After the automation and agentic AI what should be next?
I've been thinking about what the next stage of working with AI should look like.
Most of what I see focuses on making AI do more work. It writes code, generates designs, analyzes data, and automates workflows. Humans provide instructions, review the output, and decide what happens next.
But I think there's a more interesting problem to solve.
How can I combine human and AI intelligence to achieve better outcomes than either could achieve alone?
Take growth as an example. AI analyzes conversion data and finds a pattern. A growth practitioner recognizes a possible customer motivation behind it. That insight redirects the AI's analysis, which uncovers something that changes the original hypothesis.
What interests me is not just dividing tasks between humans and AI. It's how each can change what the other thinks, explores, and does next.
I think the same question applies to product design, software development, research, and strategy. Of course, more collaboration doesn't automatically mean better results. Sometimes AI should work independently; sometimes human judgment should lead. I'm interested in figuring out how to get the best out of both, rather than defaulting to either automation or human review.
I'm exploring this idea and trying to understand whether human–AI interaction can be designed more deliberately, instead of simply adding AI to existing workflows.
I'd love to hear from people building with AI:
Are you mainly using AI to execute tasks faster, or have you found ways to make the interaction between human and AI fundamentally improve how you think and work?
Replies
An agent fixed a failing test by weakening the assertion to accept the broken output instead of fixing the bug. The suite stayed green for days. Verification layers have to be tamper-proof and completely out of the agent's reach, because agents optimize for the pass, not the correctness.
@konstantin_tikhaev hey sorry but I don't think here my point is agentic failure discussion
@ajay_yadav13 The focus shifts from getting agents to execute tasks to independent verification. That means building isolated environments that prove the work is actually correct.
I see this loop in support more than in growth. One support thread from a single user showed what my funnel events could not: she played the song preview twice, never noticed the store badges below it, and wrote that she never got her free song. The welcome email said her credits were on the balance, and she read credits as money and asked us to take her card off the site. No dashboard would have told me that the word "credits" was the problem. The useful step comes after: with that one story, the AI can go back through the events and count how many other people stopped at the same place. Human for the why, machine for how many.