The 2026 AI Index in six numbers a manager can use

In short: six figures from Stanford’s 2026 AI Index matter most to managers: agents now succeed on about two in three computer-use tasks, documented AI incidents rose to 362, industry builds more than 90 percent of notable models, US private AI investment reached $285.9 billion, experts and the public are 50 points apart on AI at work, and Americans trust their government least to regulate it. Each points to a management action, not a technology choice.¹

Stanford’s Institute for Human-Centered Artificial Intelligence released its 2026 AI Index on April 13, covering developments through 2025.¹ Most coverage focused on model capability. For people running AI programs, six numbers say more.

The prior-year figures for incidents, industry’s share of notable models and investment are from the 2025 edition of the Index;² the 12 percent agent baseline is the 2026 edition’s own year-on-year comparison.¹

1. Agents went from 12 to about 66 percent

On a standard benchmark of computer-use tasks, AI agents’ success rate rose from 12 percent to about 66 percent in a year.¹ That is a dramatic improvement, and it still means agents fail roughly one task in three. Plan for capable agents that need review, exception handling and limits, not for agents that can be left alone.

2. Documented incidents rose from 233 to 362

The Index counted 362 documented AI incidents in 2025, up from 233 the year before.¹ More deployment means more failures in public. The organizations that handle an incident well have an owner, a way to switch the system off and a log before they need one.

3. Industry builds more than 90 percent of notable models

Industry now produces more than 90 percent of notable frontier models.¹ For most organizations the strategic question is not whether to build a model but how to choose, contract with and exit from the few companies that do.

4. US private AI investment reached $285.9 billion

By the Index’s count, US private investment in AI reached $285.9 billion in 2025, more than 23 times China’s $12.4 billion.¹ Concentrated investment tends to mean a few large model providers and many application vendors built on them. Vendor viability and exit planning deserve more attention, not less.

5. Experts and the public are 50 points apart

Asked whether AI will improve how people do their jobs, 73 percent of experts expected a positive effect against 23 percent of the public.¹ Your employees and customers are that public. A program that is technically sound can still stall if the people it depends on do not trust it.

6. Americans trust their government least to regulate AI

Among the countries surveyed, Americans had the lowest trust in their own government to regulate AI, at 31 percent, against 91 percent in Singapore.¹ In a market where people do not expect regulation to protect them, they judge each organization’s AI by its own conduct. Clear notice, a human route and visible accountability are worth more in that environment, not less.

What is missing from the six

None of the six numbers measures value captured inside organizations, because no public index can. That figure exists only in your own systems: the baseline you took before a rollout, the output and quality you measure after it, and the costs you actually paid. An index can tell you where the field is moving; only your own measurement can tell you whether your program is moving with it.

How to read an index like this

Three cautions keep the numbers honest. Benchmarks measure tasks in test conditions, so a 66 percent success rate on a benchmark is a ceiling for messy real workflows, not a forecast. Incident counts measure what is reported, so a rise reflects more use and more attention as well as more failure. And survey figures are associations: they tell you what people believe, which matters for adoption, but not what will happen. Our earlier look at McKinsey’s 2025 survey applied the same discipline to adoption numbers.

Using the numbers

Put the first two figures in front of anyone proposing an autonomous agent, alongside the four commitments in Singapore’s agentic framework. Put the third and fourth in front of procurement, with the vendor viability and exit questions from Module 9 of the AIMP program. Put the last two in front of whoever writes your change communications. Each one is a better argument for good management than any adjective.