Innovation in the age of AI

Coming up with ideas used to be the expensive part, but today, in the age of AI it isn’t. Your competitors can now easily generate fifty product ideas or directions before lunch, and they’re doing it with the same three or four models that everyone is, all trained with the same data. What’s scarce isn’t an idea — it’s an idea the models aren’t also handing to everybody else. Your advantage lives in how you frame the problem and in the customer contact that feeds it.

And there’s a second effect. Thinking that used to happen in a room with other people now happens in a chat window. It feels like collaboration, because something answers. But the thing answering has no stake in the outcome, no memory of the customer call you were both on, and no reason to keep telling you that you’re wrong.

Both these effects push in the same way — toward the median. The encouraging part is that the fix mostly isn’t new: the innovation frameworks that predate AI by decades were built around exactly the mechanisms a private conversation with a model removes.

The models are homogeneous, whatever persona you give them

Ask one model to act as a skeptical CFO, a growth marketer and a security engineer, and three different answers come back. The content is different, the priorities change, the objections are new. What doesn’t change is where any of it comes from: one set of weights over one corpus of data, with one sense of what counts as a reasonable next sentence. A persona changes which part of that corpus you draw from. It doesn’t add a second one.

Other people in the room provide uniqueness. They have independent experiences and history, and a stake in what gets decided, which is what makes them keep arguing. None of which is available by prompting.

This shows up in measurement, not just in principle. In a controlled experiment published in Science Advances, 293 writers produced short stories, some with access to AI-generated story ideas, and 600 judges evaluated the results. Individually, AI helped: novelty and usefulness ratings rose, and rose most for the writers who scored lowest on creativity to begin with. Stories by writers offered a single AI idea were measurably more similar to the rest of their group. The increase equalled about 10.7% of the range seen among unassisted writers.

That’s the trade off, and it’s easy to miss because only one side of it is visible from inside a team. Each person’s output gets better. The set of things the company might do gets smaller.

The same pattern appears at the level of a single session. In a study of creative ideation, different users produced less semantically distinct ideas with ChatGPT than with a non-AI creativity tool, and that ChatGPT users felt less responsible for the ideas they generated. Ownership is critical in a startup, where ideas survive because somebody refuses to drop them. Nobody fights hard for something they don’t feel they came up with.

The competitive version of all this is blunt. If you and your three closest competitors each ask a model how to differentiate, you will get similar answers.

We already have the frameworks

Innovation doesn’t need a new framework for the AI era. It needs us to use the existing ones, because each was designed to produce precisely what a solitary chat session removes. Here are four that are worth knowing.

Design sprints. Five days, one question. Developed at Google Ventures and documented in Sprint. Monday maps the problem and picks a focus, Tuesday sketches competing solutions, Wednesday makes the hard decisions and turns the winner into a testable hypothesis, Thursday builds a high-fidelity prototype, Friday tests it with real customers. The detail that matters most is easy to overlook — Tuesday’s sketching is done individually and in silence by people who then present. The sprint is a group process built around private work.

The Double Diamond. Published by the UK Design Council in 2003: Discover, Define, Develop, Deliver. Two diamonds rather than one, because there are two things to diverge on — the problem and the solution — and teams under pressure skip the first. Generating twenty solutions to a problem nobody has framed is the most reliable way to be busy and wrong.

Creative Problem Solving (CPS). The oldest of the four. Brainstorming dates to Alex Osborn in the 1940s; CPS was formalized at Buffalo State College with Sid Parnes and the Creative Education Foundation. Clarify, Ideate, Develop, Implement — with divergent and convergent thinking treated as separate acts, “the gas pedal as Divergence, and the brake pedal as Convergence,” and one explicit rule: defer judgment while diverging, because instant judgment in response to an idea shuts idea generation down.

Time-boxed builds. Hackathons, makeathons, ship-it days. No canonical method — a hard deadline, teams that form themselves across existing boundaries, and demonstrations at the end. They give you range and cross-team contact, but they don’t force a decision. That makes them a poor substitute for the other three and a good complement to them.

What innovation actually is

Strip the three structured frameworks down and the same five elements are in all of them. Time-boxed builds supply only some of the five, which is why they complement the others rather than replace them.

1. A framed problem, written down. Discover, Clarify, Monday’s map, the hackathon theme. The framing is the work; most of what follows is comparatively mechanical.
2. Divergence and convergence kept apart. Two modes, never at the same time, with an explicit rule protecting the first from the second.
3. Independent generation before discussion. Sketching in silence, deferred judgment, ideas written before they are spoken.
4. A forced convergence. A deadline and a named decision, because an idea nobody has chosen or taken ownership of never gets acted on.
5. Contact with reality. Friday’s customer test, Deliver, the demo. The output is a tested hypothesis, not a list.

Idea generation, and why AI isn’t a good fit

Item three is the one where most people want a model to take over, so it’s worth being precise about what it does to that step.

The evidence on independent generation of ideas predates the internet and gets misread constantly. Across four experiments, Diehl and Stroebe found that people generating ideas separately and pooling the results outproduced the same number of people brainstorming together. A major cause was idea generation blocking: only one person can talk at a time, so everyone else waits instead of thinking. The lesson isn’t that groups are useless. The group’s job is convergence, and generation belongs to individuals first.

AI looks like the cure here. Nobody interrupts you, the output is immediate, and the volume is enormous. But what arrives is drawn from the same well every competitor is drawing from. It removes the blocking and the variance in one move. But the variance was the part that made independent generation worth doing.

A person alone with a model is also missing three things the framework supplies, and each maps to one of the five elements. The first is pushback. A model can be prompted to argue, and the argument will be well formed, but it has no stake in the result and concedes on the first push, so the convergence step never meets real resistance. The second is independent information. The model can’t be the one who says the last two customers asked for the opposite, which is what contact with reality depends on. The third is commitment. People execute what they helped choose and quietly deprioritize what they were handed, so a decision made alone with a model tends not to get carried out.

What to do as an early-stage startup

You can’t stop for five days with eight people and a runway, and you don’t have to. The five elements compress into a single day, and most of the value sits in the first two steps.

The distinction to hold onto is between work where sameness is free and work where distinctiveness is the product. Research, synthesis, turning forty support tickets into five themes, generating variants, building the prototype, writing the test script — sameness costs you nothing there, and the leverage is real. Framing the problem and choosing the bet are the other kind. That’s where your advantage either exists or doesn’t, and it’s the last thing to delegate to a system that’s giving everyone else the same advice.

Common questions

Does using AI to generate ideas make a team less innovative? +

It makes individuals more productive and the group more similar. In a controlled experiment with 293 writers and 600 judges, access to AI-generated ideas raised novelty and quality ratings for individual work — most for the least creative writers — while stories by writers given a single AI idea were measurably more similar to the rest of their group, an increase equal to about 10.7% of the range seen among unassisted writers. For a company that trade is worse than it is for one writer: the gain lands on each person’s output, and the loss lands on the range of things the company might do, which is where competitive difference comes from.

Can AI personas replace having a diverse team? +

No. A persona prompt changes which part of one model’s training data gets sampled; it doesn’t add a second source of experience. Every persona shares the same corpus, the same blind spots and the same absence of any stake in the outcome. A colleague brings information nobody else has, an independent history, and a reason to keep arguing after the first pushback. Personas are useful for pressure-testing a plan against predictable objections, and they are not a substitute for people who have to live with the decision.

How do you stop AI from homogenizing your team's ideas? +

Fix the order and fix the inputs. Generate first and alone, before anyone opens a model, so the first pass comes from human experience instead of a shared starting point. Let different people approach the model from different angles rather than pooling around one prompt. Feed in material the model cannot have — customer quotes, support tickets, your own operational numbers — because that is the only input a competitor cannot reproduce. Then use AI heavily for the parts where sameness is free: research, synthesis, prototyping and testing.

What is the lightest innovation process a small startup can run? +

One day, structured like a compressed design sprint. Morning: a written problem statement, then fifteen minutes of silent individual idea generation, then pooling and clustering. Afternoon: converge on one testable bet with a named owner and a date. Within the same week: a prototype and five customer conversations. The two non-negotiable parts are the written framing at the start and the decision at the end — without the first you solve the wrong problem, and without the second the team learns that the exercise leads nowhere.

---

Setting the terms for how AI is used across your organization is a separate question from how a team generates ideas, and it’s the work of an AI transformation engagement. If the harder problem is how ideas and decisions move through your leadership team, that’s executive coaching.