AI agents in group chats work only if the process they're supposed to manage is already clear and well-defined. I learned this when I turned down automating a process because it was broken; automating it would have just made the mess faster. An agent coordinating travel plans or event logistics will run into the same wall. The real work comes first, before the software.
Instinct announced last week that its AI agent can now join group chats, even if not everyone in the chat uses the app. The feature is designed for the things groups actually do: planning trips, sorting out who's bringing what to an event, coordinating carpools, running fantasy leagues. It sounds useful. But I've spent enough time on the wrong end of broken automation to see the trap. The feature only works if the underlying process is already clear. If your group is still figuring out how to make decisions, adding an AI agent doesn't solve that. It just makes the confusion faster.
I once turned down a job because the process was broken
A potential client asked me to automate their workflow. I looked at how they actually worked, and I said no. Not because the automation was hard. Because automating their process would not have fixed anything. It would have only made the mess run at machine speed instead of human speed.
The problem was that they had no agreement on how the work should actually be done. Three people did the same task three different ways. There were exceptions to the exceptions. What one person called a finished job, another person wouldn't accept. No process existed yet, only habit and improvisation.
I told them what they needed first was to agree on one way of doing the work. Write it down. Test it. Make sure it actually works for the cases they care about. Only then does automation make sense. Because software has to encode one clear way of doing things. It cannot do ambiguous. Feed it ambiguity, and it will amplify it.
A group chat planning a trip is the same problem
Imagine five friends trying to book a trip together. One person wants to settle on a destination first, then dates. Another wants to lock dates, then find a place. Someone keeps changing their mind about budget. No one has said who's actually making the final call. The group chat is ninety messages of back and forth, and no one is actually closer to a booking.
Now add the Instinct agent. What exactly is it supposed to do? It can surface options. It can check availability. It can keep track of preferences. But it cannot resolve the actual disagreement: how the group makes decisions. If the group has no way of saying "this is settled, we move forward", the agent cannot invent one. It can only keep proposing variations on the same argument.
The agent will move faster than the group would on its own. Speed is not always an improvement. In fact, it can make things worse. Speed at moving in the wrong direction is just faster regret.
Systems need to know when to stop
I run a content site called BabyJunctions where articles publish with no human writing them. I learned something early on that stuck with me: an unattended system needs to be better at stopping than at starting.
When I first set it up, the system could publish. It could not say no. If a step in the pipeline went wrong, the next step would take whatever it got and keep going. Bad data in meant bad articles out. The first time I found out about it was when a reader told me so.
I rebuilt it the other way. Every step now has to verify that its own work is good. If it cannot, nothing ships. I get told instead. Skipping a day costs almost nothing. Publishing something wrong under my own name costs a lot.
A group chat agent should work the same way. When should the agent stop and say "I need all of you to agree on this before I move forward"? That has to be clear before you add the agent. If it is not, the agent will just keep pushing the group toward decisions they have not actually made.
The harder part is not the software
I have spent time connecting systems that should work together. An ERP linked to accounting. Inventory to orders. The connection itself is straightforward. The hard part comes after: the data does not match. A product has three different names and codes across the systems. The names have changed over time. There are one-off codes no one remembers the reason for. Cleaning that took longer than building the connection.
A group chat with an AI agent has the same kind of problem. The software is not the issue. The issue is: what does your group actually agree on? If you cannot answer that in plain language, you cannot ask software to enforce it. The agent needs to know what "success" looks like for a plan. It needs to know who decides. It needs to know when something is actually locked in versus when it is still open for discussion. Those are not software questions. They are process questions. And they have to be answered first.
What you need before you add an agent
- Agree on the decision rule. Not "we all feel good about this". Something testable: "the majority picks it" or "whoever books earliest wins" or "the organiser decides and everyone else confirms".
- Write down the sequence. What happens first? When do we need everyone's availability? When is the decision actually made, and who knows it is final?
- Plan for disagreement. What happens if someone's budget or dates shift after the group decides? Can they veto it or just drop out? That matters before you automate anything.
- Test it with humans first. Run the sequence a few times without the agent. Find the places where the process breaks or where people actually ignore the rule you thought you agreed on. Fix those.
Only then does adding an agent help. Because the agent's job is to speed up something that already works, not to invent a working process.
Instinct has built something that works technically. But for features like this, technical is the easy part. The hard part is in your group, not in Instinct's servers. And it has to be done before the agent arrives.
