# Patience Is a Virtue: The Case for Decreasing Velocity When You Build with GenAI [Frank Anderson](/team/fanderson) Software Engineer Date July 1, 2026 Categories [Artificial intelligence](/blog/category/artificial-intelligence) [Digital learning development](/blog/category/digital-learning-development) [Product strategy](/blog/category/product-strategy) > The companies that pushed genAI hardest last year are quietly slowing down, and landing on the same lesson: genAI speeds up the work, not the judgment needed to check it. A small team spends a weekend building something that, by Monday, looks finished. A genAI tool wrote most of it. The demo lands, leadership is thrilled, and someone asks why the rest of the product roadmap can't move this fast. Then a few weeks pass, real users show up, and the product starts generating errors in ways nobody predicted. Worse, nobody on the team can say why, because nobody can fully explain how it works. Let's get one thing out of the way: genAI is powerful, and it is worth building software applications with it. The gains are real, the tools keep getting better, and companies that sit it out will fall behind. What I see most folks miss when they jump headlong into genAI software development is that it still requires following a disciplined process to achieve good software: the incremental iteration, the reading and reviewing, the testing as you go. It’s that discipline of proven processes that makes the genAI tools pay off. Slowing down to follow a consistent process lets you speed up in the long-term. Take smaller, more measured steps. Evaluate and adjust. Then iterate again and document as you go. Sacrificing this for reaching a goal faster might feel like progress, but the time saved up front will be lost when you have to stop, untangle the code, and start understanding how to fix it months later. ## Why Sprinting Feels Reasonable Image The pressure to move fast is easy to understand. An experienced software engineer or online instructional designer familiar with the regulatory requirements and market expectations of the U.S. K-12 education market is expensive and hard to find. The cost of a genAI tool seems like a rounding error compared to that. However, the math isn’t that simple. It neglects the true value of what you were paying for: human experience, judgement, wisdom, and tacit knowledge. The output is the easy part; genAI tools can produce that in minutes now. Instead, it’s the process the human brings to the table: reviewing, understanding, testing, catching what's missing, and anticipating what’s needed. That comes from years of hands-on practice, not from a subscription. GenAI tools produce work that looks like expertise: clean code, fluent copy, a polished prototype. But if what it’s generated isn’t in your domain of expertise, you can’t tell whether it’s accurate or not. If you can't evaluate it closely, looking right and being right are the same thing. That's the trap. Try this: Ask your favorite genAI tool about something you have deep knowledge and experience with. You’ll quickly see that the genAI is 80-90% accurate, with a few things you’d likely say are missing, or should have been phrased differently, etc. Now ask it about something you know absolutely nothing about. Can you see the gaps? How do you know it’s accurate? Many across industries are starting to realize the problem this poses. Sprinting to market with a hot prototype built with genAI gets all the attention. But the hard work and discipline needed to make it into a great product, takes time, wisdom, and experience. ## Even the Companies That Sprinted Are Slowing Down If you feel the urge to sprint because everyone else is, notice that the companies furthest out front are quietly tapping the brakes. Klarna [replaced roughly 700 customer service roles with an assistant built with OpenAI](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396), only to realize it didn’t save them anything and are now [rehiring people](https://www.digitalapplied.com/blog/klarna-reverses-ai-layoffs-replacing-700-workers-backfired) and using human experience to augment genAI. Within a year of [Duolingo announcing an "AI-first" strategy](https://fortune.com/2025/05/24/duolingo-ai-first-employees-ceo-luis-von-ahn) it pivoted to [recast the use of genAI in their workplace](https://fortune.com/2026/04/13/duolingo-ceo-luis-von-ahn-ai-usage-requirement-employee-performance-evaluations/). [Ford just did the same thing](https://techcrunch.com/2026/06/28/ford-rehires-gray-beard-engineers-after-ai-falls-short/) for many of its quality engineers. The risk for our industry is blunt: pull the human experience and judgment out of educational content, and you tend to get a result built without the educator (or learner) in mind. No one is quitting on the technology. Everyone is just realizing that human judgment has to stay in the loop. ## Experience Is What Makes GenAI Work Better When I wrote [What Your AI-Generated Code Isn't Telling You](https://www.clarity-innovations.com/blog/what-your-ai-generated-code-isnt-telling-you), I talked about comprehension debt: a quiet pileup of missing security, privacy, and accessibility protections that no one on the team can explain, maintain, or fix. It compounds in silence until something breaks in public. This problem can be mitigated by an experienced engineer working through a thorough process of reading and reviewing the code *before* it is considered finished. For K-12 education, that means working with [software engineers that understand](https://clarity-innovations.com) the difference between FERPA and COPPA, WCAG 2.2 and UDL. Someone that knows which data schema and models work best for protecting student PII, and is familiar with the requirements of NIST or SOC2. None of this shows up in a genAI prototype. Or the genAI tweaked demo by someone on the marketing team. And unless there’s a human in the loop that has the experience with these mandatory regulatory requirements, they might not be addressed in live production code. In K-12 education, the stakes get concrete fast. If nobody has read the code then nobody can explain what the code does. Slow down, get help from experienced product developers familiar with the education market, and avoid the costly mistakes that sprinting to the finish with genAI can cause. If you’re unsure whether your genAI developed product is suffering from this, [please reach out](https://clarity-innovations.com/contact) and take advantage of our free one-hour consultation offer.