The Year AI Becomes Infrastructure

Five predictions for the next twelve months in AI

The future arrives as changes in behavior before it arrives as a change in consensus.

People keep waiting for one model release, one benchmark, or one product announcement that will make the AI transition official. But the important shift is already visible in what the most advanced users are doing.

They are not chatting more. They are delegating more.

In The Shift From Chat to Command, I argued that agentic AI changes the unit of work. The prompt becomes the work order. The thread becomes the workspace. The agent becomes the production unit.

In When Cognitive Labor Becomes Abundant, I pushed the argument further. Once agents can work longer, use tools, retain context, and run in parallel, useful intelligence starts to behave like compute. You stop conserving attempts. You launch a portfolio.

Those shifts are still early. The next twelve months will be when they accelerate across the entire economy, with ambitious people being “super spreaders” of AI by demonstrating the incredible power of AI firsthand.

My central prediction is simple: the coming year will be defined by AI becoming ordinary infrastructure for powerful people.

The model will keep improving, but the model is no longer the whole story. The system around the model now matters just as much: memory, permissions, tools, sandboxes, evaluations, connected data, background work, and human judgment.

That system is turning intelligence into labor.

Here are five techno-optimistic predictions for what happens next.

1. The Agent Becomes the Default Interface for Serious Work

Chat will not disappear. It is too natural and too useful.

But chat will stop being the frontier interface for work.

A chatbot waits for the next message. An agent keeps moving toward an objective. It opens the files, searches the sources, calls the tools, builds the artifact, checks the output, and returns with something closer to completed work.

That is a fundamentally better shape for most knowledge work.

The early data already points in this direction. OpenAI reports that by May 2026, 70.2% of sampled individual Codex users had assigned at least one task estimated to require more than an hour of experienced human work. More than a quarter had assigned a task estimated at over eight hours. Inside OpenAI, agents became the primary AI interface across engineering, legal, finance, and recruiting. Non-developer adoption grew fastest of all.

This is not proof that every agent completes every job perfectly. It is proof that user behavior changes when the tool can attempt real work.

Over the next year, the dominant AI interface for serious users will start to look less like a chat window and more like a command center.

You will see the mission.

You will see the active workstreams.

You will see what each agent is doing, where it is blocked, what it changed, what evidence it used, and which decisions require approval.

The interaction model moves from turn-taking to supervision.

That matters because human attention is poorly used when it has to babysit every step. The real leverage appears when the machine can work independently inside clear boundaries, then escalate at the points where judgment matters.

The winning pattern will not be full autonomy. It will be supervised autonomy.

Let the agent research for an hour, but require sources. Let it update the CRM, but not send the final email. Let it prepare the contract, but not execute it. Let it write the code, run the tests, and open the pull request, but keep deployment behind an approval.

This is how trust grows: not through promises, but through legible work.

By this time next year, millions of knowledge workers will measure AI use less by messages sent and more by agent-hours directed. The meaningful question will no longer be, “How often do you use AI?”

It will be, “How much useful work do your agents complete each week?”

That is the first prediction.

AI moves from something you consult to something you manage.

2. Small Teams Begin Producing Institution-Scale Output

The minimum viable team is about to shrink again.

The internet let a small company reach a global market. Cloud computing let it rent infrastructure that once belonged only to large firms. Software-as-a-service let it assemble an operating stack without building every system itself.

Agents add the missing layer: on-demand cognitive labor.

A founder can now assign market research, customer analysis, product prototyping, financial modeling, contract review, sales operations, support analysis, and content production in parallel. The work will not be flawless. Neither is work produced by a hurried junior team. The difference is that the founder can run more attempts, compare more paths, and enter more domains before hiring a full department.

This does not make people irrelevant.

It makes ambition cheaper.

Many valuable projects never begin because the coordination cost arrives before the proof. You need the analyst before you know whether the market is real. You need the engineer before you know whether the product works. You need the designer before users can react. You need the operations person before the process has volume.

Agentic labor changes the order.

You can build the first system, run the first analysis, create the first prototype, and test the first workflow before the organization exists. You can reach evidence sooner. Then you can hire around reality instead of hiring around a guess.

That is extraordinarily pro-founder, pro-creator, and pro-experiment.

Over the next twelve months, we will see more one-person and five-person firms produce work that looks like it came from teams five times their size. Some will be software companies. Others will be research shops, investment firms, education businesses, media companies, consultancies, and strange new hybrids that do not fit an old category.

The key is not headcount reduction. The key is output expansion.

The best small teams will not use agents to do the same amount of work with fewer people. They will use agents to pursue opportunities that would previously have been ignored. More customer segments. More product experiments. More research. More languages. More personalized service. More follow-through on the long tail of worthwhile tasks that never survived the weekly priority meeting.

This will also change what it means to be employable.

The valuable operator will be the person who can cross boundaries. Someone who understands the customer well enough to direct research, understands the business well enough to judge the analysis, understands the product well enough to shape the build, and understands quality well enough to reject polished nonsense.

The best specialists will gain more reach.

The best generalists will gain more force.

By August 2027, the idea that a tiny team can operate with the functional range of a much larger company will no longer sound like an internet slogan. It will be an observable operating model.

The individual will not replace the institution.

The individual will gain institutional leverage.

3. Software Becomes Personal, Abundant, and Disposable

Most people still think software is something a company builds and a customer buys.

I see something major happening in SaaS.

When the cost of writing software falls far enough, much of it will be created for one company, one team, one person, or one moment. It will solve a narrow problem, do the job, and perhaps never become a product at all.

This is software as a consumable.

An analyst will build a custom research pipeline for one investment question. A sales leader will create a territory-planning tool around the exact logic of her business. A teacher will generate a simulation for one class. A scientist will turn a fragile notebook into a tested workflow. A family will build a small system to coordinate care for an aging parent.

None of these projects needs a venture-backed startup behind it.

They need an objective, context, and an agent that can translate intent into a working tool.

The evidence is already moving beyond professional developers. Anthropic’s analysis of roughly 400,000 Claude Code sessions found that people across major occupations achieved verifiable success on coding tasks at nearly the same average rate as software engineers, although deeper domain expertise still improved results. OpenAI reports that knowledge workers are using Codex to build lightweight tools that previously required engineering help.

This does not mean engineering disappears. It means software demand explodes.

When spreadsheets became accessible, the world did not need less financial analysis. It produced far more of it. When cameras moved into phones, society did not decide it had enough images. It created trillions more.

Software will follow the same pattern.

We will make more tools because tools become cheaper to make. We will automate smaller annoyances. We will instrument more workflows. We will test ideas that were never valuable enough to justify a development queue.

Professional engineers move up the stack. They design platforms, review architecture, secure systems, maintain shared infrastructure, and solve the hardest technical problems. Domain experts move down the stack just far enough to express their knowledge in software.

That meeting point is where the abundance happens.

Over the next year, “build versus buy” will gain a third option: generate.

Buy the stable system of record. Build the strategic product. Generate the narrow tool that connects the two, handles the edge case, or removes the recurring frustration.

The result will be messy. There will be too much software, weak software, and tools that should never escape a sandbox. Every abundance curve produces waste.

But productive waste is part of progress. Cheap attempts let us discover what deserves to last.

By next summer, a growing class of people will no longer accept a repeated digital task simply because no software product exists for it. They will describe the need, generate the tool, test it, and move on.

The distance between frustration and automation is collapsing.

4. The Organization Becomes a Learning System

The first phase of enterprise AI was seat distribution.

Buy licenses. Turn on access. Train employees. Count usage.

The next phase is organizational memory.

A capable model without context is a smart stranger. It does not know why the company made a decision, how a customer should be treated, which exception matters, what quality looks like, or where the process tends to break.

The company that supplies that context gains an enormous advantage.

Over the next twelve months, the most advanced organizations will stop treating agent use as a collection of personal tricks. They will begin capturing the operating system of the firm.

Instructions become skills.

Past decisions become memory.

APIs become tools.

Approval rules become permissions.

Quality standards become evaluations.

Repeated work becomes a monitored workflow.

This is more than automation. It is the conversion of tacit knowledge into reusable infrastructure.

Today, much of a company exists only inside people’s heads. The real sales process is not the process map. The real close procedure is not the checklist. The real customer escalation path is not the policy document. It is a network of habits, exceptions, relationships, and judgment calls carried by experienced employees.

That makes organizations fragile.

Agents create a reason to make the hidden system explicit. To delegate a workflow, you have to define it. To evaluate the output, you have to state what good means. To grant access, you have to understand which permissions the work actually requires.

The act of making a company legible to agents can make it more legible to humans.

This will produce a new management discipline. Call it agent operations, workflow architecture, or machine-labor management. The name is less important than the function.

Someone must decide which work should be delegated, which context should persist, which actions require approval, which outputs need independent verification, and how lessons from one run improve the next.

Reliability will be the forcing function. Microsoft Research’s CORPGEN work shows how sharply current agents can degrade when they must manage many interdependent tasks. It also shows the direction of improvement: stronger planning, isolated memory, and learning from experience.

The lesson is not that agents are unusable.

The lesson is that the harness matters.

By August 2027, the best companies will have internal libraries of agentic workflows that improve with use. A good investigation, onboarding process, account review, forecast, or product launch will not vanish after the employee completes it. The method will remain available for the next person and the next agent.

Institutional knowledge will start to compound instead of evaporate.

That is a profound upgrade to the firm.

5. AI Begins to Expand the Rate of Discovery

The most important AI story of the next year may not come from an AI company.

It may come from a laboratory.

Science is full of cognitive bottlenecks. Researchers must read immense literatures, clean data, maintain old software, translate between disciplines, design experiments, interpret results, and decide which hypothesis deserves the next scarce dollar of lab time.

Much of that work is necessary.

Not all of it requires a human to perform every step.

The early systems are already moving from summarization toward participation. Google DeepMind’s Co-Scientist uses multiple agents to generate, debate, and refine scientific hypotheses. A recent OpenAI field report on scientific computing describes small research teams using coding agents to modernize fragile scientific software, speed analysis, and attempt projects that would otherwise have required much more engineering capacity.

The mechanism matters.

AI does not need to become an autonomous Nobel laureate to accelerate science. It needs to remove enough friction that researchers can run more intellectual experiments before committing to physical ones.

One agent maps the literature.

One looks for conflicting evidence.

One reproduces the analysis.

One proposes mechanisms.

One attacks the hypothesis.

The scientist decides what is real, what is novel, what is safe, and what deserves to be tested.

That division of labor is not a retreat from human science. It is an expansion of scientific agency.

Over the next twelve months, we will see the first visible wave of research programs built around continuous agentic support. Not isolated demonstrations. Operating workflows.

The near-term wins will come from shortening the loops around discovery: better literature synthesis, faster code, cleaner data, stronger replication, improved experiment design, and more systematic exploration of candidate explanations.

Some of the most meaningful gains will happen in places that receive too little attention. A small biology lab will maintain software it could never afford to rebuild. A rare-disease researcher will search a wider hypothesis space. An engineer will model more materials. A social scientist will test robustness across more datasets. A clinician-researcher will connect evidence that sits in separate literatures.

There will be errors. Scientific verification remains hard because reality does not care how persuasive an output sounds. Expert judgment, reproducibility, peer review, and physical experiments remain the court of appeal.

But this constraint makes the opportunity clearer.

AI makes cognitive exploration cheap while the real world provides the test.

The result is not instant truth. It is more shots on goal for truth.

By next summer, the strongest case for AI optimism will not be that a model scored higher on another exam. It will be that real researchers used agentic systems to save months of work, revive neglected projects, and move promising ideas into the world faster.

That is where this becomes larger than productivity.

Productivity gives us more output.

Discovery gives us more future.

The New Abundance Behavior

These predictions share one mechanism.

The cost of a useful attempt is falling.

When attempts are expensive, people conserve them. They wait for permission. They narrow the question. They pick one path. They tolerate broken workflows because fixing them costs too much. They leave ideas unexplored because the team is already full.

When attempts become cheap, behavior changes.

You test more paths.

You build the missing tool.

You investigate the strange result.

You personalize the service.

You start the company.

You run the experiment.

This does not remove scarcity. It moves it.

Judgment becomes scarce. Taste becomes scarce. Trust becomes scarce. Courage becomes scarce. The ability to choose a worthy objective becomes scarce.

That is good news for people.

The bleak framing of AI assumes human value came from the friction of producing the first acceptable draft. It did not. Our highest value comes from deciding what should exist, understanding other people, taking responsibility, setting standards, finding meaning, and acting when the map is incomplete.

AI makes those human functions more important because it gives them more leverage.

What To Do Over the Next Twelve Months

Do not wait for these predictions to become consensus.

Start operating inside them.

Replace one recurring chat with a complete work order. Give the agent an objective, context, tools, constraints, and a definition of done.

Turn one repeated process into a reusable skill. Capture the instructions. Save the references. Define the review points. Make the second run better than the first.

Run one portfolio instead of one attempt. Ask multiple agents to explore different paths, then apply your judgment to the comparison.

Build one piece of software that no vendor would ever build for you. Solve the narrow problem. Feel what happens when intent becomes executable.

Choose one domain where you have real expertise and push an agent beyond generic assistance. Make it work inside your standards, your data, and your actual operating environment.

The goal is not to use more AI.

The goal is to create more agency.

That is the optimistic case for the next year. Not a machine that does everything while humanity watches. A new labor layer that gives more people the capacity to build, learn, discover, and act.

The chat era made intelligence easier to access.

The agent era makes intelligence easier to deploy.

Over the next twelve months, deployment becomes infrastructure.

And infrastructure changes who gets to build the future.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏