Your Problem-Solving Idea Flow, AI-Augmented

How do we use AI to augment human capabilities to generate better ideas and solve problems? There's immense potential in doing this well, both personally and organizationally.

I have provided updates in the past about recent research that could be of practical, immediate use to professionals: the general vision of augmented human capabilities and how we need to lead machines to ask us questions; how we can get AI to ideate harder; the work that we have done at MIT to provide a scaffold, an exoskeleton to humans in the ideation process; the competencies that human professionals need to acquire and some relevant skills to wield the power of AI better. These contributions are worth reading as they complement what we will do in this essay.?

Here, I share practical guidance that your teams (and their machines) can execute to generate, day after day, better ideas, whether it's problem-solving (known-knowns) or creativity (all the way to unknown-unknowns). I will detail some steps to incorporate AI into current work practices, leading their people to use these steps as a standalone method or as part of their existing techniques. As you explore these concepts, remember that while I mainly refer to text-based problem-solving and creativity tasks, AI can increasingly be used for visual (and possibly spatial) and even auditory tasks.?

A few reminders first:

1. The process is not one of querying machines; instead, it is about augmented collective intelligence. There’s a real risk that humans might let poorly managed or incompetent machines take over, leading to mediocre or worse results. This is, unfortunately, a common issue in many organizations, and many of the disappointed accounts we hear in the media can be traced back to this problem. Similarly, there’s a misconception that “AI knows everything at any one time, so there’s no reason to tease its thinking through a process. As the evolution of prompting techniques demonstrates, there is great value in leading machines’ thought sequences. So, put your pilot gear on and take control.?
2. Use individuals, groups and machines in deliberate sequences. The general instinct is to have individuals query individual machines. I recommend considering different configurations, such as: (a) humans think first individually, then in group(s), and then add the AI's perspective to theirs; (b) the perspective of humans is injected into AI for the machine to build on it (c) the perspective of machines is injected into human groups for people to work on them; (d) all of the above can be done with multiple AIs instead of one, for instance, agents that have specific roles (planner, curator, critiquer, etc.) and capabilities (broader generalist or more specialized models; multimodal or not; etc.).
3. A problem-solving or ideation process is not linear. The gestation of ideas, especially great ideas, often involves a non-linear, meandering journey with recursion and back-and-forth. This is a normal part of the creative process and should be embraced.?

We need frameworks for multi-step AI augmentation. Some literature on this is emerging (like our MIT and some recent Harvard work), showing that well-designed human scaffoldings improve the quality of solutions. These early experiments hint at the possibility of multi-step flows and their potential. Here, we dive deep into what those steps can be.

The overall structure of the flow (below) will be familiar to many, especially those familiar with sophisticated problem-solving or ideation processes, such as design thinking, especially in its double-diamond form. This structure will guide us through generating better ideas and problem-solving with the help of AI.?

First Step. Exploring the Why: AI Helps Falling in Love with the Problem

The first chunk of the process consists of exploring the problem space. You might have heard the phrase, “To generate better solutions, you must fall in love with the problem before falling in love with the solution”. You might also have heard people saying that it is essential to define the Why of the problem first and the What (category of problems it belongs to). Not spending time understanding why a problem matters in detail or segmenting the problem space improperly typically leads to poor results. The good news is that AI can help us do that; the bad news is that humans don't automatically do that with AI, yet. The additional value of exploring the problem well is that, apart from helping humans, it helps the machine spend computing resources accumulating and focusing on relevant concepts that it will later "keep in mind" for the solution-finding phases. Locking in on relevant context is essential for machines to do their job well.

Below, I'll walk you through the steps, using a specific example of a complex problem (reskilling for AI) to show more clearly the applicability of these concepts. I will first provide the schematic representation for each phase and then describe it in detail.?

I recommend the following steps when it comes to discovering the Why:

1. Confirm problem's description. Enlist the machine's help to confirm the problem's description, hence trying to avoid possible "loss in translation." Language is not a very precise tool at times, but language models, if guided appropriately, can help us refine it. For instance, if your problem is one of “finding creative ways of retraining executives so that they harness the power of AI,” you should lead GenAI to ask you and your team additional questions about what you mean by that - which could result, for example, in spelling out what kind of company you're working, what kind of people work there, which people are most at risk, what kind of resources you have already applied to this problem, etc. This is the right time to add any insightful perspective on the problem that you already have prepared - don’t expect that AI “knows everything”: it probably does know much of what you would know, but for cost efficiency reasons, it is unlikely to prioritize all of the relevant information during inference time (when you query it).?One interesting twist is that the recently released OpenAI o1 models, because of their ability to think logically, could significantly improve the already impressive capabilities the machines have in this space.
2. Personas' perspectives. Then, if the problem lends itself to it, ask the machine to map the stakeholders and take the perspective of different personas involved in or experiencing the problem we are trying to solve. In our example, that might mean different seniority levels of workers (for example, CxOs or entry-level employees), in various functions (for instance, in the finance or the sales function), in different parts of the world, and needing different things (for example, a general overview as opposed to practical knowledge required for upcoming business meetings). Use AI to describe the journeys of each of those personas as they traverse the situations that cause problems, and ask AI to journal their emotions (e.g., an employee trying to find time to learn during a busy week and then struggling with identifying the most relevant and appropriately-sized piece of content for that specific day).?Or, if you have already obtained traditional client or customer preference data, include it as context.
3. Decompose. Third, apply some decomposition of the problem, for example, breaking the problem down into parts and types, identifying the precursors and the consequences of that problem, as well as moving up and down the abstraction ladder to identify levels of specificity or generality that could yield a better vantage point on the problem. For instance, here, learning new skills for AI means a set of specific learning steps but also includes upstream problems like the curation of contextual practical knowledge that can help adult learners understand the problem in the context of their work and downstream issues like the inability to teach subordinates and colleagues what one has learned. Abstractions are essential when solving seemingly intractable problems - since, in this case, AI could help us understand the broader set of problems related to ours, for instance, general knowledge management issues or knowledge transfer challenges present in many new-technology innovation periods.?Again, new, more logically-reasoning AI models like OpenAI o1 can help here.
4. Inject doubt. Enlist the AI tool to poke holes in our description of the problem. Generative AI is often quite good at critiquing thought processes, especially in textual form, and could help identify gaps or white spaces that could be meaningfully explored before moving on to the next steps. You can also ask AI to search for knowledge on important dimensions of the problem to bring some more insightful angles of attack to the fore.?
5. Summarize and filter. After this, it is essential to take stock of what has been covered and cluster and further structure ideas to streamline your output, consolidate your understanding, and weed out duplications. This is also when you - individually or as a team - filter perspectives that don't seem promising (insightful, practical). Identifying what’s particularly insightful helps prioritize aspects to be kept front and center of the AI memory - which, while being increasingly vast, benefits from being shown what good insight is. AI is increasingly good at helping filtering, especially if you clearly define what is important for you (e.g., only surprising insights or MECE - mutually exclusively and collectively exhaustive ideas, etc.). This is also an area where new, strong logic models can help.

Second Step. Exploring the What: Guiding AI to Interesting Solution Spaces

The second part relates to discovering the What, that is, understanding what category of problems and solutions are relevant. I like to do a thorough job at this often-skipped phase, and people move directly to generate ideas. From experience, subtle yet powerful insights can be generated by looking at the problem through a creative, unusual, categorical lens.?As an example, there is an old innovation saying - "if you are segmenting your market the wrong way (typically, the traditional way", you will never be able to truly innovate" - which illustrate the value of taking the right type of perspective. There are many ways of doing that, and AI is very useful here.

1. Analogize. Start by finding insightful analogies. AI has become better at this, especially when given enough context. In our example of up/reskilling for AI, analogies are: individual plants in a forest that become aware of and react to the threat of a new parasite; or community workers in rural villages of developing economies needing to learn how to administer vaccines. Good analogies often shed light on poorly understood parts of the problem (for instance, the behavior of decentralized organizational units in our example). They can yield interesting avenues for solving those problems cost-effectively (for instance, in our example, by hinting at the energy communities of employees can muster).
2. Use lenses. Then, apply other lenses that force your AI to think of the problem in new ways. At the MIT Center for Collective Intelligence, we developed some of them, dealing with the design of organizations, as part of the so-called “supermind design methodology”. For instance, what kind of organizational structure (“group type”) is currently trying to solve the problem? (In our example, most learning is managed hierarchically, with some opportunities for communities of interest to form.) Another lens is the supermind’s cognitive process, which examines what the collective brain of the current organizational design is trying to do. (In our example, the whole organization tries to learn by creating capacity in the appropriate parts of the collective brain, while specific units are “sensing” the environment to identify the correct use cases people should learn about.) Similarly, understanding what technologies are being used and their advantages and shortcomings is helpful for the subsequent solution-seeking phase. Finally, I can’t emphasize enough the value of encouraging AI to apply to the problem existing, human-made relevant management or scientific (or other) frameworks, to nudge AI into emulating some symbolic/abstract thinking embedded into the vast literature that exists around those frameworks? - e.g., in our example, using Christensen’s Disruption Theory to unearth interesting dynamics relative to skill-based competition (for instance, AI “disrupting from below” the work that humans do, by taking first mundane and repetitive tasks, with human incumbents doubling down on their efforts to perfect their current capabilities instead of migrating to a different value proposition while they have enough resources).?
3. Inject doubt. As in the previous phase, leverage AI to poke holes and identify gaps in this phase’s analysis.?"What are we missing?" is an excellent question for your AI and yourself.
4. Work with constraints. Truly creative thoughts may emerge by forcing the thought process to navigate hard constraints. For instance, in our example, you could use AI to detail extreme scenarios, such as those where learning resources are extraordinarily scarce or abundant, where the learners are highly sophisticated or extremely junior, or where all learning opportunities happen asynchronously and remotely, as opposed to others where all learning happens in person with real instructors. Constraints also come in the form of customer preferences (say, durability vs. performance), industrial constraints (e.g., materials availability, production location), or budget boundaries?(e.g., limited or conditional budget).
5. Summarize and filter. After this, as in the previous step, AI will be leveraged to summarize and cluster ideas and filter out less-interesting thoughts with your and your team’s active participation.?

Third Step. Exploring the How: Co-creating with Machines

Now that you've "fallen in love with the problem," it is a good time to venture into solution-finding. This last phase relates to activities that help generate solutions using AI as a thought partner. That means avoiding using it as an “oracle” and instead ensuring that you and your team are active participants in iterating ideas as they form. Those who resist the urge to move too fast into this phase typically deliver more novel and effective ideas.

1. You can start by asking AI to re-read the individual Why and What threads as context and deliberately identify possible solutions. A few good ideas might emerge even at this stage if the problem exploration was sufficiently insightful, for instance, through truly eye-opening analogies. After producing the first ideas, do not hesitate to point AI’s attention to the output of specific previous steps (e.g., “What does the analogy XYZ make you think?”). Ensure you don't anchor the following steps on these first results to avoid missing genuinely creative ideas. ?
2. As in the previous phase, one should force AI to generate solutions through lenses derived from frameworks and management (or other) theories, such as the following.?

• Groups: The supermind design’s “group type” described above is one (for instance, highlighting and facilitating the teaching role of practice leaders as part of communities of interest).?
• ACI pillars. The Augmented Collective Intelligence Pillars (more at Supermind. design) is another, leading AI to identify ideas that solve for specific high leverage points in collective-intelligence systems design (in our example of upskilling at scale, by for instance, enabling an AI-assisted, decentralized curation-at-scale required to surface relevant, contextual learning examples; or identifying the most important network nodes, humans and machines, in a network and enabling them accordingly; or creating the right incentives for a community; or providing a proper collaboration infrastructure).?
• Tech use. One can also ask AI to provide ideas about using specific technologies (say, augmented reality or edge computing) and apply them to parts of the Why and What components.?This is the area that most people intuitively think about first. Still, technology adoption is most effective when tied to organizational design components, which means using it on the right other "how" levers.
• Fact bases. If a database of solutions examples is available, with highly curated and relevant data sets, it could also offer valuable stimuli for AI (for instance, a Tech Crunch database of all education startups or a database of collective-intelligence organizations such as the one available on the Supermind. design website).
• Recombine. Most ideas are of limited interest, but their recombination could yield surprising value. AI can help recombine ideas at scale. They can converge into alternative solutions described as one would typically do with so-called “concept posters.” Models that are better at reasoning, like OpenAI o1, can complement the creativity of older ones by combining and deduplicating ideas, etc.
• Constrain. Once again, have AI develop ideas under specific constraints, such as available resources and capabilities or specific customer requirements and preferences.
• Filter ideas once more. You want to encourage the AI to over-emphasize specific types of ideas. For example, privilege the most creative or surprising or use the most understandable and potentially useful. Keep in mind that many models will try to satisfy you by offering ideas that do not depart too much from normalcy, which means that you and your team must be prepared to forcefully prod AI to go into uncharted territory (or you might need to fiddle with system parameters such as model output temperature).

Once you have filtered, with the help of AI, the numerous inputs from this phase, you are ready to refine them. Several activities could get you there with the help of AI tools. I want to underscore how valuable AI can be in this phase: humans are often not great at candidly critiquing ideas, as they’re bounded by what they know and want to respect civility conventions - but AI can find holes into seemingly good ideas, if it is asked to look into logical cracks, play contrarian or antagonize, among others. As many innovation experts say, “Ideas are cheap, but finding flaws in them is very hard and valuable.”?

1. First, once again, use AI to identify blind spots. Has your process inadvertently missed out on a specific range of potential solutions? Are there implicit biases? For instance, in our example, has AI suggested solutions that are out of reach for older employees??
2. Then, analyze how each idea or component of ideas could fail to meet its objective. AI can support a thorough exploration of the failure modes. For instance, in our example, AI could highlight managers’ reluctance to give their teams guidance and time, limiting the uptake of any reskilling program.
3. You can also take the perspective of different personas and critique each idea from their viewpoint. For example, what would the CFO think of the ideas? What would a junior employee in South America think of it (e.g., “Are you providing self-improvement coaching in Spanish?”), as opposed to a senior leader in Japan (e.g., “How do you provide constructive educational feedback in group forums while showing respect and deference to authority?”)
4. Finally, use some constraints again, but in this case, their value is to weed out potential solutions, for instance, those that require too many resources or have too long of a deployment cycle.
5. The last filtering exercise can be clustering and structuring the remaining ideas and asking AI to assess their feasibility, desirability, and economic (or other) viability. AI can do that in a range of reasonably well-understood domains and could help even in more specialized domains by using first-principle thinking, especially if given access to a range of successful or unsuccessful examples from existing databases.?

Enlisting the Full Power of Humans and AI to Push Hard

The process that we just described can be particularly successful when leveraging a few opportunities:

1. Human in the loop. Tools can create many ideas quickly, but designing a human-centered experience and training your users to remain firmly in the loop is essential. That means providing continuous and intellectually aggressive feedback to the machine to avoid drifting towards platitudes. You can also ask part of your team to do the same exercise as AI does, in parallel, and then converge the results. Or you do the exercises first, with AI asking you probing questions.?
2. Using different agents. It is not hard to build AI agents with specific capabilities that you can use for specific steps. For instance, an AI agent could be configured to be a particularly insightful “gap finder” by giving specific examples of how to do so. Other agents could be beneficial in understanding how to apply management framework, including finding the relevant ones. Some agents could be configured to be good raters of possible solutions. AI agents could credibly take the perspective of individual personas.?
3. Asynchronous batch processing for “fractalization.” Conversely, as AI is increasingly cost-effective, you can build workflows that potentially follow many “rabbit trails,” branch out from those, and provide intermediate and final results at specified intervals. For example, some of these machines could use brute force to combine many ideas, e.g., trying to combine every group structure and technology type. This only works if your filtering mechanisms are reliable, but that should be the case in specific domains, especially over time.
4. Recursive loops. As I mentioned at the very beginning of this essay, do not mistake this process as a step-by-step, one-directional checklist. Solid problem-solving and creativity require the ability to trace back multiple steps, start again, deemphasize specific trajectories, and try initially neglected ones. And that is, for now, firmly the work of a human.

Below is an overview of the entire process we discussed. Remember that this picture is not exhaustive: other problem-solving and creativity techniques, particularly those that you and your teams are used to, could be inserted as part of the overall process. That's especially powerful if they’re framed correctly and accompanied with sufficient context so that AI understands them and if you see them as a two-way street facilitating a dialectic between humans and AI, not a "download" from the machine onto us.?

This won't stop. Get Ready for the Change

We will continue to discover how to make our tools more effective at interpreting each step. It is also possible that we will find different pathways for creativity - the idea of fractal exploration mentioned before, coupled with the usage of pre-existing mental frameworks, especially when harnessing AI to do filtering at scale, hints at possibilities that were inconceivable when using only humans as part of the creation process. Some form of truly "alien" intelligence could help us in ways we don't fully imagine today.

Time will tell, and we will discover pitfalls above and beyond the potential threats we intuit right now (such as an exponential increase of mediocre output, the increasingly ineffective use of truly bright, creative human minds, etc.).

For now, however, it seems sensible to continue the discovery of AI-assisted problem-solving and creativity, especially by exploring the synergy between humans' capabilities, individually and in groups, and the capabilities of one or many machines.

This essay 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. There is more here.

Contact me if you want these capabilities to enhance your organization's learning, problem-solving, and innovation abilities.