In 2015, I spent the better part of three years reading ten years of writing about how social enterprises grow their impact. It was my doctoral research, a longitudinal content analysis of the Stanford Social Innovation Review, and the question underneath all of it was deceptively simple: when a mission-driven organization has something that works, what makes it spread, and what keeps it from spreading?

That research was not about artificial intelligence. Generative AI as we know it now did not exist when I defended my dissertation. So I want to be clear at the outset that I am not claiming I saw any of this coming. I am claiming something I find more useful, and a little uncomfortable: more than ten years later, watching associations try to scale AI, I keep recognizing the same patterns I documented back then, wearing different clothes. The tool is new. The problem of scaling mission-driven work is not.

One of the first things my research surfaced was that the language itself was slippery. "Social enterprise," "social entrepreneur," and "social entrepreneurship" meant noticeably different things to different writers, and that vagueness was not harmless. It let people sound aligned when they were not, and it quietly slowed the work.

I see the same fog in the AI conversation now. "AI strategy," "agent," "AI-native," "scale" — all of them get used loosely, and the looseness costs us. People walk into a conversation saying "we need an AI strategy" the way they would order a logo, as if it were a deliverable, when what they actually need is clarity about what they are trying to accomplish.

The lesson then is the lesson now: get specific before you try to get sophisticated. Name the outcome you want, in plain language, before you reach for the vocabulary. The vagueness feels like alignment. It usually isn't.

The most persistent theme across all those years of writing was measurement. Nearly everyone agreed that social impact had to be measured. Almost no one found it easy to do in practice. So organizations did the natural thing: they counted what was countable, reported that, and let the harder question drift.

This is the pattern I find most alive in association AI work. We count the things that are easy to count: staff trained, tools deployed, hours saved on a draft. Those are real, and worth tracking, and they tell you the program is running. What they do not tell you is whether any of it moved a member outcome or advanced a goal your strategic plan would recognize.

The answer is to track outcomes anyway, honestly and imperfectly. Capture a baseline before you start. Watch the member-facing numbers that matter, name AI's plausible contribution to them without overclaiming it, and let the picture build over a year or two rather than a quarter.

The third pattern is the one that surprised me most in 2015. The barriers to scaling were rarely the innovation itself. A genuinely good solution would stall, and when I looked closely the reason was almost never that the idea was weak. It was that the context was wrong, or that the people and partners whose buy-in the work depended on were never brought along.

I learned this one a second time, the hard way, in my own work. Last year I led the build of a member microlearning system, using AI to turn webinars into short, usable learning. The build went well. The thing we made was good. And it quietly underperformed, because we never established the operating rhythm that gets a good thing used: a named owner, a promotion cadence that outlived the launch, tracking from the first week.

That is the gap I watch associations fall into with AI now. A pilot works, a few staff get faster, and then it stalls — not because the tool failed but because nobody built the rhythm to carry it.

If the scaling problem is old, that is actually reassuring, because it means we are not starting from nothing. We have a fair amount of hard-won knowledge about what closes that gap, and most of it has very little to do with the tool. It has to do with being specific about what you want, honest about what you are measuring, and deliberate about the coordination that turns a working pilot into a lasting capability.

So the question I would leave you with is the one I keep asking myself: now that you can see the old pattern coming, which one are you about to repeat with AI, and what would it take to break it this time?