AI’s effect on the front end of innovation

When AI drives up the number of ideas at the front end of your innovation funnel there are two problems we can identify. The first is lack of diversity and variation of the ideas and secondly that the effort in innovation shifts to selection, prioritisation and validation in the funnel.

Innovation performance is impacted because there won’t be enough expertise and resources available to perform that task efficiently to the same quality standards. In this article we dive deeper in both the challenge and the solution.

Generating an idea is now almost free

In a Wharton experiment, GPT-4 produced about 800 ideas an hour. A student working alone produced about 20. Each idea cost roughly $0.63 from the model and $25 from the student. Buyers’ stated purchase intent was 46.8% for the model’s ideas and 40.4% for the students’. Of the 40 best ideas in the pooled sample, 35 came from the model (Girotra and colleagues, 2023).

The experiment covers one task and one group of ideators, so treat the exact figures with care. While AI driven ideation seems attractive, meaningful innovation is more involved, it’s an orchestrated approach that gathers data, turns that into insights and lets the insights drive innovation. But the direction is clear. As early as 2023 the authors wrote that the critical task in innovation practice “may shift from idea generation to idea evaluation and selection.” They added that AI models “do not yet appear to be particularly well suited” to that task.

Consider what this means for a moment. When ideas were expensive to produce, the effort of producing them filtered the pipeline before any committee reviewed or selected the best ones. A weak idea often died because nobody had time to do the appropriate discovery, document the findings and find support for it. That filter is gone, and every weakness in your selection process now meets a much larger volume of ideas at the first selection gate.

More ideas does not mean more choice

A wider funnel helps only if the ideas are significantly different from one another. A study in Nature Human Behaviour found that ChatGPT raises the creativity of individual ideas while reducing the diversity of ideas across a pool (Meincke, Nave and Terwiesch, 2025).

Innovation teams describe the same pattern. Researchers writing in Harvard Business Review observe that every team now uses the same few AI models. Some teams report a creative renaissance. Others report “a flood of homogenized, forgettable ideas that all sound like they came from the same person” (De Freitas and colleagues, August 2026).

Earlier work found that standard GPT-4 ideas lack the diversity of human brainstorming groups, and that prompt design changed this. Prompts that made the model reason step by step gave the most diverse output and came close to matching human groups (Meincke, Mollick and Terwiesch, 2024).

Insight driven innovation, using diverse perspectives to inform ideation is the counter to that deficiency. We will cover this in an upcoming article.

Which ideas get a pass to the next gate

Selection is where the effort now increases, and the research on human evaluators is showing worrying signs. A randomised study at a leading research university covered 2,130 evaluator-proposal pairs. Evaluators “systematically give lower scores to research proposals that are closer to their own areas of expertise and to those that are highly novel” (Boudreau, Guinan, Lakhani and Riedl, 2016).

The setting was research grants. An innovation portfolio review faces the same task: judging novel ideas on limited information, including ideas that sit close to the reviewer’s own field.

AI potentially adds a second bias. The same authors argue that polished AI-generated pitches can be mistaken for better ideas, which produces false positives among human evaluators. Novel ideas are marked down, and ideas that read well are marked up. False positives have cost consequences if they are only filtered out in later stages when resource commitments increase by default. Understanding how to address and prevent this bias is what innovation teams need to solve.

AI makes evidence cheaper and the criteria more important

AI does reduce the screening load when combined with reweighting past expert ratings (Kireyev and colleagues, 2025). Once the system is properly trained it stands to reason that the top (great ideas) and bottom (bad ideas) scores can be trusted while the middle ratings deserve more scrutiny. The aim is to reduce the volume of ideas that fall within the middle rating category. That confidence needs to be built over time as ratings and models improve.

De Freitas and colleagues name another danger, the automation trap. AI tools now take on ideation, screening, preference simulation and consumer interviews together, so the feedback signals organisations use to learn what works become increasingly AI-produced. A company that generates ideas with a model, screens them with a model and tests them on simulated customers has closed the loop. Nothing in it has touched a real buyer or original insights.

AI reduces the screening load. On what customers want, its output is a hypothesis until a customer confirms it. Desirability is still the most important category of questions that need to be addressed at the front end.

Decision-grade evidence

The remedy is decision-grade evidence: evidence sized to the decision it supports, linked to named sources, and explicit about what was not measured. It has two parts. The first is the set of questions each idea must answer. The second is the standard of evidence each funding decision demands.

Strategic Fit

Desirability, feasibility and viability are familiar, and AI can help to answer each. They are questions of evidence. Strategic fit is a question of choice. It connects to which future your company intends to compete for, and your leaders must answer it before any idea is scored. Future Back begins there: choose the future, then work back to the bets that must be placed first to start to learn and validate the underlying assumptions on which your strategy was built. It's the right way to test hypotheses, reduce risk and make progress when executing a strategy.

Fit also carries a risk. If it means resemblance to today’s business, the criterion builds a novelty penalty into the test itself. Define fit against the future you want to lead, and codify it before any scoring begins aligns strategy and innovation. An idea that passes the other three tests and fails on fit should not simply disappear. Record it with a review date, because the future you chose may change or the idea may be spun out through open innovation mechanisms.

Keep the gates in balance

Providing evidence has a cost, and that cost limits how many ideas a team can resolve. Compare two rates. Idea velocity is the number of ideas ready for evaluation that enter the funnel each period. Validation throughput is the number a team takes through the evidence standard to a decision in the same period. Their ratio is the load ratio. Above 1, the backlog grows every period.

Take a team that admits 40 ideas a month and resolves 10. The backlog grows by 30 a month and reaches 180 after six months. An idea admitted in month six then waits about 18 months for a decision.

Output is set by the slowest step, and speeding up any other step only adds work in progress. AI accelerates generation, the constraint or bottleneck is at the gates when tough decisions need to be made.

The remedy needs to be systematic. Use AI at the gates but provide enough oversight to improve the outcome, i.e. codify the criteria and evidence standards up front, train the system and build confidence over time to increase throughput.