Redefining Agile In The Age Of AI

Tim Lee is the Founder and CEO of Bookipi, an international SaaS company trusted by millions of small businesses worldwide.

gettyFor as long as I’ve built software, the rule has been the same: Deliver fast, fail fast. Get a minimum viable product (MVP) out the door, then gather user feedback to improve the next iteration. Agile took that instinct and turned it into a discipline, with shorter cycles, tighter feedback loops and quicker releases.

AI has changed that strategy, and not in the way most people expected. Product managers, designers and engineers can now spin up a working prototype in record time using code generators. When everyone can shorten development time with AI, being fast stops setting you apart. Research shows that AI can reduce the time developers spend on some development tasks—so those hours aren’t where the contest is won. Quality and user experience are.

That doesn’t make agile obsolete. What it means is that agile is shifting its center of gravity away from sprints, trouble tickets and ever-shorter release cycles and toward consistency of what ships.

Agile was built for a different era—one when engineering was the scarcest resource in the building. Breaking work into small pieces and running it through scrums and sprints was how teams offset the sheer cost of human coding hours.

AI has knocked most of that cost out. Generating the first version of an application is easy now, and a machine can handle the working prototype, the tests and the documentation. The bottleneck now sits in human judgment, verification and integration. A 2025 randomized controlled trial by METR, a nonprofit research group, captured this sharply: Experienced developers using AI tools took 19% longer to finish their tasks, even though they believed the tools had made them faster. The speed was only real in feel, not in fact, because the effort simply relocated from writing code to reviewing, reconciling and validating what the machine produced.

Agile also rested on the idea that mistakes were expensive, so the cheapest way to test a product was to ship the MVP and find out. That logic weakens when AI lets you generate several versions, compare them and validate before launch rather than after.

The way we measure productivity has to move with it. Lines of code, tickets closed and release frequency—these count motion, not progress. A team can look busier than ever and produce nothing more coherent or more valuable to a customer. Expand your capacity to build without sharpening your sense of what’s worth building, and all you’ve really done is learn to create waste faster.

The original point of rapid releases was learning. You shipped in order to test assumptions against real customers. AI changes the economics of that too because now you can test more assumptions, simulate more edge cases and generate more scenarios before anything goes live. Handled well, that should mean a more mature, higher-quality product reaches the customer in the first place.

I’ve started calling this approach “move fast, launch slow.” Iterate rapidly in private, as fast as the tools allow, but release less often and only through well-defined quality gates. The aim isn’t to ship the first version that works. It’s to ship the first version that delivers real value and earns a customer’s trust.

It borrows from both traditions. It keeps the speed and iteration of agile but brings back some of the discipline of waterfall—more thinking up front, more attention to the whole workflow and stronger verification before release. You could call it an AI-native waterfall, although the label matters less than the operating principle behind it: Let AI compress the iteration, but keep a human accountable the whole way through.

Code generators haven’t made agile obsolete either. They’ve changed what’s worth optimizing for.

Velocity metrics are giving way to quality and outcome metrics. It isn’t enough to ask whether a release shipped on schedule. It’s whether it improved task completion and reliability, and whether it moved customer satisfaction and retention.

Verification matters more than it ever has. AI-generated code needs clear standards to check performance, security, usability and consistency against, which means named quality gates and documented criteria rather than a general sense that things look fine.

There should also be fewer handoffs, but not at the expense of expert review. Because product managers and designers can now build working prototypes and engineers can weigh in on product decisions earlier, the work doesn’t have to be passed from hand to hand. The goal is a shared product the whole team can test and improve together.

Above all, the end-to-end experience is what counts. Customers don’t care about sprint increments or how quickly the latest release landed. They want a product that works and that they’re glad to use. AI makes software easier to build, but it doesn’t make it distinctive or trustworthy on its own.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?