The Shift Nobody Has Operationalized Yet
Ethan Mollick's latest piece lands with real weight because it's backed by an OpenAI-commissioned study of how work is actually changing inside OpenAI itself — not a think-piece, an empirical snapshot. The conclusion: the chatbot model of AI work is already obsolete. The new model is task assignment to long-running, self-correcting agents. If your team is still organized around who writes the best prompts, you're optimizing for a workflow that's being deprecated.
This matters most for non-tech operators, because enterprise software vendors and tech-forward teams are already rebuilding around agents. The gap is widening every quarter you don't.
What 'Harness' Actually Means for Operators
Mollick introduces the concept of a 'harness' — the combination of tools, permissions, environments, and guardrails that an agent operates inside. The harness is now the differentiator, not the raw model. Two companies using the same underlying model can produce wildly different outcomes depending on how well the harness is designed.
For a practical translation: your CRM access rules, your data hygiene, your approval workflows, your escalation logic — these are your harness. Agents built on a clean, well-structured operational environment outperform agents built on chaos, regardless of which model powers them. This reframes AI investment from 'which tool do we buy' to 'how do we structure work so an agent can execute it without a human in the loop at every step.'
That's a product and operations problem, not a technology problem.
The Actionable Reframe for H2 Planning
If you're in budget or headcount planning right now, the Mollick framing gives you a concrete lens:
- Stop asking: Who on the team is good at prompting? What AI tool should we add?
- Start asking: Which discrete workflows can be handed to an agent end-to-end? What does the harness need to look like — data access, exception handling, output review — before we can trust it?
- Audit your current AI spend for chatbot-pattern usage: one-off queries, manual copy-paste into interfaces, ad-hoc generation. That's low-leverage. The high-leverage version is defining a repeatable task, building the environment for an agent to run it, and monitoring outcomes rather than managing each instance.
Companies still structured around chatbot workflows aren't just leaving efficiency on the table — they're building organizational muscle memory around a model that's already being replaced. The right move in H2 is to pick two or three high-volume, well-defined workflows and redesign them for agent execution. Start with the harness design, not the model selection.