Problem-Finding AI Agents and Exponential Serendipity

2025 will be the year AI agents move into production in many places, mainly to tackle initially small but increasingly meaningful problem-solving tasks. However, an essential part of their work happens upstream from problem-solving. That is, problem-finding. Problem-finding involves identifying the why and the what before working on the how. In many cases, this is at least as important as solution-finding, and quite often, it’s what senior leaders must focus on the most.

Interestingly, it is also where most people feel that machines, including AI, have the least capabilities - a feeling that Pablo Picasso encapsulated well when he said, "Computers are useless; they can only give answers." But - should that be true today?

Pablo Picasso said, "Computers are useless; they can only give answers." Is that true today?

Beyond the concept and intuition, I want to propose an architecture for doing this at scale—one that can be industrialized and yield exponential results using technology we already have and, indeed, the technology we’ll have in 2025.

Ideas—including those related to problems—have a structure that machines can interact with

The core notion here is that ideas have structure (morphology). They can be divided into different parts and connected. Humans sense those components intuitively in their minds. Technology can do that via knowledge graphs or other representations that AI could generate if enabled and asked to. Then, when exploring an idea—its why, the what, and the how—we can ask AI to analyze it deliberately, including the structure of ideas that relate to problems, not solutions.

Even if AI machines handle some of it in their answer (at inference time), this additional reasoning step and the tokens involved can yield tremendous value in finding more accurate and conceptually non-trivial solutions.

Today, these exercises are carried out in workshops where participants give each other ideas or systematically seek them through some form of “information and knowledge feeder,” a curation engine (the like of Perplexity, of Gemini DeepResearch.) But all of this is slow, manual, and suffers from the limitations of human's "field of vision" which makes the serendipitous discovery of new ideas clunky and inconsistent.

What can be done today? Likely, if what we seek is novelty, "extrapolation", we can't just ask AI point blank. So let's take a step back and reflect on the appropriate thought process through something that oddly resembles chemistry and physics.

Forcing idea collisions

In a previous essay called AI Can Ideate Harder, we saw that we can use AI-enabled ideation processes to force ideas and their components to collide with others deliberately. In that process, we also use lenses—conceptual frameworks based on solid human reasoning encapsulated in a framework—and smash ideas against them. We do that both to find solutions and problems by exploring and exposing different dimensions of the problem to discover new angles of attack.

We can use AI-enabled ideation processes to force ideas and their components to collide with others deliberately. We do that both to find solutions and problems.

A great way to do this traditionally is to seek out interesting people and their ideas and then compare ours with theirs. By the end of that conversation, we better understand the problem at hand (the Why) and the potential categories of possible solutions (the What). This then informs the downstream problem-solving work.

We do this routinely in our offices, in meetings, at water coolers (physical and possibly virtual), at conferences, on social media, and so on. In a way, our societies thrive because there is this "perpetual motion machine" of idea collisions across our networks, and those ideas get harvested by organizations—companies, academia, and even entire markets.

Enter Problem-Seeking AI Agents

What’s exciting is how AI can exponentially amplify these processes in non-trivial ways. Given enough resources and guidelines (including ethical, bias, and intellectual-property), AI can relentlessly canvas people's “halo”—the corpus of knowledge surrounding them. Machines can identify interesting individuals, examine their ideas (in this case, their challenges), and "smash" ours against theirs. That works very well for problem-finding too. As a result, AI can serve us knowledge that is a net-new, relevant addition to what we already know instead of forcing us to wade through many things before we find something really accretive.

Machines can identify interesting individuals, examine their challenges, and "smash" our capabilities against theirs to find meaningful recombination. That works very well for problem-finding too.

In problem-finding, an AI agent could locate people whose ideas benefit from our broader capabilities and do a first round of iteration to discover intersections, either existing or potential (that is, requiring additional work), before handing them over to humans and their other AI tools (such as scaffoldings and exoskeletons.)

A first simple application that could make each of us a bit of a superhuman problem-finder: an AI agent sifts through all relevant newsletters, identifying the key trends that others are trying to solve and giving you not just a deduplicated digest of the zeitgeist but a "net-new" one that has filtered out anything that you know already, hence increasing the output usefulness and reducing your cognitive effort.

Now open the aperture: imagine you run an innovation ecosystem and know (your AI's corpus knows) the talent and technologies it typically comprises. You can now send AI agents into the halos, the knowledge spaces of potential target companies and those who lead them. The AI can find combinations between what those companies seek, their unresolved needs, and what my ecosystem can potentially offer with additional innovation efforts because the "why" and the "what" are mapped thoroughly.

You can extend the examples to many other spaces:

1. In business-to-consumer, your problem-finding AI can continuously canvas the synthetic personas of target customers, especially as they dynamically evolve because of market trends. In the media industry, an AI could scan online streaming habits, fan forums, and influencer trends to uncover unserved audience interests. It might detect, for instance, a rising fascination with eco-thriller narratives, suggesting a new genre mashup that existing studios or content creators haven’t explored yet. By reviewing consumer transaction data, credit bureau updates, and macroeconomic indicators, an AI could surface potential underserved financial segments—like gig economy workers lacking stable cash flow options. By analyzing electronic health records, research publications, and local environmental factors (e.g., pollution levels and dietary trends), an AI agent can detect emerging disease clusters or unaddressed care gaps.
2. In business-to-business, your problem-finding AI interacts with the body of knowledge accumulated around target clients, both the organizations (for instance, through their product/service portfolios and their earnings calls) and the buyers themselves (looking at their public statements)
3. Your problem-finding AI can also scour machine-generated data, such as those from IoT sensors (think weather data), and combine it with others. For instance, insurers detect weather and home-improvement trends to address climate-related risks in the residential market. Problem-finding AI can also sift through global shipping data, real-time traffic feeds, and weather forecasts to identify new types of bottlenecks well before they happen.
4. New scientific and patent data can also be engaged with, whose "why" reveals the emergence of new partial solutions that hint at new problems. For instance, decades ago, Corning's Gorilla Glass could have hinted at new user interface designs. Or, today's lightweight edge AI hints at new, solvable problems, from industrial logistics to environmental monitoring. An AI agent can survey scientific studies, sustainability metrics, and community impact reports and discover that specific recycling initiatives are failing due to poorly chosen plastic types.
5. And, of course, the sky is the limit. Imagine how this can support strategy (also including M&A) and finance teams, as well as their CEOs. But they could also enable public bodies to identify upcoming challenges and inform their scenarios.

The chart below illustrates a possible architecture of these systems.

The result is a better understanding of which problems are worth solving—problems that are both desirable and whose solution is increasingly feasible.

Agent, meet my agent—and talk through my stuff

To recap, a problem-finding agent is an AI-driven system that [1] scours relevant data sets (knowledge halos), [2] identifies new or unmet needs, [3] cross-references areas that are attackable by the categories of solutions that exist (even if the exact solutions don't), and [4] presents them to humans for deeper evaluation.

Conventionally, this kind of synergy happens through people who meet each other—either systematically or serendipitously. But now, machines can go and “meet” with the knowledge base of those people. In a not-so-distant future, these machines could even meet with other people’s AI agents, allowing for an initial and thorough scan of the potential for intersection and identifying additional problems to be solved. Once they find enough overlap—enough interesting problems that should not only be solved but can be solved—then humans can meet, possibly assisted by machines with a thorough context.

There has been much discussion of agents independently engaging other agents. This is one of the more straightforward use cases I can imagine, as it benefits from scale and is low-risk.

This is how companies can start:

1. in concert with your technology team, business teams should identify the AI agent capabilities you can deploy in the next 3 months. Don't undertake anything that can't be done in a short sprint
2. design AI-finding agents human-centered, by involving users early in determining what would be truly desirable to them (use cases, user experience, human-in-the-loop feedback) and what data (knowledge bases across silos) you can access, and avoiding unnecessary over-specs
3. start small with a proof of concept, but don't try anything for which you don't have a path to scalability

Individual and organizational resilience hinges on detecting inflections early in an increasingly fast-moving world. AI problem-finding agents can help there. In the process, we might heed Picasso's concern and use computers to do more than just give answers.

If all of this sounds like the inception of an Iain M. Banks novel, it may very well be. And in 2025, the pieces are in place for it to be a reality.

This article is part of a series on AI-augmented Collective Intelligence and the organizational, process, and skill infrastructure design that delivers the best performance for today's organizations. More here and in the white paper here.

Get in touch if you want these capabilities to augment your organization. Build learning, problem-solving, innovative, intelligent organizations. Build Superminds.