A long chat with an AI gives you more ideas and no decision. Here's a four-phase session that ends with one idea chosen, one test designed and a clear line on what you still decide yourself.
Don’t ask for a hundred ideas in one chat. Write your own ideas first, run two or three AI rounds, cluster everything by the bet it makes about the customer, challenge the shortlist in a fresh conversation, then score the finalists and pick one. AI generates, groups and attacks. A named person scores, chooses and owns the test.
The first ten ideas from an AI brainstorm are usually the ten anyone would get from the same prompt. They’re sensible, they’re well worded, and they read like a list a colleague would also have produced. The usual cause isn’t the tool. It’s a session with no structure, where a long chat keeps producing and rarely decides.
A marketing lead asks an assistant for fifty ideas to win back trial users who went quiet after the first fortnight. The list arrives in seconds. She scrolls it, copies four that sound reasonable into a doc, and brings them to the next meeting, where most of the time goes on the same questions the team would have argued about without AI: which of these is realistic, who would do it, and how would we know it worked. Nothing was wrong with the ideas. Nobody had decided what the list was for.
Three things tend to go wrong in an unstructured session, and they compound each other:
What works better is a session with four phases, each with one job. Diverge widens the pool. Cluster shows you what you actually have. Challenge attacks the survivors. Converge picks one idea and turns it into a small test. AI does different work in each phase, and so do you. The rest of this guide is the script, with prompts you can paste and a scorecard that ends in a decision.
I’ve written it for the two situations I’d expect most marketers to be in: working alone, and running a session with a team in the room. Where they differ, I say so.
Before the first phase, write the question in one sentence and name the constraint that makes it real. “How do we grow?” produces noise. “How do we get trial users who went quiet in week two to come back, using email and in-product messages only, with no discount?” produces ideas you can compare. If the problem is bigger than a campaign, such as which market to enter or which product to back, start with the structured analysis in our guide to using AI for strategic planning and bring the brainstorm in afterwards.
A Future Factors session design based on our own judgment about where AI helps and where a person has to decide. It is not a research-validated method.
Keeping the phases apart matters because they pull in opposite directions. Diverging rewards quantity and tolerance for bad ideas. Challenging rewards suspicion. If you do both in one chat, the tool tends to defend what it just wrote, and you tend to stop reading critically once the list looks long. Treat each phase as a separate sitting, even if the sittings are twenty minutes apart.
The session should end with three things: one idea chosen, one small test designed, and a short list of ideas parked on purpose. A pile of ideas isn’t an output.
Phase one starts without the tool. Spend five to ten minutes writing your own ideas on a page, including the bad ones. This isn’t a ritual. Once a model’s list is in front of you, it becomes the frame you think inside, and your own half-formed ideas look small next to forty tidy ones.
There’s some research behind the instinct, with a limit worth stating. In a Science Advances experiment on creative writing, people given AI-generated ideas produced work rated as more creative individually, while the stories across the group became more similar to each other.[1] That was a writing task, and the finding is about averages across many writers, so I’d read it as a reason to protect your own starting point and not as a verdict on AI brainstorming.
Your ideas go on the page before the AI’s do.
Your own list is worth more than its length suggests, because it holds the material the model doesn’t have: an objection a prospect raised last month, a complaint that keeps turning up in support tickets, something a competitor did that annoyed you. Put those in, even as fragments. They’re the raw material for the ideas that end up feeling specific to your business.
Then brief the tool. The prompt below does the work a vague “give me ideas” request leaves out: it gives the problem, the limits and what you’ve already tried, it shows your own ideas so the model doesn’t repeat them, and it asks for different families of approach so you don’t get fifteen versions of one.
| I’m a [role] at [type of company]. The question: [one sentence]. The audience: [who they are and what we know about them]. Constraints: [channel, budget, timing, things we can’t say or do]. Already tried: [list]. My own ideas so far: [paste]. Give me 15 new ideas. Group them under five different families of approach (for example timing, community, product, partnership, content; rename the families to fit the question). No two ideas may rely on the same mechanism. Skip anything on my list or in ‘already tried’. For each idea, write one line on how it would work and one line on what it assumes about the customer. Don’t rank them and don’t recommend any. |
Run it three times. Round two: add ‘Now give me 15 that a cautious head of finance would approve.’ Round three: ‘Now 15 that would make our team uncomfortable but are still legal and on-brand.’ Delete the obvious ones after each round and add anything the list sparked in you.
Stop when a round starts repeating clusters you already have, not when you hit a number. For most questions that’s two or three rounds, which is a few dozen ideas. More than that mostly adds variations on bets you’ve already seen.
A list of fifty ideas is hard to judge because the ideas aren’t independent. Many are the same bet in different clothes: three of them assume the customer stalled on effort, four assume they doubted it would work for their kind of business. Judging them one by one means judging the same bet repeatedly, and it favors whichever idea is worded best.
Cluster ideas by the bet they make about the customer, not by topic.
Grouping by topic gives you “email ideas” and “event ideas,” which tells you nothing about whether they’d work. Grouping by bet gives you something you can check against what you know. Hand the whole list to the assistant in one go, because grouping is a job it does quickly and well, and keep the naming for yourself.
| Here are all the ideas from our session: [paste every idea, yours included]. Group them into 4 to 7 clusters. Name each cluster by the bet it makes about why the customer behaves as they do, written as a sentence (‘they stalled on effort, not interest’). List which ideas sit in each cluster. Flag any idea that fits two clusters and any cluster where the ideas are near-duplicates. Then list any bet we haven’t covered at all. Don’t rank, don’t drop anything. |
Paste your own ideas in with the AI’s, unlabelled, so the grouping treats them equally.
Expect to edit the clusters. The tool will sometimes split one bet into two clusters because the ideas are worded differently, or lump two bets together because they share a channel. Merge, split and rename until each cluster is a sentence you could check against something real, such as a support theme, a survey comment or a sales note. Where you can’t check a bet, mark it as a guess, because that tells you where the cheapest test is.
Here’s what the result looks like for the trial-user question, after a person has edited it. The last column is the part the tool can’t fill in.
| Cluster | The bet it makes | Ideas inside | Your call |
|---|---|---|---|
| Lower the effort | They stalled on setup, not interest | Starter template, short onboarding call, import wizard | Keep. Cheap to test. |
| Show someone like them | They doubt it works for a business like theirs | Peer story email, peer webinar | Keep. Needs a customer’s permission. |
| Give a reason to act now | Interest is fine, urgency is missing | Event deadline, extended trial after one completed task | Keep the second only. |
| Ask why they left | We’re guessing, and a direct question beats a guess | One-click reply question, short exit survey | Keep. Fastest way to test every other bet. |
| Discounts | Price is the barrier | Three variants | Park. It cuts against the pricing page. |
An illustrative table for a made-up trial-user question. Parking a cluster on purpose, with a reason, belongs in the output as much as keeping one.
By now you have a shortlist of three or four clusters, and every idea on it still has the shine of being new. Phase three is where the session earns its keep, because the cheapest time to find out an idea is weak is before anyone has built it.
Open a fresh chat for this phase and paste in only the question, the constraints and the shortlist. It’s a judgment call on my part and not a measured result, but a model that has spent an hour helping you generate ideas tends to be generous about them, and a clean conversation without that history gives you a harder reader. The prompt asks for specific objections and a cheap test, because “this might not work” helps nobody.
| Act as a skeptical head of marketing who has watched ideas like these fail. The question: [one sentence]. Constraints: [list]. What we already know about the customer: [paste any real data, support themes or sales notes]. The shortlist: [paste]. For each idea: (1) the strongest reason it would fail with our customers; (2) what we would have to believe for it to work; (3) the cheapest test that would settle it within two weeks; (4) whether anything in what I’ve told you already contradicts it. Be specific and don’t soften. Don’t suggest new ideas and don’t rank. |
Giving it real customer facts matters more than the persona. If you have none, say so in the prompt, because then every answer is a guess and the test column becomes the main output.
Now you do the part only you can do, which is to sort the objections into three piles:
| Idea | Objection the AI raised | Verdict | What changes |
|---|---|---|---|
| Peer webinar | Quiet users rarely open event invites, so attendance will be thin | Real | Record a short customer conversation and email it instead. |
| Starter template | Templates get ignored unless they match the user’s use case | Answerable | Offer two or three, chosen by the answer to one question at sign-up. |
| One-click reply question | Reply rates from lapsed users are usually low | Generic | Nothing. Low volume still tells us the top reasons. |
An illustrative sort. The verdict column is where lived knowledge, such as what sales hears from lapsed users, outranks the model’s confidence.
If a real customer-facing person, such as someone in sales or support, can look at the same shortlist for ten minutes, ask. Their objections will be better than the model’s, and they’ll often name the one the model didn’t raise.
Phase four is where most AI brainstorms quietly stop, because choosing feels like the human’s job and nobody builds a method for it. A short scorecard turns “I like this one” into something a team can argue about. Score each finalist from one to five on the five questions below.
| Criterion (5 is best) | One-click reply question | Starter templates | Peer story email |
|---|---|---|---|
| Fit with what we know about the customer | 4 | 4 | 3 |
| Evidence we already have | 3 | 3 | 3 |
| Size of the upside if it works | 3 | 3 | 4 |
| Cheapness of a first test | 5 | 4 | 4 |
| Safety for the brand and for the rules we work under | 5 | 5 | 4 |
| Total | 20 | 19 | 18 |
Illustrative scores. Read the totals as 20 or more: run first; 15 to 19: run if you have capacity; under 15: park with a reason. The bands are our rule of thumb, and the point of the numbers is to start the argument, not end it.
Let the assistant add up the columns and draft the test plan, but score the rows yourselves. Two of the five criteria, customer fit and evidence, depend on what you know that the model doesn’t. If you want a second opinion, ask the tool to score the same table separately and compare. Where its numbers differ from yours by two or more points, you’ve found the place worth a conversation.
AI can rank the options. One named person chooses.
One habit makes the scoring more honest: score one criterion for every finalist before moving to the next, rather than scoring each idea start to finish. Otherwise the idea you like gets a halo that carries across all five rows. If two finalists tie, the cheaper test wins the first slot.
A close call between two ideas is fine. Run the cheaper one first and write down what result would make you run the other. What isn’t fine is leaving the session with a shortlist and the sentence “let’s think about it,” because that’s how a brainstorm turns into another meeting.
The four phases stay the same whether you’re alone or in a room. The risks differ, and so does who should hold which seat.
The risk is that the AI becomes the only other voice in the room, so every idea gets tested against one reader with no stake in your market. Split the session across two sittings, with phases one and two in the first and three and four in the second, so you come back to the shortlist cold. Use the fresh-chat challenge pass, then put the shortlist in front of one real person before you score it. A one-line message to a sales or support colleague is enough.
Production blocking is the risk. In a set of experiments published in 1987, groups brainstorming together came up with fewer ideas than the same number of people working separately, and the main explanation the authors supported was that people had to wait their turn to speak.[2] Those were lab groups of an earlier era, so it doesn’t predict how your hybrid team behaves. It does support a practical habit: have everyone write ideas privately and in parallel before anyone shares.
The other team risk is easy to recognise. Someone puts the AI’s list on the shared screen, and the room starts reacting to the list. The quieter people’s ideas often don’t get said, because the screen has set the agenda. Run phase one on paper or in a shared doc first, add the AI’s round second, and cluster everything together.
For the team version, the prompts are worth saving where everyone can reach them. Our guide to building an AI prompt library for your team covers where they live, and the split of work between roles is covered in how marketing teams divide the work with AI.
The idea that wins a session like this often isn’t in any of the model’s rounds. It comes from someone who has sat through the support calls, run the last three campaigns, or watched a customer abandon the product halfway through setup. That’s why phase one starts without the tool, and why the challenge phase needs a person who’s close to customers. In my judgment, the model’s job is to make sure that idea gets compared fairly, not to produce it.
Where AI does real work in a session, it’s worth writing down what stays with a person. This is the split I’d use.
| Phase | What AI does | What the human still owns | How it gets checked |
|---|---|---|---|
| Diverge | Generates rounds of ideas in different families | Writes first, and adds the idea that comes from lived knowledge of the customer | Your own ideas are on the page before the first prompt |
| Cluster | Groups ideas and flags duplicates and gaps | Names the bets, merges or splits clusters, parks some on purpose | Each cluster’s bet is checked against customer evidence |
| Challenge | Raises specific objections and cheap tests | Decides which objections are real, answerable or generic | A customer-facing colleague sees the shortlist |
| Converge | Adds up scores and drafts the test plan | Scores the rows and picks one idea | The decision log names an owner and a review date |
The person closest to the customer usually has the idea the model can’t produce, and phase one is built to leave room for it.
Then record the decision somewhere that outlives the meeting. This is the log for the winning idea from the example above, filled in.
| Field | Entry |
|---|---|
| Question | How do we get trial users who went quiet in week two to come back, with email and in-product messages only and no discount? |
| Chosen idea and its bet | A one-click reply question in the follow-up email. We’re guessing why they left, and asking is the fastest way to test every other bet. |
| Test | Send to one segment of lapsed users for two weeks. Tag each reply by reason. |
| What would change our mind | If replies cluster around one reason, run the matching idea next. If almost nobody replies, switch to the starter template. |
| Owner and review date | Lifecycle marketer. Review in two weeks. |
| Parked and why | Discounts (cuts against the pricing page). Peer webinar (quiet users rarely attend). |
The “parked and why” row is what stops the same ideas coming back next quarter as if they were new.
You can judge the process in the three stages we use for any new way of working. Use: did the session end with a chosen idea and a test? Persistence: are you still running it a month later, and did the parked list get reused? Impact: did the tested idea change a decision or a number? A team can run all of this internally from the prompts above. A workshop shortens the learning curve and gets a whole team using the same phases and the same scorecard, which is hard to do with one enthusiastic person and a shared doc.
For this week, pick one question that’s stuck. Block ninety minutes, or two sittings of forty-five. Write your own ideas first, run the three rounds, cluster, challenge in a fresh chat, score, and write the decision log before you close the doc.
Ask for around fifteen per round and run two or three rounds, each with a different lens. Stop when a round mostly repeats clusters you already have, not at a target number. A bigger single request tends to give you more variations of fewer bets, which is the problem the phases are built to fix.
Phases one and two work fine in one conversation. Start a new chat for phase three and paste in only the question, the constraints and the shortlist, so the challenge doesn’t come from a model that’s been agreeing with you for an hour. That’s our judgment, not a measured effect, and it costs you two minutes.
Use it as a second scorer, not the first. Score the table yourself, ask the tool to score it separately, and look at where you differ by two points or more. Those gaps usually point to something one of you knows that the other doesn’t, which is more useful than either set of totals.
Describe the customer and the problem in general terms, leave out names, figures and anything covered by an agreement, and use the tool your company has approved. The prompts work with a description like ‘a mid-size software company with a free trial’ as well as they do with the real details. Check your own policy first when customer data is involved.
When the answer is a fact you can look up, when the decision has already been made, or when only one option is allowed. A brainstorm needs a real choice at the end. If the options are fixed, spend the time on the scoring and the test plan instead.
The two studies cited here were checked against primary or repository records on October 5, 2026. Diehl and Stroebe’s 1987 findings on production blocking come from the paper’s abstract and conclusions in an author-hosted copy. The Doshi and Hauser Science Advances paper was read through its UCL repository record, because the journal site blocked automated access, so I’ve described its direction of effect and left out its effect sizes. The four-phase session, the three rules, the scorecard bands and the decision log are Future Factors working methods based on our own judgment, not research-validated scales. I left out the widely quoted figures from other AI ideation studies, since this guide doesn’t depend on them. The scenes and tables in this guide are illustrations, not client stories, and the scores are examples.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.