In September I rebuilt this website. It took just over a day. From my "ok go" to the new site live was about 24 and a half hours, and in that time I reviewed four design directions, two working builds, and a blind test where I picked between three AI writers' drafts without knowing which was which. Earlier in the same week, I had built the first issue of a video newsletter, This Week in Education, a channel built to run itself. First message to a full-length prototype video took under 14 hours. Approval of the finished issue took about 31 hours, and around 7 of those were my night's sleep.
Before AI, these jobs ran on a different clock. HubSpot's guide to website redesigns, a vendor's guide updated in 2024, says many redesigns take 3 to 4 months. A video agency's own guide, Yans Media, undated, says 6 to 10 weeks for a 60 to 90 second animated explainer, with a first storyboard arriving about two weeks in.
Where the time went
I ran Flip Education with one other person and a set of AI agents. I'd say the saving is mostly the handoffs. The more people involved, the more that added in terms of turnaround times. A brief goes out, a draft comes back a week later, comments go back, and each round costs days of waiting even when the actual work is short.
With a small team and agents, the waiting between my question and the answer more or less disappears.
In February the PR desk for Flip Education sent its first batch of pitch emails about two hours after I asked for the work to start, and the first reply from an editor came the next day. It was a no. Less than 48 hours in, I judged the round promising and widened the list from the biggest sites to middle-sized ones, because the rule above all others was whether an outlet's typical reader could use the product today. More on that desk in the Digital PR piece.
What a day buys
A first version in a day changes the month. If one try costs a month, you get one try, and it has to be right. If one try costs a day, you get many tries, and each one can be wrong.
The research points the same way. Koning, Hasan and Chatterji, in the journal Management Science in 2022, found that start-ups which adopted A/B testing, showing two versions to different users and keeping the better one, did 30% to 100% better after a year, measured in visits to their websites. They built more new products, and they gave up sooner on ideas that got bad signals. Camuffo and colleagues ran randomised trials, first with 116 Italian start-ups and then with 759, published in the Strategic Management Journal in 2024. Founders taught to test their ideas like scientists dropped bad ideas more often, and changed direction a few times rather than never or again and again.
That second part matters. The win was more tries, and also earlier, cleaner decisions.
The website showed me this in one afternoon. Four AI reviewers scored the four design directions, and all four ranked last the one I chose, a calm, senior page that recreates a well-made existing design with room for evidence. They read the brief's "jaw-dropping" as loud. I read it as calm and senior. The video showed it too. The first round built four finished videos, one per style, before anyone saw a frame. That cost about $210 of computing at list price and four hours, and it was the wrong method. I stopped it and changed to stills first, then a 15-second clip, then the full piece, with my pick between each step. I chose the style from the stills. Every one of those times sits next to a decision of mine. Speed alone didn't do the work.
The limit moves
Noy and Zhang, in the journal Science in 2023, ran an experiment with 453 professionals and found AI cut the time for writing tasks by 40% and raised quality by 18%. I believe it, for writing. But the counterpoint deserves equal space. METR, a research group that measures AI, found in early 2025 that experienced software developers working on their own large, mature code took 19% longer with AI tools, while believing they were faster. The gain depends on the task. New work, first drafts, first versions of things, that's where I see the hours come back. Polishing something old and complex, probably less so.
So here's where I land. When building costs weeks, building is the limit, and deciding feels cheap because there's little to decide on. When building costs a day, deciding becomes the limit. Judgement is still the work.
The judgement includes the boring kind. In September an agent ran a wrong command that closed every open work session and browser tab on my computer. I asked only for everything to be restored, and never asked for a fix. A check that refuses that kind of command was in place under 13 minutes after the mistake, because a standing rule of mine is that every failure is a gap in the setup, and the agents help close it. That rule is a decision, made once, and it pays every time something breaks. More in My personal agent setup.
If your team is deciding how your company should work, I wouldn't start with the tools. I'd count your handoffs, and ask what a month looks like when each idea gets three tries instead of one.