10 AI Lessons Most Companies Learn Too Late

This past week I was in a lot of rooms, on a lot of calls, and deep in a lot of conversations about AI. Some were workshops, some were strategy sessions, and a few were just catch-ups with people building interesting things. The same themes came up across all of them, and I found myself sharing the same hard-won lessons in every room.

So here they are. Ten things I actually said to real people this week. No theory, no hype, just what's working and what I'm watching out for.

• A prototype is worth a million meetings: Business users can now describe what they want in Claude, get a clickable prototype, and hand that directly to engineers. Someone I caught up with this week redesigned his entire website in 90 minutes: screenshot, screenshot, screenshot, description, goals, framing, boom. That entire translation stack, from requirements to design to dev handoff, can be skipped. The prototype is the brief.

• Ask AI how to ask AI: When you're stuck, ask Claude what you should be asking it to accomplish your goal. It reframes the problem, surfaces approaches you hadn't considered, and gets you unstuck in seconds. I do this daily, and it's the highest-leverage habit I've found. The meta-move of asking the AI to help formulate the right question beats any prompt library I've seen.

• Sidecar the skeptics: When IT or security blocks AI on company data, don't fight it head-on. Build a high-visibility win on non-sensitive data first. Pick a highly acute pain and a very visible problem, build something independent of your data, and get the organization moving in the right direction. Earn trust, then expand. Fighting for access on day one is a losing game.

• Prompt, context, harness: Here's a maturity framework I share in almost every workshop. Most people are stuck at prompt engineering: crafting the right question. The next level is context engineering: giving the model the right background so it can do its job. The level above that is harness engineering: baking in guardrails so the whole thing works consistently without you babysitting it. Most teams fail because they never get to the harness. They're endlessly optimizing prompts when the real leverage is upstream.

• Three harness lessons from two years of AI-native engineering: Two years of building AI-native workflows have cost me some lessons I'd rather you not repeat. First, don't kill all process. I went too far anti-Agile early on and created chaos because people were stepping on each other with no coordination. A lightweight work queue is enough to prevent that. Second, build a shared harness so every engineer on your team is operating with the same guardrails, not their own improvised version. Third, optimize token spend before it eats your margin, because it will.

• Token burn is the next budget line you're not watching: The major AI providers are heading toward IPO, and subsidized rates won't last forever. Right now, everybody's getting hooked on cheap compute, and when pricing normalizes, a lot of companies are going to feel it. You can sometimes pay a junior developer less than what you're paying an LLM to do basic tasks. Token optimization strategies matter now, before rates go up and the math changes on you.

• Synthetic customers replace A/B test spend: Train agents on your customer data and test ad messaging, creative, and positioning against synthetic personas before spending real media dollars. My background includes running A/B testing at scale, and the amount of money we spent on paid acquisition just to learn basic things about our audience was staggering. Synthetic customers aren't perfect, but they're a fraction of the cost and they can run 24/7. It's a fundamentally different way to think about research spend.

• Push through the trough of disillusionment: People revert to old habits when new AI workflows feel slower at first. I hear it constantly: "I could do it faster if I just sent this email myself." But that's the moment you have to push through, not retreat. One team I work with took about 30 days of discomfort before it clicked, and then they went from operating like 10 people to operating like 50 or 60. Seth Godin calls this The Dip. AI adoption has one too, and most people quit right before the other side.

• The altitude change: AI moves your role from executing to orchestrating. You're going from doing the thing to designing the system that does the thing. The people who say "me crafting every single pixel is my value" are fighting the wrong battle. That's not what you need to bring anymore. You have to go upstream: designing the systems, defining the outcomes, and trusting the machines to handle the execution. The value shifts to the architect, not the laborer.

• Your email agent learned your voice in an hour: This one surprises people. There's tooling now that can ingest your historical emails and start auto-drafting responses in your voice and tone, including attaching the right documents without being told. Within about an hour of ingestion, it's writing emails that are close to 99% accurate. For someone like me who lives in email, this isn't a convenience feature. It's a force multiplier.

All of these tips point to the same thing: the competitive gap between companies that are actually implementing AI and companies that are still debating it is widening fast. The teams we work with who've pushed through the discomfort and built real harnesses aren't just more efficient. They're operating at a fundamentally different speed.

Building great experiences in the age of AI means designing the systems that make those experiences possible. That's what we're here to help with.

If you enjoyed this post, check out my newsletter. Each week I share thoughts and ideas on building great experiences in the age of AI: subscribe here.