Where AI Is Going Next

Picture a slide from a strategy deck written in 2023. It is a staircase. At the bottom, Today: models and assistants. One step up, Next: reasoning and multimodal AI. Above that, Then: agents and tool use. Higher still, AI-native workflows, and at the top, AI in physical systems. This one is a reconstruction, not a quote from any company. It was careful, hype-free and, by the standards of its day, sensible.

Its middle steps came quickly. OpenAI’s first publicly available reasoning model, one that works through a problem before it answers, shipped in September 2024¹. A month later a model that could operate a computer by reading the screen, moving the cursor and typing was released, described by its own maker as “still experimental” and “error-prone”². By mid-2026, 40 percent of respondents at organizations with more than 1 billion dollars in revenue told McKinsey they were scaling AI agents, up from 27 percent a year earlier³.

The slide was not wrong about direction. It was wrong about the unit. It staged the future by capability names, and names arrive quickly. What decides whether any of them changes your business is something the slide never measured: whether the capability works reliably enough, cheaply enough and safely enough for a particular piece of your work.

The capabilities are here; the thresholds are not

The argument of this chapter fits in two sentences. Today’s AI already reasons, perceives and acts. What changes next is how reliably, how cheaply, how autonomously and how far into the physical world it does those things.

Those four dials, not a list of product categories, are the trajectory to watch. Each one moves at its own speed, and each one matters only when it crosses a line that your workflow cares about: an error rate you can live with, a cost per task below what the work is worth, a degree of independence your controls can supervise, a physical setting where a mistake is survivable. The executive skill is not predicting when that happens. It is knowing in advance which line matters, watching it, and having decided what you will do when it is crossed.

What is already here

It is worth being exact about the starting point, because many plans still treat it as the future. Three capabilities are in routine production use in October 2026.

Reasoning. Models now spend extra computation working through a problem before answering, which is why their results on mathematics, science and code have jumped. How they do it is covered in Large Language Models and Training vs Inference. Perception. The same systems read images, charts, scanned documents and speech, as Multimodal AI explained. Action. Through tools and connections to other software, models can look things up, fill in forms, run code and complete multi-step tasks, the shift traced in From AI Assistants to AI Agents.

The 2026 AI Index summarizes the result plainly: models now meet or exceed human baselines on PhD-level science questions, multimodal reasoning and competition mathematics⁴. The same report notes that top models read an analog clock correctly only about half the time. That is the jagged frontier described in AI Is Changing Everything: brilliance and brittleness side by side. It is also the clue to where the trajectory goes next. Note what kind of evidence this is. Benchmark scores show capability on a test; they do not show that the same system delivers results in a real business process. That evidence comes only from deployments, and from tests on your own work.

Reliability: from “can” to “every time”

A demonstration shows that a system can do something. A business process needs it to do it every time, or to know when it has not. The distance between the two is the first and most important dial.

The progress is real. On OSWorld, a benchmark of everyday tasks across real operating systems and applications, AI agents went from about 12 percent success in 2024 to about 66 percent in 2026, against a human baseline of about 72 percent⁴. Read the same figure the other way, though, and agents still fail roughly one task in three. OSWorld is still a benchmark: it measures capability under test conditions, not results in a live process.

The research group METR makes the gap more precise. It measures the length of task, in human expert time, that a model can complete. When it first published the method in March 2025, the best model could finish tasks of about an hour with 50 percent success. Asked to succeed 80 percent of the time, the same model managed tasks of only about 15 minutes. Across models, the 80 percent horizon was roughly five times shorter than the 50 percent one⁵.

That ratio is the executive point. Few business processes can run on a coin flip, so the useful measure is not the most impressive thing a system has ever done but what it does on the thousandth ordinary case. Reliability is also a property of the whole system, including retrieval, checks and escalation, not only of the model, as Accuracy, Hallucination and Reliability showed. Expect the next few years to be spent closing this gap, and expect the closing to be uneven across tasks. The question to ask of any workflow is simple: what error rate can it tolerate, and what does one error cost?

Cost: the same capability keeps getting cheaper

The second dial is the price of capability. Here the trend has been steep and consistent.

Epoch AI tracked what it costs to reach a fixed level of performance on a range of tests. Depending on the task and the level, prices fell between 9 and 900 times a year. Reaching GPT-4’s level on PhD-level science questions became about 40 times cheaper each year⁶. The fastest declines were the most recent, so nobody should extrapolate them with confidence. Why AI, Why Now? gave a similar figure from another source: the cost of a GPT-3.5-level answer fell more than 280-fold in two years⁷.

Two cautions keep this from becoming a simple story. First, a falling price per unit of capability does not mean a falling bill. Reasoning models do more computation per answer, and agents call models many times per task, so the cost per completed task can rise even as the price per word falls, as Training vs Inference explained. Second, price is only one line of total cost; The Economics of AI covers the rest.

The practical consequence is still large. A use case you parked because it cost too much per transaction may be viable a year later without any new invention, simply because the price of the capability it needs has fallen past your line. Cost can turn yesterday’s “no” into today’s “yes”, and it is an easy dial to forget to re-check.

Autonomy: longer tasks, same governance question

The third dial is how long and how independently a system can work toward a goal. METR’s measure gives the clearest long-run view.

From 2019 to early 2025, the length of task frontier models could complete with 50 percent success doubled about every seven months; fitted only to models released in 2024 and early 2025, it was about 110 days, or three and a half months⁵. By May 2026, METR estimated that its best-performing model could complete tasks taking a skilled person at least 16 hours, and warned that results above 16 hours are unreliable with its current set of tasks⁸.

Read this carefully. METR itself stresses that the time horizon is not how long an AI can work on its own; its error bars span about a factor of two in each direction; its tasks are mostly software and research; and horizons on visual computer-use tasks are 40 to 100 times shorter⁹. The trend does not tell you when an agent will run your month-end close. It tells you that the ceiling on autonomy is rising fast, and that it is rising fastest in software work.

As the technical ceiling rises, autonomy becomes less a question of what the system can do and more a decision about what you will let it do. That decision is governed by the controls set out in Agentic AI and Autonomous Actions: least privilege, approval for consequential steps and actions that can be undone. Those controls do not loosen because the models improve; the case for them grows as agents take on longer chains of work.

Physical systems: slower, narrower and real

The fourth dial is AI acting in the physical world, through vehicles, robots and machines. It moves more slowly than the other three because errors there are physical, testing is expensive and the law is stricter. It is also no longer hypothetical.

Driverless taxis are the clearest case. Waymo reported 500,000 paid rides a week across ten US cities in March 2026, up from 50,000 a week in May 2024¹⁰. A peer-reviewed comparison over 56.7 million driverless miles found 85 percent fewer crashes involving suspected serious or worse injuries than a human benchmark on the same roads¹¹. In factories, the International Federation of Robotics counted more than 600,000 industrial robots installed in 2025, an 11 percent rise, taking the global working stock to a record 5 million. It credits AI, machine vision and easier programming with raising what robots can do and lowering what they cost to deploy¹².

The pattern is the one to remember: physical AI scales only inside a well-defined operating domain, a set of cities, a factory cell, a warehouse, and widens that domain step by step. Its risk surface is different too. A wrong answer can be corrected; a wrong movement may not be reversible. AI + Robotics, later in this module, covers that ground in depth.

Plan for thresholds, not predictions

If the four dials are what moves, the planning discipline follows. Replace predictions such as “agents will handle our procurement by 2028” with conditional statements: if a capability crosses a stated threshold for a named workflow, then we will take a stated action. A threshold has four parts, quality, cost, speed and acceptable risk, and it belongs to a workflow, not to a technology. The card below is an illustration, not a real company’s thresholds.

Thresholds need a home. A simple radar, adapted from the rings that Thoughtworks has long used for its Technology Radar¹³, sorts every capability you track into one of four positions.

The radar protects against two opposite failures. The first is technology fever: every release enters the roadmap, and the organization reprioritizes constantly without learning anything. The second is complacency: the current stack is declared sufficient, and nobody notices when the economics change underneath it. Both fail for the same reason. Neither has written down what evidence would change its mind.

Two further habits make the radar work. Remember that model releases arrive in months while redesigned work takes years, a distinction drawn in The Speed of AI Adoption. And test the plan against more than one future, which Preparing for AI Uncertainty and Strategic Change develops into a full method.

Story: the feature that was parked correctly

This is a composite drawn from common patterns rather than a single company, and its details are illustrative.

A mid-sized company sells procurement software to manufacturers. In early 2024 its product team tested AI to pull payment terms, renewal dates and penalty clauses out of supplier contracts, a feature customers had asked for. On a test set of scanned contracts, the model missed too many clauses, and the cost per contract was too high for the product’s price point. The team recommended waiting. The leadership team agreed. It was a good decision.

Eighteen months later, during a renewal negotiation, the company’s largest customer mentioned that its own operations team had built contract extraction into its workflow with an AI coding tool, and asked why it was still paying for manual uploads. The renewal closed at a discount. The customer was not unusual: in McKinsey’s 2026 survey, 32 percent of all respondents, not only those at large firms, said their organizations had decided against buying a software product or feature because they could build it with agentic coding tools³.

The post-mortem found no bad decision, only missing machinery. The 2024 test had no written threshold, so nobody could say what result would have changed the answer. The feature had no owner after the product manager moved on. The test set had been discarded, so re-testing would have taken weeks rather than an afternoon. And the company’s watch on AI was a newsletter of model announcements, which told it a great deal about what had been released and nothing about whether its own line had been crossed. The fix took a month: a threshold card for every parked AI idea, a named owner, a kept test set and a quarterly re-test.

What this means for leaders

The future of AI is best treated as four dials moving at different speeds, not as a staircase of product categories. That changes what you ask for. Ask your teams for the reliability their workflows need, the cost per completed task they can afford, the autonomy their controls can supervise and the physical settings where a mistake is survivable. Then ask them to watch the gap between those lines and today’s capability, rather than the stream of announcements.

Check yourself

1. Reasoning models, multimodal AI and agents that use tools are still mostly future capabilities.MythFact
2. An AI agent that succeeds on two-thirds of everyday computer tasks is ready to run a business process unattended.MythFact
3. The price of reaching a fixed level of AI performance has been falling by many times a year.MythFact
4. Falling prices per unit of AI capability guarantee a falling cost per completed task.MythFact
5. METR’s time horizon measures how long an AI system can work on its own.MythFact
6. Physical AI tends to scale inside a well-defined operating domain and widen it step by step.MythFact

Reflection: your parked list

What comes next

Each of the four dials depends, in the end, on the models underneath. Why have they become better at reasoning, cheaper to run and able to see and act, and what does that mean for the choices you make about them? The next chapter, The Evolution of AI Models, looks inside the technology itself.