AI should plan with your group, not for it

Ask an AI assistant to plan a weekend away and you get a tidy answer in seconds: a route, three places to stay, a list of things to do. Ask it to plan a weekend away for eight friends and the answer looks just as tidy. That is the problem.

A plan for one person is a puzzle. A plan for a group is a negotiation. The hard part was never finding a beach; it was agreeing on which beach, which weekend, how much everyone can spend, and who is actually coming. AI is very good at the first kind of problem and, left alone, quietly bad at the second.

What AI is genuinely good at

Generating options is cheap for a model and expensive for people. Ten weekend ideas within three hours of Nairobi, each with a rough cost and what it suits, is a minute's work for an assistant and an evening's work for whoever volunteers to "look into it".

AI is also good at the boring middle of planning: turning "let's go to the coast in December" into a checklist, spotting that the dates clash with a public holiday, drafting the message that asks everyone to confirm by Friday. None of this needs taste. It needs patience, and software has plenty.

What it gets wrong about groups

A model plans from what it is told. In a group, the most important facts are the ones nobody says out loud: one person can't afford the nicer option, another is only coming if their friend comes, a third has quietly decided they won't drive. A confident, polished itinerary can steamroll all of that, because it looks finished.

It also flattens disagreement. If an assistant picks the "best" option, the group stops talking and starts complying, or someone feels outvoted by a machine. Neither ends with a good trip.

Four lines a useful AI shouldn't cross

1. It suggests; the group decides. Options, trade-offs and a clear "here's what you'd give up" are useful. A final answer chosen on the group's behalf is not. The decision, and the moment of agreeing, belongs to the people going.

2. It says what it is. Anything drafted by AI should be labelled as such, so nobody mistakes a suggestion for something a friend wrote or a price someone checked. Guidance on trustworthy AI, such as NIST's voluntary AI Risk Management Framework, is about building AI people can trust; being plain about what a machine made is the simplest place to start.

3. It is careful with what it is told. Group plans carry personal details: names, phone numbers, who owes what, sometimes health or family reasons for a choice. In Kenya, the Data Protection Act 2019 gives people rights over their personal data, enforced by the Office of the Data Protection Commissioner. An assistant should see only what it needs to help, and nothing it doesn't.

4. It never handles the money. Suggesting a budget is fine. Moving money is not something to hand to a model's judgement. Contributions, refunds and payouts should go through rules people agreed to and can see.

What good looks like in practice

Picture the same eight friends with an assistant that respects those lines:

• Someone asks for coast ideas for the first weekend of December. The assistant offers four, each with a rough cost per person and travel time, and marks them as AI suggestions.
• The group votes. Two people quietly pick "cheaper, please"; the assistant reworks the shortlist around a lower budget without naming who asked.
• Once a place is chosen, the assistant drafts the checklist and the message asking everyone to confirm and contribute by a date. A person reads it and sends it.
• The money goes into a shared pool everyone can see, through a payment service, under rules the group set. The assistant can remind people. It can't touch the pool.

The trip that comes out of this isn't the one the AI would have picked. It is the one the group agreed on, which is the only kind worth planning.

The point of the tool

The best planning tools make the group chat quieter, not smarter. They take the busywork off whoever always ends up organising, and leave the parts that matter, choosing and showing up, to the people involved.

That is the test we hold ourselves to as we build AI into GrupChat: it should save the organiser an evening and cost the group nothing in say. If it ever makes the decision for you, it has done the wrong job.