systems that make agents useful inside organizations?
The most seductive answer is “context graphs.” The idea: capture enough decision traces, enough institutional knowledge, enough reasoning, and you can build agents that understand your organization. Some version of this idea is behind nearly every enterprise AI roadmap. Jaya Gupta called it “AI’s trillion-dollar opportunity.” Animesh Akoratana proposed agents as “informed walkers” building world models from their trajectories. Kirk Marple argued we should adopt existing ontologies and focus learning on what’s novel. Gil Feig countered that the graph is the easy part; selection and coordination logic is the product. Parcadei synthesized the debate into a flywheel: ingest, store, resolve, retrieve, serve, capture, compound.
The premise of context graphs is that you can capture what employees know and encode it for agents. But can you?
Part 2 of this series synthesizes the debate and should come out in a week or so.
Part 3 offers my answer, grounded in what we’ve learned building an AI-native operating model in Voxdale (a boutique design and engineering house).
Thanks for writing this, it really clarifys how AI agents could function in practice. I wonder how Audrai adapts to new types of problems, beyond just flagging known issues?