Has AI ever become capable of something much sooner than you expected? If so, you are in good company. Most leaders who build an AI strategy plan around a pace of change that no longer matches reality.
We expect progress to arrive at a steady pace. A project that took a year last time should take about a year this time. That instinct works well for most of business life.
AI breaks it. Our research suggests AI changes 5 to 10 times faster than mature industries, and up to 20 times faster on some technical measures. That gap is large enough to make a plan built on steady progress out of date within months.
Linear versus exponential: a 30-step example
A simple example shows the difference.
- Linear growth: take 30 steps forward and you have moved 30 steps.
- Exponential growth: start with one and double it 30 times, and you pass one billion.
Each gain builds on the one before it. It compounds. Researchers call our habit of underestimating this kind of growth exponential growth bias. It helps explain why something that looks years away can suddenly become possible much sooner.
The Half-Life of an AI Strategy, our research paper, puts numbers on this. In two years, the longest task AI finishes reliably went from 4 minutes to 12 hours. The paper also cites research showing that the length of real software tasks AI can complete is doubling about every 4 to 7 months. Work considered beyond AI's reach today could become routine within a year.
Cost is moving in the same direction. The paper reports that the cost of reaching a fixed level of AI performance is falling about 13 times a year. When capability rises while cost falls, a business case written twelve months ago can expire before anyone revisits it.
AI memory is growing too. The paper describes working memory of 10 million tokens, which is about 7.5 million words, or roughly 94 novels. Earlier in the field's history, image-model training time dropped from 10 days in 2017 to 2 minutes 43 seconds in 2019. Each step looked unlikely until it happened.
Look at how quickly one AI company has gone from founding to the public markets. Anthropic was founded in 2021 and, according to Axios (https://www.axios.com/2026/10/26/anthropic-ipo), filed for its initial public offering (IPO) on October 26, 2026, with shares possibly trading before Thanksgiving.
Valuation expectations are still moving. Reuters reported (https://www.reuters.com/legal/transactional/anthropic-opens-books-ais-explosive-growth-staggering-costs-2026-10-26) that some investors now value the company at about $1.5 trillion, down from roughly $2 trillion a week earlier. Some analysts are skeptical of the higher figures, as Yahoo Finance noted (https://finance.yahoo.com/technology/ai/articles/anthropic-ipo-could-2026-most-193014357.html).
You do not have to agree on the right price to see the point. A company barely five years old is being sized up alongside some of the largest companies in the world. A few years ago that outcome would have sounded unlikely.
There is another reason to expect surprises. AI is now helping build better AI, a process called recursive self-improvement. When the tools that create progress improve themselves, the rate of progress can climb even faster.
"Don't build your AI plans around the idea that tomorrow will look much like today." (Jon Nordmark, Co-founder and CEO, Iterate.ai)
You cannot out-predict exponential change, but you can build habits that keep you ready for it.
1. Keep testing what is possible. A task that failed six months ago may work now. Re-run your old pilots against current tools and see what has changed.
2. Review your assumptions often. Put dates on the assumptions behind your AI plans, such as cost, accuracy and build-versus-buy, and revisit them on a schedule.
3. Be ready to change direction. Favor flexible architectures and short commitments over rigid multi-year plans tied to a single model or vendor.
Reviews only work if they are routine. Once a quarter, gather the people who own your AI plans and work through four questions:
- Which tasks did we rule out last quarter that we should test again?
- Which costs or performance numbers in our plan have changed?
- Which commitments lock us in for longer than the technology has stayed stable?
- What would we do differently if AI could handle tasks twice as long as it can today?
Write the answers down and date them. Over a year, that record shows how quickly your own assumptions aged, which is a more persuasive argument than any forecast.
AI keeps surprising us because we measure it with a linear ruler. Once you accept that progress compounds, planning gets easier. You stop asking what is possible today and start asking what you will do when the next capability arrives sooner than expected.
This idea comes from The Half-Life of an AI Strategy, a research paper from Iterate.ai. Read the full white paper here: https://iterate.ai/resources/white-papers/the-half-life-of-an-ai-strategy
If you want to talk through what faster AI change means for your roadmap, contact our team or request a demo. You can also follow Jon Nordmark on LinkedIn for the rest of the 30-part series.