Amazon, eBay, and Google Were Built on Dial-Up. None of Them Waited for Broadband.

If you’re over thirty-five, you’ll remember this: pages that took thirty to forty-five seconds to load, downloads that cut out halfway through and had to be restarted, and buying something online meant filling out an endless form with your fingers crossed that the transaction wouldn’t crash before you finished.

And yet, in those same years, with that same terrible connection: Amazon was already selling books in 1995. eBay was already connecting buyers and sellers in 1995. Google was born in 1998. PayPal was founded that same year and launched its first product in 1999.

None of them waited for broadband to exist. They built with the technology available at the time, learned faster than the rest of the market during that uncomfortable stage, and that learning became an advantage that those who came later — even with better infrastructure — could never close.

AI in 2026 is in its own version of that dial-up stage. Hallucinations, inconsistent outputs, governance gaps, a regulatory cloud that still hasn’t fully settled. Some recent data confirms this with more precision than anyone would like. Only 29% of developers trust what an AI tool generates, compared to more than 70% just three years ago. A controlled study by research organization METR found that, working on code they already knew well, experienced programmers ended up being 19% slower using AI — not faster. And code generated by AI contains, on average, 2.7 times more vulnerabilities than code written by a person.

This is not an argument for waiting. It’s exactly the opposite argument.

Companies waiting for AI to be “production-ready” before committing are making the same bet that the late-mover retailers lost when they waited for broadband before getting into the internet. By the time the connection improved, Amazon already had years of purchase behavior data, eBay already had a trust network between buyers and sellers that no newcomer could replicate overnight, and Google had already learned more about how people search for information than any competitor arriving later with better technology.

This doesn’t mean moving forward without care. It means something more specific: building with what exists today, in real processes, even if it’s still imperfect — instead of waiting for a finished version that probably won’t arrive exactly when you’re expecting it. Today fewer than one in four teams already deploying AI agents has a mature governance model to manage them. That gap won’t close by waiting. It’ll close by learning, with real data, while the technology keeps improving under everyone’s feet at the same time.

This column is, essentially, permission. Not permission to be careless with AI — but to stop treating it as something that needs to be finished before you can use it. Imperfect AI, integrated today into a real operation, will almost always beat perfect AI that someone is still planning for next year.

What’s the “dial-up problem” in your company right now? Are you using it to learn faster than everyone else, or as an excuse to keep waiting?