Putting innovation discovery onto AI rails

The test was not a survey. People paid with their own money in a live trial. And the same sweep that generated the concepts threw off a pipeline of real companies for the client's investing and M&A teams.

The client

The non-core growth unit of a global consumer group, chartered to find and validate new businesses outside the core. The question on the table: could discovery and conviction, its slowest and most expensive stages, run materially faster without losing rigour?

What I built

I moved the entire discovery front end onto AI rails. Clustering reads thousands of market signals and groups them into themes. Generation produces candidate ideas inside each theme on demand. Filtering ranks the field. Conviction prep turns the survivors into testable propositions with the evidence requirements attached.

The method was built deliberately model-agnostic. The same workflow ran across five platforms: Grok, Gemini, Perplexity, ChatGPT and Claude, with outputs reconciled against each other. Disagreement between models usually marked the questions worth human time. Scoring blended a machine-generated read of the market data with human judgment, weighted so that neither side could overrule the other on instinct.

The live run

I ran it for real, in a high-frequency consumer category. The pipeline opened with a ranked, sourced map of opportunity spaces. Then the leading concepts went under real-world pressure: a multi-hundred-respondent study across quantitative and qualitative formats, a live tasting where people paid their own money and drank rather than answered a hypothetical, and more than twenty expert and stakeholder interviews on top of a full read of the category value chain.

The by-product

Reading thousands of market signals to generate new-business ideas does something else: it surfaces real companies. The sweep threw off a substantial pipeline of businesses worth evaluating as investments or acquisitions rather than as things to build. That opened a separate workstream with the client's investing and M&A teams: one route from discovery to a buy, invest, or build decision.

What it proves

The discovery front end now runs as one AI-native pipeline across five models, and it improves as the models improve. The same sweep that ranked the concepts also threw off a pipeline of real companies for the investing and M&A teams. The tasting was paid, not surveyed.

One of a set of engagement and systems case studies. If a problem like this is on your desk, start with a conversation.