AI Made the Fun Part Fast. Now Comes the Boring Part.

AI made shipping features fast and cheap. The cool part got easier. What got harder is the stuff under it: how you experiment, how you move your data, and how you pay the legal bill nobody put on the roadmap. Three pieces this week all point at the same place. The work that decides whether your AI features survive is the unglamorous work. Let me catch you up.

Stop spending engineers on pixels

The old playbook was tweak the funnel, measure for two weeks, move on. Elena Verna says that game is mostly over. AI products are conversational now. There are fewer screens to optimize. She uses Granola every working hour and isn't sure she ever opens the UI. So what exactly would you A/B test there?

Her rule is blunt: using growth engineers on interface tweaks is a crime. Hand the small stuff to AI, which gets pricing-page best practices right about 80% of the time. She even skipped a test on showing paid features in the free plan, because thousands of companies already proved it works. Save the real testing for the big bets.

The lever everyone is scared to pull

Here is the shift that matters for your roadmap. AI features cost real money to run. Every prompt burns tokens. So monetization hits earlier in the funnel and costs more per action. Verna calls skipping monetization experiments the single biggest mistake you can make right now.

At Lovable, her monetization bets are not aimed at revenue. They aim at engagement. Revenue-neutral changes are her biggest winners, because engagement turns into revenue a month or two later. She's testing 5 credits versus 10 on day one of the free plan. It looked terrible for 30 days, then the retention caught up. The catch: you have to run these for one to two weeks, then watch the cohorts for one to two months. Those two-week wins your execs love are now the worst thing you can do.

Your data has to be clean before any of this works

None of those long-horizon experiments mean anything if you can't trust your numbers. That's why the migration playbook from the team at Human37 matters more than it sounds. They've run hundreds of migrations. The teams that struggle aren't the ones with the most data. They're the ones who skip planning.

Two moves stand out. First, lock in the why before you move a single event, and write it down so you can measure success later. Second, run both platforms in parallel for 30, 60, or 90 days. When your old "5" shows up as "5.1" in the new tool, that's not a bug. It's your new baseline. Treat the migration as spring cleaning too: inventory every event, settle what "conversion rate" actually means, and bring only the history your business needs. Old data is technical debt from day one.

The bill you forgot to budget for

Then there's the cost that never makes the roadmap. Jay Stansell at Product Coalition calls it the hidden tax on every product decision. The EU AI Act can fine you up to 7% of worldwide turnover or $43 million, whichever is higher. Plenty of big model providers already failed to comply. That's not a checkbox at the end of the race anymore. It belongs at the starting line.

The expensive part isn't even the fine. It's the audit trail. As one guest on his podcast put it, in regulated work you need to go back to the moment you made a decision and say why. That infrastructure costs real money. Webflow's answer is worth copying: legal and privacy teams review products at the spec stage, and PMs run features through custom GPTs to catch flags early, before anyone commits resources.

The deep cut

Here's the thread tying these three together. AI shifted your real costs from build time to everything around the build. The feature is cheap now. The monetization model, the clean data to measure it, and the compliance trail to defend it are where the money and risk live.

So change one habit at your next review. When someone pitches an AI feature, don't ask how long it takes to ship. Ask three things: what's the per-use cost once this scales, can we actually measure it on a 60-day horizon, and what's the compliance line item. Borrow Ryan Singer's framing from the Product Coalition piece: appetites start with a number and end with a design. Set your appetite to include the boring costs, or you'll be firefighting them later.

Three questions for your team

1. What are our growth engineers building this quarter, and how much of it is UI tweaks that AI could handle so they can take on the harder bets?
2. Can we trust our analytics enough to run a monetization experiment for 60 days, or do we have a data migration or definition problem to fix first?
3. For our next AI feature, what's the compliance and audit cost, and who reviews it at the spec stage instead of after we've already built it?