The Twilight of the Chatbots, Seen From Inside the Fortune 100

Ethan Mollick's latest piece describes a real shift. The chatbot era, where a person prompts a model and reads the output, is giving way to agents: systems that run for hours, use tools, correct their own mistakes, and report back when the job is finished. His evidence comes from inside OpenAI itself. A joint OpenAI-academic study found that legal, HR, and other non-technical functions have adopted agents at nearly the same rate as engineering. Working alone for fourteen hours, Opus 4.7 built a software package that would have taken a human engineering team two to seventeen weeks, for $251 in tokens. A quarter of OpenAI's staff run four or more agents at once, every week. His conclusion: the valuable skill now is defining the job and checking the output, not prompting well.

I sit in a different room than Mollick's data comes from. He measured a company built by the people who build the models, staffed by people paid to master them. I spend my time with innovation leaders at companies where the AI team is three people and a shared Slack channel, running programs where the chatbot pilot from eighteen months ago still hasn't cleared legal review. Mollick names this limitation himself: his findings aren't independently replicated, and he doesn't know if they transfer outside a lab built for this exact purpose. That caveat is the whole story for most organizations I work with.

Where the real gap sits

The technology described in Mollick's piece already works. The gap I see in the organizations I advise sits in adoption, and it shows up in three places every time.

Leaders confuse agent autonomy with an on-off switch. A pilot either stays locked in a sandbox for a year, reviewed to death, or gets handed the keys to a production system after one good demo. The companies making real progress treat autonomy as a dial. Narrow scope first, high-stakes review at every step, wider latitude only after the agent earns a track record, not after an executive gets excited in a meeting.

Reskilling programs still teach prompt writing, a skill with a shrinking shelf life. Mollick is right that the job now is defining the task and checking the result. Most corporate training I review hasn't caught up to that. People are learning how to phrase a request to a model instead of learning how to scope work an agent can own end to end and how to audit what comes back. Skills expire faster than the org chart admits, and this is exactly the relearning I wrote a book about.

Agent adoption gets treated as an IT rollout instead of a change to how the business runs. The detail in Mollick's piece that stood out to me most was the pattern across departments: legal and HR picked up agents at nearly the same pace as engineering. That only happens when a technology isn't gated behind one specialist team. Most enterprises route every AI initiative through a center of excellence, and the center becomes the bottleneck the technology was supposed to remove.

The checkpoint the agent era removes

Here is where I would extend Mollick's argument. The chatbot era had a safety mechanism built in by accident: a person read every output before it went anywhere. Agents remove that checkpoint by design, and the failure mode that follows is quiet. Scope creeps a little at a time. A small error compounds three steps downstream before anyone notices.

The organizations that benefit from this shift are building the output-checking discipline before they expand agent autonomy, not after the first costly mistake forces the conversation. That discipline is a culture investment, and culture investments do not show up in a benchmark. They show up eighteen months later, in whether the organization trusted the technology enough to use it and built the judgment to catch what it misses.

Mollick is describing where the technology is going. The harder and more urgent question, for every organization that isn't paid to build these models, is whether the people and the culture around the technology are ready to go there with it.

Alex Goryachev is an AI keynote speaker and former Fortune 100 innovation executive, author of "The Great Relearning," and has delivered 310+ keynotes on agentic AI and organizational transformation. Explore his AI innovation keynotes or start a conversation about your organization's AI readiness.