Electricity took thirty years.
This took three.
Every general-purpose technology before this one gave companies a decade to adapt. Here's what actually happened between 2022 and now — and, for each stage, what it made possible for a business the size of yours.
Adoption speed is the whole story.
Anthropic's economic research makes the comparison directly: electricity took over thirty years to reach farm households after cities were electrified. Generative AI went from a research demo to something roughly two in five American workers used at work inside about three years.
That compression is why the usual strategy — wait, watch what competitors do, adopt once it's proven — produces a different outcome this time. The window it assumes doesn't exist.
Anthropic Economic Index, Sept 2025
How long other general-purpose technologies took to spread.
Comparison as framed in Anthropic's Economic Index (September 2025), which notes AI's adoption speed is unprecedented relative to prior technologies. Bar lengths are proportional to elapsed years.
Four years, four different worlds.
A chat window changes the public conversation
ChatGPT's release turned a research capability into something anyone could try in a browser. For most businesses this registered as a curiosity: impressive, occasionally wrong, not obviously connected to how work got done.
Almost nothing yet — and that was the correct read at the time. The mistake wasn't ignoring it in 2022. It was continuing to apply the 2022 assessment in 2025.
Capable enough to help, not to be trusted alone
Models became substantially more capable and gained the ability to work with images, documents, and long context. Adoption ran ahead of governance: employees started using these tools whether or not their employer had a policy.
- Share of US employees using AI at work doubled from 20% to 40% between 2023 and 2025. Source: Anthropic Economic Index, September 2025
- Business AI adoption climbed from 78% in 2024 toward near-universal. Source: McKinsey State of AI, as reported in industry compilations
Individual productivity gains, unevenly distributed and mostly invisible to the P&L. Your people were already using it. The question was whether anyone was capturing the benefit deliberately, or whether it was evaporating into slightly easier days.
From answering questions to completing tasks
Models moved from conversation into workflows: writing and running code, calling tools, working through multi-step problems. Spending followed. This is the year the technology stopped being a chat product and started being infrastructure.
- Generative AI spending rose to roughly $37B in 2025, about 3.2× the prior year's $11.5B. Source: Menlo Ventures annual report
- 52% of executives said their organizations had deployed AI agents in production; 74% reported ROI within the first year. Source: Google Cloud ROI of AI Study, 3,466 senior leaders across 24 countries
- Enterprise users reported saving 40–60 minutes per day. Source: OpenAI, State of Enterprise AI
The first year that doing nothing became a competitive decision rather than a neutral one. Not because AI became mandatory, but because competitors who moved began compounding a capability advantage you couldn't buy back later at the same price.
Agents in production, and a widening gap between those who got value and those who didn't
Autonomous and semi-autonomous systems moved into real production use across support, research, analysis, and engineering. And the split we now consider the defining fact of this period became measurable: adoption is near universal, value capture is not.
- Around 8 in 10 organizations report AI agents have delivered measurable ROI — while independent research finds only about 6% capturing significant enterprise value. Sources: Anthropic 2026 State of AI Agents Report (500+ technical leaders); McKinsey State of AI
- Reported obstacles: integration (46%), data quality (42%), change management (39%). None of them are model capability. Source: Anthropic, 2026 State of AI Agents Report
- Over 40% of agentic AI projects are forecast to be cancelled by 2027, on unclear ROI and weak risk controls. Source: Gartner
- Meta became the first major technology company to formally tie employee performance reviews to AI usage in February 2026. Source: Bloomberg
The bottleneck has moved. It is no longer capability, cost, or access — those are solved and cheap. It is whether your data, processes, and ownership can support what the technology can already do. That's an organizational problem, and it's the one we work on.
Capability doesn't arrive. It accumulates.
The most useful finding we've seen isn't about how fast models improve. It's about how fast people improve at using them. Anthropic's Economic Index analyzed real usage and found that longer-tenure users had a meaningfully higher success rate — roughly four percentage points — and the effect held after controlling for the tasks attempted, the models chosen, language, and country.
Their reading is learning-by-doing. Experienced users also brought harder, more valuable work and collaborated with the tools rather than simply delegating to them.
Source: Anthropic Economic Index, March 2026. Success measured as the model's assessment of whether a conversation achieved its aim; effect persists under controls for task and request cluster, model, use case, and country.
The same curve, inside your business.
You can't influence how fast the technology improves. You can influence how fast your organization climbs its own learning curve — and that curve is the one that determines your outcome.
Start narrow
Google Cloud's read on top performers matches ours: the strongest results come from the narrowest first scope. One well-defined task, not a general assistant.
Measure the before
Establish the baseline first. Without it you can't prove the result, and without proof you can't fund the second thing.
Fix the plumbing
Integration and data quality lead every obstacle list published — including the ones published by the AI labs themselves. A week of assessment saves a quarter.
Build the habit
Invest in the capability, not just the technology. The compounding lives in your people's practice, not in your license agreement.
Read our full analysis of the research behind this timeline →
Where are you on this timeline?
The Modern Readiness Assessment places your organization against the failure points in the 2026 column. Thirty minutes, and you keep the scorecard either way.