AI is in every headline. In real businesses, it is 1 in 4.

In the 1990s, linking two business systems was a big-company project. TIBCO became its own company in 1997 just to pass messages between large systems. In 2000, Roy Fielding described REST, the pattern almost every web app still uses to talk to others.

The 2010s handed the job to people who do not code. Gartner was writing about iPaaS by 2011. Zapier launched in 2011 to 2012, Integromat (now Make) in 2016 and n8n in 2019. Zapier went from 2,000+ connected apps in 2021 to 9,000+ in 2026.

Then AI arrived. ChatGPT launched on 30 November 2022, and function calling (June 2023) let models operate software. Anthropic open-sourced MCP on 25 November 2024. OpenAI adopted it four months later and Microsoft within six. In December 2025 it moved to the Linux Foundation. By our count, the gap between big waves fell from about 14 years to about 2.

For a century, the time it took a new technology to go from 10% to 50% of US homes kept falling: 43 years for the telephone, about 5 for social media. The trip from half to almost everyone never sped up. The internet, PCs and social media had not reached 90% when the data series end (2016 to 2019).

Businesses move slower than people. E-commerce was 0.6% of US retail in 1999 and is 17.1% today. COVID pushed it to 16.3% in mid-2020, and it has barely moved since. Cloud took 19 years to reach half of EU firms.

AI is on both curves at once. People took to it fast: weekly ChatGPT users went from 100 million in November 2023 to more than 900 million by February 2026. Businesses did not: under a quarter of US firms use AI, almost four years after ChatGPT.

Trust the government number. In late August 2026, 23.8% of US businesses used AI, up from 17.3% in November 2025 (US Census Bureau). 27.6% expect to use it within six months.

Other surveys count other crowds: 56.1% of the tech-forward firms that use Ramp cards pay for AI, 46% of US small employer firms use it in some way and 88% of McKinsey's global survey respondents do. Always ask who was counted.

Small firms are finally moving: firms with 1 to 4 staff went from 17.3% to 23.0%. But firms with 250+ staff went from 30.3% to 41.0%, so the gap is growing. Europe shows the same pattern: the gap between large and small firms grew from 22 points in 2021 to 38 in 2025.

And use is shallow. 83% of small-business AI users write with it, 21% automate a process and 7% call it fully built in.

BCG's yearly studies, each worded a little differently, put the share of companies getting little or no value from AI at 74% in 2024, 60% in 2025 and about half in 2026. How many "get value" depends on how high you set the bar: 74% of leaders at large US firms report positive returns, 39% of McKinsey's respondents see any profit impact and 7% of KPMG's report established ROI.

Lab results shrink in real life. In studies of single tasks, AI raised output or cut time by 15% to 40%. Across all US work hours, it saved 2.2% in Q2 2026 (St. Louis Fed). Danish chatbot users say they save about 2.8% of their hours, and payroll records show no measurable change in pay or hours. Little of a working day is the kind of task AI is good at, and checking its work takes time.

People also feel faster than they are. In a METR trial with 16 experienced developers, they believed AI made them 20% faster. The clock said 19% slower.

Gartner expects $2.67 trillion of AI spending in 2026. More than half goes to chips and data centres. Stand-alone agents and assistants get $29.2 billion, about 1.1%, forecast to more than double to $65.5 billion in 2027. Agent features built into other software are counted elsewhere, so the real agent spend is larger. Treat big forecasts with care: Gartner raised its own 2026 number by 32% in a year, partly by changing what it counts.

The automation market is smaller and steadier. Gartner puts iPaaS at $5.9 billion in 2022, over $9 billion in 2024 and over $17 billion by 2028. Software robots (RPA) show how hype fades: growth fell from 63% in 2018 to 14.5% in 2024. Research firms disagree 8x on iPaaS in 2035, so we never print one of their numbers alone.

Agent revenue is real but small. Salesforce's Agentforce grew from $540 million to $1.2 billion a year in six months, about 2.2x, before Salesforce widened what it counts. Valuations move faster: n8n's roughly doubled, from $2.5 billion to $5.2 billion, in seven months.

METR measures how long a task an AI agent can finish half the time. It was 4 minutes for GPT-4 in March 2023 and about 17 hours for an early Claude Mythos Preview checkpoint in April 2026 (METR's range: 8 to 55 hours), doubling about every 129 days since 2023. These are software tasks. At 80% success, the bar a business can trust, the same model manages 3.1 hours.

One pass is not a system. On a 2024 retail test, the best model tested (gpt-4o) passed 61.2% of tasks once but fewer than 25% every time across eight tries. On long office workflows, the best agent finishes 20.6%. Claims shrink as the bar rises: 62% of big US firms are "building, deploying or developing" agents, about 1 in 4 of McKinsey's respondents is scaling one somewhere and no more than 1 in 10 scales agents in any one function.

The weak points are plumbing and cost. Big companies connect 27% of their apps, and 96% of IT leaders say agents depend on those connections. The average price businesses paid per million AI tokens fell 41% from its March 2026 peak, partly because firms moved to cheaper models. Yet Gartner forecasts the inference cost of each agentic workflow will rise more than fivefold through 2028, because agents use far more tokens per job.

Upwork's client spending grew 3.6x from 2016 to 2022, then went flat at about $4 billion a year. Its active clients fell 6% while spend per client edged up to a record $5,230. Fiverr shows the split clearly: buyers fell 36% from 2022 while spend per buyer rose 41%. Fiverr told investors that low-value work is shrinking faster than it can move upmarket.

AI work on Upwork keeps growing, but the boom has cooled: growth topped out around 50% in late 2025. Its largest AI category, integration and automation, slowed from over 90% to over 50%, while AI strategy and consulting grew over 50% in Q2 2026. Simple AI content work got 90% more contracts but paid 13% less per contract, while freelancers doing more complex work with AI earned 45% more year on year.

Our own one-day snapshot of 58 Upwork automation posts on 4 October 2026 found a $50 median fixed budget (16 of 28 fixed-price posts were under $100) and hourly ranges of $20 to $40. n8n posts appeared up to 8x as often as Zapier posts. It is a small sample, so read it as a snapshot. Official partner badges stay rare: about 500 each for Make and Zapier, and 48 n8n expert partners.

Forecasters have a weak record. Of 29 dated predictions we audited (our own selection, not a random sample), about 4 came true, 3 partly did, 14 missed and 8 were written so they can never be checked. Gartner's headline agent forecasts keep arriving sooner, though the definitions shift too: 33% of enterprise software apps with agentic AI by 2028 (said in 2024), then up to 40% of enterprise apps with task-specific agents by the end of 2026 (said in 2025).

So we give ranges, built by us on measured trends, with named forecasts inside them. The starting line: 23.8% of US firms used AI in the last two weeks, about 1 in 4 of McKinsey's mostly large respondents is scaling an agent somewhere and agents are reliable for about 3 hours of software work.