A misty valley at dusk with golden lights glowing from scattered houses, surrounded by dark rolling hills and distant mountains under a deep blue sky streaked with clouds.

When the Machines Don’t Sleep

Preparing for the age of AI-only firms that never clock out

written by
When the Machines Don’t Sleep: an image by

Bottom Line Up Front (BLUF):

AI-only firms—companies run entirely by autonomous agents with no human employees—don’t yet exist, but the technology, capital, and coordination models are rapidly aligning to make them inevitable. Their advantage won’t just be lower cost; it will be Speed to Value—turning ideas into outcomes faster than any human organization can match. For leaders in regulated, safety-critical industries like utilities, the challenge isn’t to panic but to prepare: become truly AI-first by rebuilding core workflows around agents, fortify human strengths like empathy and stewardship, and position your enterprise as the trust anchor and ethical backbone in a coming economy where the machines don’t sleep.

When the Machines Don’t Sleep

It was a school‑morning kind of chaos—the marine layer pressed low over the San Gabriels, waffles toasting, one kid hunting for a missing shoe like it had fallen into a black hole. Lunchboxes were open, water bottles lined up like a tiny assembly line where each kid was making their lunch. In the middle of it, I glanced at my phone and saw the headline that snapped me awake: “A fully autonomous, AI‑only private insurer launches with 50% lower premiums and settlement times measured in minutes.” No employees. No night shifts. No HR. Just agents talking to agents, code shaking hands with code, and customers getting answers at 6:07 a.m. while I’m trying to get the next generation off to school.

“Dad, do robots have to go to bed?” one of my kids asked.

“Some do,” I said, thinking of scheduled maintenance windows. “But the smartest ones? They don’t sleep much.”

He nodded like that was both fascinating and unfair—right up there with screen‑time limits. And I felt the double pull I think a lot of parents feel: pride that my kids will grow up in a world of incredible tools, and a quiet ache that the world will ask more of them, sooner. We’re not just managing systems at work; we’re raising people who will inherit whatever we automate, streamline, and unleash.

The headline kept needling me as I made my way to my home office. If an AI‑only firm can drop into a trust‑sensitive market like insurance with half the cost and round‑the‑clock responsiveness, what happens when that same kind of company shows up in our own value stream — or perhaps someplace up and down the value chain around us — faster than we can convene a change advisory board? It’s one thing to refactor workflows; it’s another to explain to your child what it means when the machines don’t sleep and the markets don’t wait.

Somewhere between “Have a great day!” and “Don’t forget your team project,” the thought arrived: This isn’t just about new software. It’s about new kinds of ways of doing business. We’ve talked for years about becoming “Digital Company.” But this headline pointed beyond that frontier—toward AI‑only firms that operate without a single human on payroll. That may sound like science fiction. But plenty of yesterday’s sci‑fi now sits in our pockets wrapped around a cellphone. And for those of us in IT inside a regulated, safety‑critical business, and for those of us raising kids in the middle of it, the question isn’t whether to panic. It’s whether to prepare—wisely, calmly, and early—so our people and our children’s future can flourish in what’s coming next.

From “Frontier Firm” to “No‑Humans‑in‑the‑Loop”

In April 2025, Microsoft put a name to what many of us were building toward: the “frontier firm,” an AI‑first organization where AI agents execute core outcomes and humans orchestrate, oversee, and safeguard. The forecast was bold: most organizations would begin that transition within five years. But trajectories don’t stop at their first plateau. Beyond the frontier, another shape is taking form—a firm made entirely of coordinated AI agents, no human employees in the loop. Think of a hierarchical network of hyper‑specialized agents—strategy, marketing, IT, operations—each cluster led by a “lead agent” that allocates work and resources by spinning agents up or down as needed.

At the apex sits an AI CEO, charged with setting strategy and owning outcomes. Early studies of AI‑only teams suggest productivity peaks in designs that emphasize specialization with tight coordination. Over time, these firms may abandon human‑inspired org charts altogether. For now, the best designs still rhyme with structures we recognize—just without us on payroll.

To feel the stakes, put yourself in the shoes of that insurance executive. By noon, the AI‑only entrant has already converted curiosity into customers. Their premiums are 50% lower, their service is 24/7/365, and claims settle in minutes. The market reacts. Your stock slides. You don’t have months; you have days.

No one can yet tell whether AI‑only firms are five years out or fifteen. But given the scale of potential impact—competition, jobs, and humanity’s hands on the economic wheel—CEOs and IT leaders should plan now. The assignment isn’t to surrender. It’s to reimagine ourselves as AI‑first players ready to compete with what comes after.

A Wave of Investment Is Clearing the Runway

Let’s say the quiet part plainly: a true AI‑only firm does not yet exist. But the capital flowing into agentic and autonomous AI says the smart money sees disruption beyond today’s AI‑first phase. Venture funding for AI hit roughly $116 billion in H1 2025, surpassing all of 2024’s $101.5 billion. More than 100 startups describing “autonomous capabilities” have launched since 2022, backed by 120 lead investors, raising about $41 million on average.

Roughly half are building agentic “replacements” for specific roles; more than 10% target automation of entire teams or departments. This isn’t just slideware. This isn’t demoware. This is products getting deployed into production, and producing value.

The technology curve is tilting. When OpenAI released its first reasoning model in September of 2024, it marked a significant milestone by being able to to stay on task for 22 minutes 50% of the time. By September 2025, that task‑length jumped to more than 30 hours for software development, blowing past the human workday, and nearing a human work week before needing a break. The line is moving, with an expectation that sometime between 2028 and 2029 they will be able to work on tasks for a full month without needing a break.

Are there hurdles? Absolutely. Reasoning on messy multi‑step business processes, multi‑agent coordination, and external interactivity (from APIs to actual robots) are still hard problems. But the progress rate—and the betting—suggest these challenges won’t stay insurmountable.

But this isn’t anything unusual—this is like building character and the marshmallow test. The lesson we’re watching play out in AI isn’t that different from what psychologists once observed in children staring down a single marshmallow. The instructions were simple: wait fifteen minutes (for human children, but AI can not wait 30 hours), and you’ll get two. The challenge wasn’t comprehension; it was endurance. The same applies here. The path to truly autonomous firms is clear—we can see where this is going—but getting there demands patience, discipline, and the ability to hold steady while others rush in. Each incremental advance—longer task windows, more coherent reasoning, better coordination—tests our collective ability to delay gratification for the bigger reward: a world where machines don’t just assist us, but operate alongside us with purpose and precision. The question, as always, is whether we can wait long enough to earn the second marshmallow.

The Hurdles Between Here and “All‑Agent”

1) Reasoning Capabilities

Many core processes—lead generation, analysis, underwriting, outage triage—are chains of sub‑decisions requiring different skills. Consider an outbound sales agent: identify prospects, prioritize them against target profiles, pick the time to reach out, personalize the message, decide when to stop. The trajectory here is positive. Leading models like OpenAI’s GPT‑5 and Google’s Gemini 2.5 now show reasoning abilities equivalent to humans with an IQ of ~130. They plan and execute longer, more complex work. And the “task length” limit has been doubling every three to seven months, with signs of acceleration.

2) Coordination and Orchestration

Real firms are ensembles. An AI‑only firm’s lead agents must coordinate tasks and information flows across teams, resolve conflicts, and scale agents up or down with demand. Progress is real: recent studies show promise with orchestrator/evaluator agents and agent voting for conflict resolution. A wave of startups claim full team/department replacement using these architectures.

3) Memory and Knowledge Systems

Enterprises run on vast, living knowledge—client data, legal records, internal documents. AI-only firms will need infrastructure that stores, retrieves, and applies real-time and historical data accurately. Once limited by fixed context windows (e.g., ~100,000 words for GPT-4), research now shows theoretical “infinite” context is possible. Knowledge graphs add structure and lineage. And at enterprise scale, tools like Microsoft’s Semantic Index already handle tens of millions of documents—plug-and-play scaffolding for AI-only knowledge.

But raw data alone isn’t enough. Metadata hygiene, document cleansing, and records management become the quiet backbone of intelligence. Every mislabeled file, duplicate record, or stale version that tells the truths from years in the past erodes trust in the system’s reasoning. AI-only firms will need disciplined pipelines that tag content consistently, remove redundancy, normalize formats, and enforce retention rules—turning chaos into clean signal. The firms that master this invisible work won’t just have better data; they’ll have sharper judgment, faster insight, and the ability to act on truth at scale.

4) External Interactivity

AI‑only firms must participate in markets alongside (human or AI) customers, partners, regulators, and service providers—by email, phone, forms, payment rails, CRMs, legal and tax portals. The Model Context Protocol (MCP) has emerged as a USB‑C for AI apps standard, letting agents plug into applications, databases, and prompt libraries.

Physical‑world execution is harder. Robotics are widespread in industrial logistics—about 75% of Amazon’s orders are fulfilled by ~750,000 robots—but many tasks requiring fine motor skills and adaptive judgment still need humans. AI‑robot protocols must mature: goal‑level commands, real‑time feedback loops, multimodal sensors. Expect AI‑only firms to appear first in digital‑native industries (software, algorithmic trading, digital marketing). In kinetic domains, we’ll likely see hybrids: AI agents decide and coordinate; humans and semi‑autonomous robots execute.

5) Regulatory and Societal Hurdles

No jurisdiction currently recognizes AI as legal persons. AI can’t be a CEO with fiduciary duties, sign binding documents, or sit as a director. Some governments are probing the edges: the UAE permits AI board observers as part of its 2031 AI Strategy; a US bill has been introduced that could qualify certain medical AI systems to prescribe drugs. In the near term, human founders will create AI‑only firms, define mission, and assume fiduciary and signatory roles.

Society’s pulse matters, too. As with autonomous vehicles, people may resist ceding control—especially in trust‑sensitive sectors like healthcare and finance. Fears about job displacement, privacy, and loss of human control could spark boycotts or political headwinds, slowing adoption regardless of technical readiness.

What Advantage Would an AI‑Only Firm Have?

Speed to Value

Shifting work from human‑only teams to AI agents isn’t just a cost play—it’s a cycle‑time compression play. When outcomes are produced by compute, value arrives sooner: ideas move to prototypes, to tests, to production in hours or days instead of weeks. The spend profile shifts to energy, infrastructure, and compute, and that matters because those inputs can be provisioned instantly and scaled in parallel. A16Z notes AI startups allocate 80%+ of capital to compute—fuel for more experiments, run concurrently, which shortens the distance from hypothesis to customer impact. As models get cheaper and more efficient—roughly 10× price/perf gains per year driven by 30%–40% annual hardware improvements, price competition, and model commoditization—teams can iterate more often and ship value faster on the same budget. The headline isn’t “lower cost”; it’s more outcomes per unit time.

You can see this in everyday tooling. GitHub Copilot Enterprise (~$50/month) and OpenAI’s Codex (~$200/month) can be provisioned in minutes and, according to some benchmarks, produce better code than most humans. That pencils out to ~$0.05–$0.27/hour, but the real advantage is not the rate—it’s the start time. No hiring queue. No on‑ramp. Teams get time‑to‑first‑commit in minutes, time‑to‑test the same day, and time‑to‑release on a cadence that matches business demand.

For incumbents, treat this like moving into the HOV lane for outcomes: measure and manage Speed to Value directly—time‑to‑first result, time‑to‑resolution, lead time for changes, and time‑to‑customer impact. AI‑only firms will make these their default operating metrics. Match them there, and you don’t just spend less—you arrive sooner.

Customer Experience

Compute doesn’t sleep. AI‑only firms would deliver 24/7/365 responsiveness. Work that takes humans hours can shrink to seconds—some AI‑first insurers report 50× reductions in claims processing times. Every interaction becomes a training datum; with perfect recall across agents, service variance drops and quality ratchets up.

Over time, expect hyper‑personalization at the product level—micro‑offers tailored on demand—and even anticipatory experiences. Walmart’s shopping agent, Sparky, is an early hint, generating personalized meal plans and auto‑adding ingredients to your cart—nudging the age‑old “what’s for dinner?” before you ask.

Adaptability

Most companies struggle to adapt: slow planning cycles, under funded efforts, silos, principal‑agent gaps, and human resistance. An AI‑only firm is fully digital labor. It can copy, deploy, or retire agents instantly to match demand. And when the CEO‑agent updates policy or structure, the change propagates immediately with perfect alignment—no misinterpretation, no passive resistance. Agents update continuously with new capabilities and patches.

Strategy itself could become weekly, not annual—driven by always‑on, parallel simulations that test pricing and marketing tactics and select statistically optimal short‑term paths. Reaction times to market moves collapse—copying innovations or countering offers nearly in real time.

Preparing for the New Kind of Reality

So how should leaders—especially those of us at the Utility respond without losing our center? Three moves, in order.

1) Become Genuinely AI‑First

You can’t contend with AI‑only rivals by leaning only on human strengths. First, narrow their edge by making AI your operating system, not a bolt‑on. Don’t just add AI to everything, rewire your processes to be AI Driven. Do this in several domains—IT, marketing, customer service, and all internal ops areas. Process‑wide deployments of agents that understand your processes, data, and applications are feasible today and already pays off. Redesign your business flows so that agentic AI does the research, drafting, and scheduling at machine speed, while human reps orchestrate agents and focus on the emotionally nuanced, trust‑heavy outcomes with external parties—where, for now, those parties show a preference for human contact.

Few organizations are truly AI‑first yet. Those farthest along restructure work itself—redesigning end‑to‑end workflows around autonomous agents and embedding continuous learning into the fabric, rather than sprinkling bots and agents over legacy applications and processes. Three leadership imperatives follow:

  1. Name the AI‑first North Star. Identify the processes most advantaged when redesigned around agentic AI.
  2. Launch end‑to‑end transformations of the highest‑impact workflows—don’t nibble the edges. Make sure these processes are fully reimagined for agentic execution, not just overlaid with tools, and measure success by time‑to‑value and safety outcomes. This means new deliverables, new metrics, new governance, new skills, and new rhythms.
  3. Scale the wins across the organization, compounding learned patterns by building similar agents at either end of the now reworked process.

There’s also a greenfield option: set up a dedicated organizations that internally replicates the slices of your org best suited to agentic transformation, then scale up from there. Make sure this team acts with empathy, agility, and close partnership with the core business. You are looking for a priest not a cowboy. The goal is to learn fast and transfer those lessons back into the mainline organization.

2) Double‑Down on Uniquely Human Work

Yes, today there are domains where humans still hold an edge—complex reasoning in unfamiliar contexts, deep domain judgment. But those moats will narrow. Don’t retreat to “things AI can’t do yet.” Instead, focus where new constraints appear because AI is succeeding.

These constraints won’t last forever. The cycle will. As AI advances, new bottlenecks emerge. Our job is to keep asking: What new constraints have the latest AI developments created in our industry? And: Which ones invite uniquely human contribution right now?

3) Become a Critical Node in the AI‑Only Ecosystem

Don’t view AI‑only firms solely as enemies. There is value in competition and cooperation. Incumbents can position themselves as irreplaceable partners that solve the constraints AI‑only players can’t (or shouldn’t) shoulder.

  • Physical Supply Chain Integration. AI‑only firms still need kinetic execution—factories, fleets, field operations. Utilities, manufacturers, and logistics pros can be the execution layer, integrating AI intelligence with real‑world workflows: think manufacturing, fulfillment, field service, and grid‑aware scheduling.
  • Trust and Relationship‑Building. Relationship capital—earned over decades—doesn’t materialize from a freshly spun agent network. One study showed 89% of executives say in‑person meetings seal deals (Harvard Business Review Analytic Services, 2018). Incumbents can act as relationship brokers for AI‑only firms in B2B contexts. On the consumer side, we can lend brand equity and even underwrite commitments, acting as the human face and guarantor for otherwise machine‑driven ecosystems.
  • Ethical Stewardship. As AI‑only firms touch more people, expect demand (and mandates) for governance. Incumbents can offer ethics‑as‑a‑service: fairness audits for finance agents, oversight boards, or clinician review platforms for health decisions—keeping the license to operate intact.

What This Means at at the Utility

We live in a world of non‑negotiables: safety, reliability, regulatory compliance, customer trust. That doesn’t insulate us from AI‑only rivals; it shapes how we respond.

  1. Prioritize agentic wins at the seams. Outage communications, field‑crew logistics, materials forecasting, vendor billing—all are ripe for end‑to‑end agentic redesign. Don’t sprinkle tools; restructure workflows.
  2. Tie AI workloads to grid realities. If AI‑only firms’ costs are compute‑heavy and compute is energy, then energy strategy is their business strategy. Align data center demand, demand response, and scheduling with our operational rhythms.
  3. Treat knowledge as a first‑class asset. Build retrieval‑ready corpora (policies, procedures, historical tickets, safety lessons learned) designed for agent consumption. Move beyond “document storage” to knowledge graphs, meta data and enterprise‑scale semantic indexing.
  4. Stand up orchestration patterns early. Adopt MCP interfaces for internal systems so agents can plug in safely. Use orchestrator agents and evaluator patterns to keep human oversight where it matters.
  5. Codify governance and ethics. Publish an AI Operating Manual: where agents run autonomously, where a human‑in‑the‑loop is mandatory, and how exceptions escalate. Think “unit tests as seatbelts.” Cheap to use, priceless when something jerks the lane.
  6. Invest in human craft. Train our people for roles that grow in value as AI scales: curation, evaluation, escalation, relationship stewardship, and the human touch in moments that matter. Celebrate “human‑made” where it adds meaning.

The Broader Question We Can’t Outsource

The firms that thrive will track AI progress closely, pilot agentic tech ahead of peers, and start honest conversations about where human capabilities now create the most value—whether in competition with, or partnership alongside, AI‑only players.

But step back and ask the larger question: Do we want AI‑only firms to become a reality? There will be real consequences—job displacement at scale, risks of privacy erosion and inequality, and even the specter of catastrophic AI misalignment. We could lose meaningful human control over swaths of the economy and be forced to rethink the role of humans in work and society.

And yet, the incentives are enormous. The world’s most valuable companies are staking their future on AI agents and the models and hardware that power them. Geopolitics add urgency: sitting out may simply cede advantage to other nations more willing to embrace AI‑only firms.

Because autonomous agents will likely take root first in corporate contexts, business leaders—you and me—carry a unique responsibility. We can foster a debate that includes legal, sociological, and ethical dimensions—not as a brake on innovation, but as a steering wheel. That posture feels, to me, like stewardship: strength and restraint held together.

A Field Guide for Leaders at the Utility in the Next 18 Months

  1. Name three end‑to‑end workflows to rebuild around agents (not tools). Charter teams. Set outcome metrics tied to time‑to‑value and safety.
  2. Stand up an agent platform with MCP connectivity, orchestrator/evaluator patterns, and observability baked in.
  3. Refactor knowledge. Create a prioritized backlog of corpora to move into retrieval‑ready formats; link to Semantic Index or equivalent at tens‑of‑millions scale.
  4. Define the governance line. Publish where agents can act autonomously, where human‑in‑the‑loop is mandatory, and how to halt a process safely.
  5. Quantify compute as cost of goods. Treat energy and compute like a single portfolio. Model scenarios assuming 10× annual price/perf improvements.
  6. Launch human‑craft training. Curation, hypothesis evaluation, escalation judgment, crisis empathy, relationship stewardship. Celebrate excellence publicly.
  7. Establish external roles. Where can we be the physical execution layer, the trust anchor, or the ethics steward in an AI‑only ecosystem? Pilot one partnership in each category.
  8. Stress‑test against an AI‑only entrant. Pick a sensitive business function (e.g., customer claims or vendor onboarding). Simulate a rival with 50% pricing and instant service. Decide now what you’d change in week one.
  9. Invite the broader conversation. Convene a quarterly forum with regulators, community leaders, educators, and labor to surface risks, opportunities, and responsibilities.

What We Keep—and What We Let Go

We keep safety as the first sentence. We keep reliability as the second. We keep the humility that comes from running critical infrastructure for neighbors we actually see in the grocery store. We keep our commitment to the people who do the work, from field crews to analysts to the folks who answer the phone at 2 a.m.

What we let go is the illusion that sprinkling AI on the edges or on top of things will suffice. It won’t. If AI‑only firms bring a new physics of business then we must become AI‑first with the same seriousness we bring to wildfire mitigation or grid hardening. That doesn’t mean we become unrecognizable. It means we become truer to our purpose: serving people with resilience, stewardship, and care—only now with new tools, new rhythms, and new partnerships.

Back Where the Morning Began

That evening, after homework and the soccer cleats were finally parked by the door, we did the small things that make a family hum. The dishwasher took its quiet shift. A permission slip got signed. I read the last chapter of a book out loud while a small head leaned against my shoulder and tried not to fall asleep before the hero made it home. “Dad,” came the muffled question from the pillow, “will AI take your job?”

I smiled in the half‑dark. “My job is to love my neighbor,” I said, “to help keep the lights on, and to teach you to grow up and become a better parent than I am. The tools we use will change. That the other parts won’t.”

Somewhere, an algorithm is grinding through a night that never ends. Somewhere else, a lineman is checking a sling, a dispatcher is rerouting a crew, a customer is waiting for restoration, and a manager is deciding whether this was the month to upend an old process and try something new. Our kids will grow up in a world where machines rarely sleep. But the call on us—the people who build, secure, and steward the systems—remains: to choose wisdom over hurry, courage over cynicism, and service over spectacle.

We don’t get to choose the pace of technology. We do get to choose our posture. So here’s mine, and maybe it can be ours: move toward the frontier with open eyes and steady hands. Build the agentic roads inside our house. Strengthen the human work that gives meaning to all the rest. Partner where partnership multiplies good. And keep stewardship at the center—because the grid is more than electrons, the economy more than code, and a home is more than the sum of its devices.

Lights out. A last check on the door. A whispered prayer for wisdom and safety for ourselves and for the people we serve—neighbors, customers, and the small ones asleep down the hall. In a world where the machines don’t sleep, we can still practice the ancient rhythm: we prepare, we practice, and then, together, we empower each other.