In 2024 the talk about AI in the company was simple: the manager sat down with ChatGPT, asked a question, read the answer, adjusted, and got on with the work. Researcher Ethan Mollick named that phase co-intelligence and now, in a recent article, he says: that phase is over. Not because nobody uses a chatbot anymore, but because, in parallel, a class of AI emerged that does the entire task without you following step by step. Mollick calls the new stage co-existence: living alongside an AI that, on certain tasks, is better than your employee, and on others is still worse than the intern.
This post translates his idea into operational language. For the business owner and the department director, this is not a technology matter. It is about three new decisions that are going to enter your week: when it pays off to refuse AI's help, when it pays off to hand over the entire task to it, and what to do when AI also becomes the audience that reads what your company writes. Let's take it in parts.
1. What changed from 2024 to now, no jargon
Think of two different figures on the team. The first is the assistant you consult: you ask, it answers, you decide. The second is the employee who executes: you hand over the task, they disappear for a few hours, they come back with the finished work. In 2024, AI was almost only the first figure. In 2026, it also became the second, and on many tasks it is cheaper and faster than hiring.
The most documented example comes from software companies. Mollick notes that teams which adopted AI agents to write code are multiplying the volume produced by seventeen, and that inside Anthropic, the maker of Claude, most of the new code already comes out of the agent, with each developer delivering several times more than they delivered before. The exact number matters less than the pattern: AI stopped being a consultation tool and became an outsourced executor, with everything that has of good and of complicated.
What does this have to do with your company, even if you don't program? Everything. The same movement is arriving in customer service (agents that handle the entire ticket), in sales (agents that prospect, qualify and book meetings), in legal (agents that draft contracts), in finance (agents that close reconciliation). The question stopped being "how do I use ChatGPT?". The question became: "when does it pay off to let AI do it alone, and when does that get too expensive?".
2. New decision number one: when to refuse the help
It sounds strange, but it is the most underrated decision. On some tasks, even with good AI available, it pays off for your team to do it by hand. Not out of stubbornness, out of consequence. If you outsource to AI the reasoning that keeps your team sharp, a year from now the team no longer knows how to do it, and when AI fails (and it does fail), nobody in your house is able to fix it.
Where this weighs: strategic analysis, personnel decisions, reading an important client, a delicate negotiation. It is not that AI can't help in these contexts. It is that your team's training depends on them going through these decisions. Like the gym: the machine does the lifting for you, but you come out weaker. Refusing AI at selected points is a way to protect the company's intellectual capital, not a way to be conservative.
3. New decision number two: when to hand over the entire task
The opposite of the previous one. There are tasks where insisting on human control becomes expensive waste. Line-by-line bank reconciliation, first-pass resume screening, meeting transcription, drafting the first version of a standard sales email. These are tasks where AI today is faster, more consistent, and almost never worse than a tired human doing it on Friday at seven in the evening.
The delicate point, and where most companies get stuck: handing over the entire task is different from "letting AI suggest and the human approve". Human approval at high volume becomes a rubber stamp. The employee approves everything on autopilot, and AI's failure slips right through. If you are going to hand over the task, hand it over for real, with audited sampling afterward and a shut-off button for when the error rate climbs. Case-by-case approval, for low-risk tasks, is just friction that gives the illusion of control.
The practical criterion: if the cost of getting it wrong once is low (sending a generic email with a mistake), hand it to AI with sample auditing. If the cost of getting it wrong once is high (sending a contract with a wrong clause), keep a human in the middle, but pay that cost consciously, knowing you are paying to reduce risk, not to pretend the human is watching.
4. New decision number three: AI also became your reader
This is the part that catches the marketing manager off guard. More and more, the audience of your website, your post, your sales material, is not only human. It is an AI agent reading in place of a human who asked it to summarize, compare against competitors, or recommend. Mollick recounts that he adapted his own site to have a version meant to be read by AI, after discovering that old SEO tricks (hidden text for AI to see) no longer work with new models, which in fact identify these attempts as a suspicious instruction and ignore them.
Translating to your operation: if the potential client asks their AI to "find a supplier of X", the one answering them is the AI, based on what it managed to read from your site. A confusing site, with no clear numbers, no explicit use case, no reference price, becomes an invisible supplier. This changes the criteria for writing a page, a product description, a FAQ. It is not last decade's keyword SEO. It is writing in a way that an AI can summarize your proposition well for a human who is going to decide in thirty seconds.
5. Why this is an operations problem, not an IT one
The most common reading, and the most wrong, is to treat this transition as a technology project: hire a vendor, "implement AI", train the team, done. It is not. The three decisions above are operating policies, not a choice of tool. Who decides when to refuse AI is the department manager. Who defines what can be handed over entirely to AI is the process owner, together with risk and legal. Who rewrites the site with the AI reader in mind is marketing together with sales.
The technology vendor (us included) comes in afterward, to execute the decision. The expensive mistake is the reverse: buying AI first and then finding out, process by process, where it fits. It comes out more expensive, takes longer, and produces the classic frustration of "we invested in AI and I didn't see results".
6. What to do this week, without big projects
Three cheap moves you can run without hiring anything:
- Map five tasks in your operation and mark each one as "keep human for team training", "hand over to AI for cost", or "not sure yet". The "not sure yet" column is the most honest, and it is where a pilot makes sense.
- Ask an AI to summarize your own site (ChatGPT, Claude or Gemini) as if it were a potential client asking "is it worth hiring this company?". Read the summary with a critical eye. If the AI couldn't summarize it well, your site is written for old search ranking, not for today's reality.
- Define who decides among the three quadrants above. Who in your house has the authority to say "this task goes to AI, with a monthly sample audit"? If the answer is "nobody yet", that is the next seat to design, before buying any tool.
7. The reframe
The discussion inside companies is still stuck on "let's implement AI" as if it were installing a new system. Mollick is saying, and practice confirms, that the right metaphor has become another one: negotiating a coexistence with a strange collaborator who changes capability every three months, is better than your team at some things, worse at others, and who also started reading what your company publishes.
There is no ready-made solution. There is a new posture. Companies that understand this first will decide ahead of time where to refuse, where to hand over, and how to talk to an audience that is half human and half machine. Those that keep thinking about "which is the best AI tool for me to buy" will reach 2027 complaining about the same problem under another name.
Inspired by "Co-Existence and the End of Co-Intelligence", by Ethan Mollick, published in One Useful Thing. Read the original here.