Brainstorming, Planning and Decision Support with AI

Some of the most valuable uses of AI do not involve producing a finished document at all. They involve thinking: generating a wider list of options than you would have come up with alone, finding the hole in a plan, or laying out the trade-offs of a choice so you can see it clearly.

They also come with a particular hazard. The model is agreeable, confident and does not bear the consequences of what you decide. Used well, it sharpens your thinking. Used carelessly, it launders a hunch into something that sounds like analysis.

• How to brainstorm with AI without ending up with generic ideas.
• How to build and stress-test a plan.
• How to use AI for decisions while keeping the decision yours.
• How to counter the tendency of models to agree with you.
• Which decisions AI should not make.

Good brainstorming has two phases, and mixing them makes both worse.

1. Diverge: produce many options without judging them.
2. Converge: choose using criteria.

AI is strong at the first phase because it is fast and has no ego. Its weakness is sameness: if you ask for "ten ideas", you often get the ten most typical ones, which is also what everyone else gets.

• Write your own list first. Spend five minutes listing what you already think of. Then ask the assistant for ideas that are not on your list. This stops it from anchoring you on its first suggestions and keeps your thinking active.
• Ask for variety explicitly. "Give 15 ideas in five different categories: low cost, high effort, unusual, for a beginner, for an expert."
• Use constraints that provoke. "What if we had to do this with zero budget? With one day? With only students as helpers?"
• Ask for the worst ideas, then flip them. "List five terrible ways to onboard a new employee. Now invert each into a good practice."
• Use perspectives. "How would a sceptical customer, a new joiner and a finance manager each see this idea?"
• Use a known technique. For example, SCAMPER prompts ask you to substitute, combine, adapt, modify, put to another use, eliminate or reverse parts of an existing idea. You can ask the assistant to apply each to your product or process.

Cluster the ideas, remove duplicates and score the survivors against criteria you set. A table works:

You can ask the assistant to fill the table, but your judgment on impact and risk should override its estimates. It does not know your context.

A plan is a set of decisions about order, time and risk. AI can structure it, and you must supply the facts.

Goal: launch a simple portfolio website by 30 November.Constraints: I can work 6 hours a week; I know basic HTML and CSS; I have no budgetfor paid tools.Known tasks: choose three projects to show, write short descriptions, build thepages, deploy, share with five people for feedback.Please: (1) break this into tasks of at most 2 hours each, (2) put them in asensible order with dependencies, (3) show a week-by-week schedule that fits myhours, and (4) list the three tasks most likely to take longer than expected.

AI estimates of time and effort are guesses. People are also poor at estimating, and the combination can feel authoritative. Compare with your own history ("the last site took me 14 hours") and add a buffer for tasks you have not done before.

A pre-mortem is a technique where you imagine the plan has already failed and work backwards to the causes. It is a strong use of AI because the model can quickly list many plausible failure reasons:

Imagine it is 1 December and this plan failed: the website is not live. Listthe 8 most likely reasons, with a one-line preventive action for each. Includereasons related to my constraints (time, skill, motivation), not only technical ones.

Read the list and pick the two or three that apply to you. This gets you the benefit of foresight at almost no cost.

• Make the next action explicit. A plan is useful when its next step takes ten minutes to start.
• Review weekly. Ask: "Here is what I finished and what slipped. Adjust the schedule, and tell me what to cut if I only have 4 hours next week."
• Cut, do not stretch. If time shrinks, remove scope before extending the deadline silently.

For decisions, a good structure is:

The values part cannot be delegated. Whether salary, learning or location matters more depends on your life, and no model knows it.

Goal. Choose between two internships.

Weak request.

I have two internship offers. Which should I take?

You will get a generic list of "factors to consider" and perhaps a recommendation with no basis.

Better request.

Help me think through a decision. Do not tell me what to choose.Options:A: Startup, 3 months, stipend Rs 15,000/month, work on a real product, remote.B: Large company, 2 months, stipend Rs 30,000/month, structured training, on-site inanother city (I would pay about Rs 12,000 for rent).My criteria and weights (out of 100): learning and ownership 40, net money 20,career signal on my CV 25, practical hassle 15.Please: (1) show a table scoring each option 1 to 5 per criterion with aone-line reason, flagging where you are guessing; (2) list facts I should find outbefore deciding; (3) give the strongest argument for each option; (4) tell me whatwould change the ranking.

Example response (illustrative). A scoring table, with notes such as "net money: B Rs 18,000 vs A Rs 15,000, but only for two months", a list of questions ("Will A convert to a job offer? Will B pay for relocation?"), and arguments for each option.

Check and use.

1. Verify the numbers yourself in a spreadsheet. Net money for B is 30,000 minus 12,000 = Rs 18,000 per month, for two months, Rs 36,000 total, versus A at 15,000 x 3 = Rs 45,000. A model may have compared monthly figures and overlooked the duration. Check the arithmetic.
2. Treat its scores as hypotheses. It guessed at "career signal". Replace with your own view or ask people in the field.
3. Do the research on the facts it listed: talk to former interns, check the offer terms.
4. Make the decision, then write one sentence on why. A short decision journal (what I chose, why, what I expect) lets you learn from the outcome later.

Notice that the AI never decided, and that your own checking changed the picture: the higher monthly stipend looked decisive and total earnings favoured the other option.

Language models lean towards agreeable answers, as article 4 explained. For decisions, this is dangerous: if you reveal your preference, the model tends to build a case for it.

Counter it:

• Hide your preference. Present options neutrally, without saying which you favour.
• Ask for the strongest case against your favourite. "Argue as a sceptic that option A is a mistake."
• Ask what would change the answer. "Under what conditions would option B be better?"
• Use a second framing. Ask again with the options reversed or described differently, and check whether the advice flips with presentation.
• Talk to a person who will disagree. A friend, mentor or colleague gives you something the model cannot: a real stake in your outcome and knowledge of you.

Use AI to prepare (questions to ask, terms to understand, options to consider), but not to decide, in areas like:

• Health: diagnosis, medication, whether to seek care.
• Legal matters: your rights, contracts, disputes.
• Money with serious consequences: investments, loans, tax filing positions, major purchases.
• Safety: anything where a wrong answer can cause physical harm.
• Decisions about other people: hiring, firing, grading, disciplining. Models can reflect biases, and the consequences fall on someone else. Bias, Copyright and Human Judgment says more.

For these, use authoritative sources and qualified professionals.

• Using the first ten ideas without asking for variety.
• Letting the AI set the criteria, and not noticing they are not yours.
• Trusting estimates and scores it had to guess.
• Revealing your preference and receiving agreement.
• Skipping the pre-mortem because the plan looks good.
• Sharing personal or confidential details about yourself or others in a decision prompt.
• Delegating the decision.

Imagine it is 15 December and my plan to learn SQL well enough for interviews bythen failed. Context: I work full time, I can study 5 hours a week, I tend to losemomentum after two weeks, and I have no one to practise with. List the 8 mostlikely reasons it failed, ordered by likelihood given this context, and one smallpreventive action for each.

I am considering leaving my job to join a startup. Without telling me what to do:1. List the strongest arguments for staying and the strongest arguments for going,   with equal effort on each side.2. List the facts I should find out first (about the startup's funding, my notice   period, my savings runway).3. Describe what a regret-minimising person and a risk-averse person would each   think.4. Tell me what would make going clearly a mistake.

• Next: AI Coding Assistants: Modes, Strengths and Your Responsibility begins Module 5, for developers.
• For career decisions in particular, see Job Descriptions, Skill Gaps and Portfolio Projects with AI and When Should You Switch Jobs? A Decision Checklist.
• For building a repeatable planning routine, see Design Your Personal AI Operating System.