In this article

• Does your AI setup get better the longer you use it, or just older
• What makes an AI foundation compound instead of decay
• How to build so capability stacks instead of resetting
• What a foundation that compounds looks like in practice
• Why a compounding foundation can look flat and still be getting stronger
• How to keep the foundation compounding instead of starting over
• Common questions

A compounding AI foundation gets stronger the longer it runs, because each cycle of work is captured and reused so capability stacks instead of starting over. Most software does the opposite: it gets older and harder to change as complexity piles up. The useful question is not how new your AI is. It is whether your setup is getting better or just getting older, because those are two different curves.

The hopeful curve has a name. Under Wright’s Law, technologies that get used a lot get cheaper and better at a constant rate as cumulative production doubles: solar fell roughly 20 percent for every doubling of installed capacity across more than four decades, a 99.6 percent decline, on a learning-by-doing cycle. That is an illustration of compounding through accumulated experience, not direct proof about AI, and the honest caveat rides with it: a curve that looks like learning is not by itself evidence that experience caused it. A foundation gets older on its own. It only gets better by design.

The difference between a foundation that compounds and one that decays is one condition: whether the work each cycle produces is captured, reused, and maintained so it feeds the next cycle, rather than piling up unused. Meet that condition and capability stacks. Miss it and you get the default, which is decay.

The default is expensive. On the Stripe developer survey, engineers spend about 42 percent of the work week on maintenance and bad code, and roughly a third of it on technical debt specifically, around 59,000 dollars per engineer a year. That is the cost of aging software that nobody paid down: a growing share of effort spent just to stand still. A foundation that compounds has to beat that default on purpose, because nothing beats it by accident.

A foundation gets older on its own. It only gets better by design.

Capability stacks when each unit of work is built to make the next one easier. In compound engineering, solved problems get codified into reusable tools, patterns, and documentation that future work retrieves automatically, so each unit of work makes the next one easier rather than harder. A fix does not just close one issue; it removes a category of future issues, and the setup gets easier to understand and extend over time.

That is the by-design condition stated as a build rule. The opposite is the codebase everyone has worked in, where every new feature is a negotiation with the ones before it, so the team ends up fighting the system instead of building on it. The same choice sits under an AI setup. If what the system learns this month is captured and reused next month, the foundation accumulates leverage. If it is discarded, you accumulate debt and call it progress.

In practice, a compounding foundation shows up as returns that climb the longer the setup runs and learns. On one vendor benchmark for AI customer service, returns average 41 percent in year one, climb to 87 percent in year two, and pass 124 percent by year three as the system learns from real interactions and teams optimize their knowledge bases. Treat that as directional, one function and one vendor, not a universal number. The shape is the point: the organizations that treat a deployment as continuous improvement, not a one-time install, see the strongest results.

The mechanism underneath is undramatic. Every call handled, every correction made, every play that worked feeds the next cycle, so the setup that ran for a year is not the same setup that shipped. A setup you can watch improve on a real number is a setup that got stronger rather than merely older.

The early number often hides the compounding. Drawing on a study of 540 companies across 16 industries, one analysis argues that hours saved stops telling the truth as AI matures, because productivity shows up first while growth, trust, and insight take longer to surface and compound in the background. The same source is honest about the limit: AI ROI measurement is still immature, and most companies never reach the compounding value at all, because the ownership and measurement to capture it are missing.

Maturity is where the gap shows. Segmenting adopters by sophistication, Deloitte finds 67 percent of the most mature adopters report large or very large ROI across 46 measures versus 61 percent of the least mature, and notes that even marginal differences can compound into measurable value over time. A foundation can look flat in the hours-saved column and still be getting stronger underneath, because the value it is building has not surfaced yet. This is the Resolution end of the FOMO to Fatigue to Resolution arc: past the fear of missing out and past the fatigue of tools that never compounded, the founder stops chasing the newest thing and starts asking what the setup keeps.

Here is where the fix becomes concrete, because compounding is an architecture choice before it is a result. The foundation compounds only when the durable part, the business the system has learned, is held separate from the swappable part, the models underneath. Works is built on that separation. The Areas, Categories, and Notebooks it accumulates as you run the work sit in a layer kept apart from the intelligence, so when the underlying models change, the business the system already learned stays learned. New models, new connectors, and new workflows land inside the setup you already have, and a six-month-old workspace runs on today’s version with no re-setup. The investment compounds instead of resetting. That is the No AI Tax difference, a foundation built for you rather than rebuilt by you every time the ground moves.

If you want to build on a foundation that compounds, sign up for early access. The specifics of how a setup absorbs each new model without a rebuild are their own subject, covered in how a setup absorbs new models without starting over. This idea sits inside a larger one, that AI can be an appreciating asset rather than a sunk cost, which runs through Compounding AI and its sibling on why your AI investment should hold its value, and it is what an AI Business OS is built to deliver.

You do not want the newest AI. You want the foundation that is worth more next year than it is today, because it kept what it learned.

Why does most software get harder with age while some AI setups get easier?

Most software gets harder because complexity accumulates. Under Lehman’s law of Increasing Complexity, each change raises internal disorder unless effort is spent counteracting it, so development slows and every change costs more, felt as everything taking longer than it used to. An AI setup gets easier only when it inverts that default on purpose, capturing and reusing what it produces so each cycle feeds the next.

Is compounding just a buzzword?

Often, yes, until one condition is met. A cost curve that looks like learning is not by itself proof that experience caused the improvement, because experience and time usually move together, and accumulation can cut the other way: a data moat tends to erode as the corpus grows, because each new record costs more to acquire and adds less value. Compounding is real, but only when context is actively captured, reused, and maintained.

What makes a foundation compound by design rather than by accident?

By design means every cycle’s output is kept and fed back in, not discarded. The test is concrete: does the work the system did last month make this month’s work easier, or does the setup start from a blank slate each time. If nothing carries forward, there is no base for anything to compound on, no matter how much activity runs on top of it.

1. Our World in Data, Learning curves and Wright’s Law (Max Roser), 2023
2. TechnicalDebtCost.com, How Technical Debt Destroys Developer Productivity, 2026 (primary Stripe Developer Coefficient, 2018)
3. Every, Compound Engineering (Kieran Klaassen), 2025-2026
4. Fin (Intercom), The ROI of AI Customer Service Agents, 2026 Benchmarks
5. Make, The AI maturity payoff (Daria Hvizdalova), 2026
6. Deloitte Insights, Three ways mature AI adopters can capture more digital value
7. Laws of Software Engineering, Lehman’s Laws of Software Evolution (Milan Milanovic), 2026 (primary Lehman and Belady, 1980, IEEE)
8. New Things Under the Sun, on Peter Thompson, JEP 2012 (Matt Clancy)
9. Andreessen Horowitz, The Empty Promise of Data Moats (Casado and Lauten), 2019