Superhuman Knowledge Workers? AI Exoskeletons and Scaffoldings

To control, harness, and compete with AI (and AI-powered organizations), we need to make people better - individually and in their organizations, through their work processes. Even with today’s technology and practices, there are signs that AI can help do that with knowledge work, as it serves as an augmentation technology when used well.

Some call the result of this evolution cyborgs, reflecting a tightly integrated approach to human-AI collaboration, with continuous, synchronous exchanges at the subtask level. That is opposed to centaurs, i.e., a model of human-AI collaboration where there is a clear division of labor between human and AI capabilities, and often, one set of tasks gets fully automated, and collaboration often happens somewhat asynchronously. Both are useful and will happen. This article focuses on the former: what they are, why they matter, and how to build them.

Early promise

Recent empirical studies underscore how Generative AI-based tools have the potential to enhance productivity and work quality across professional domains, but not by default.

Early studies show that AI tools like GitHub Copilot and GPT-4 can significantly boost productivity across fields, enabling faster task completion, higher quality outputs, and increased creative performance. Developers, customer support agents, business professionals, and consultants reported notable efficiency and quality improvements, handling more tasks in less time while enhancing overall output standards. Generative AI can also make individual contributors, especially those with lower skills, less dependent on others (and more efficient) and able to explore with more tools.

Yet, augmentation still has risks, and collaboration between humans and machines doesn’t automatically result in better outcomes than each would achieve individually, as an extensive recent analysis showed.

How do we design and build tools to get human augmentation right? To answer that, let's first clarify what kinds of supporting tools exist.

What are AI exoskeletons and scaffolding for people?

First, let’s clarify what scaffolding and exoskeletons in AI mean.

Scaffolding is primarily designed as a temporary support, focusing on cognitive assistance and learning processes. It is designed to encourage a more reflective and iterative approach to problem-solving. It adapts dynamically to user progress, gradually reducing support as proficiency increases, and aims to foster independence. Scaffolding is commonly applied in educational technology, software development assistance, and guided data analysis, targeting specific tasks or learning objectives. For example, an AI that shortens the initial learning curve on any topic or asks questions to develop specific critical skills. 

In contrast, AI exoskeletons are more persistent enhancement systems. They aim for continuous augmentation of both cognitive and analytical capabilities across various domains without necessarily intending to be removed over time. Exoskeletons seek a symbiotic integration with human cognition and decision-making processes, providing ongoing support in real-time information processing, language tasks, complex modeling, etc. While scaffolding works towards user autonomy, exoskeletons maintain constant augmentation, becoming integral to the user's enhanced capabilities—for example, an AI "whisperer" or a problem-solving GPT.

They represent distinct approaches to augmenting human capabilities but sit on a continuum. Some scaffolding can also be used as ongoing support for execution, and some exoskeletons can help people learn.

How to build them right

The first clear realization is that we must “lean into the 70”, that is, 70% of the AI transformation work that isn’t about the technical, core-IT side - process and people. AI can help make them better. And most organizations don’t allocate enough resources to do that. While much of the headlines these days are about autonomous AI agents, the lower-hanging fruit is the design of the interaction of people, individually and collectively, in workflows. 

This means focusing on people's capabilities, including what I call Augmented Intelligence, which is the ability to work with machines seamlessly and act as a "System 2" (slow, abstract) to AI's "System 1" (fast, recognizing data patterns). In this article, however, I want to talk about the design of the workflows that help people achieve more, and better, especially but not only the problem-solving and innovation process.

To start with, building scaffoldings and exoskeletons right requires understanding how to guide AI along an appropriate cognitive process (more on this here and here).

It also means focusing on which AI capabilities are already good enough to improve the human ones. The framework below shows the types of activities people and groups perform (called "supermind cognitive processes"), hints at which GenAI is good at, and helps understand the potential scope of intervention.

Some experimental evidence from prototypes

Our research at MIT's Center for Collective Intelligence started in 2020 and has since shown the value of designing scaffoldings and exoskeletons well. Our latest study, for instance, compared the effectiveness of MIT Supermind Ideator, an AI-driven idea-generation and problem-solving tool we built, to ChatGPT and solo human effort. In addition to an intentional process reflected in specific prompts, the Ideator partially operates on a fine-tuned large language model (OpenAI’s latest) trained on case studies of innovative organizational practices. This makes it more context-aware and responsive to domain-specific challenges than general-purpose tools.

We found that individuals using it generated significantly more innovative ideas and engaged in deeper interactions than those using ChatGPT or working independently. The findings underline the advantage of specialized AI interfaces for improving human-AI collaboration in targeted, complex problem-solving contexts. By offering structured steps and prompts, Ideator is a scaffolding tool where people learn how to perform specific tasks optimally—and might even be an exoskeleton, usable constantly—that helps users overcome cognitive biases and explore alternative perspectives.

Interestingly, the data shows that AI tools like Ideator might be especially beneficial for less naturally creative people, at least initially. The study also emphasizes that, despite AI’s support, human judgment remains essential for curating and refining AI-generated ideas to reach optimal solutions. It also suggests the need for designing the appropriate turn-taking and sequencing of human-machine interventions to avoid drifting into platitudes and risking dumbing down people over time.

Turning it into real-world impact

Combining domain expertise (business, innovation practices) with easy-to-build and easy-to-iterate AI tools has great potential. Low- and no-code AI technologies (like OpenAI custom GPTs, Langchain, or Wordware, to mention some) will eventually do what Excel did to all of us thirty years ago: allow people to translate business logic into repeatable actions, hence forever changing how work is done. 

This is especially important as the "art of the possible" (the "jagged frontier") keeps moving, because continuous experimentation from laypeople, not just technologists, will drive identification of use cases and iteration toward viable product-market fit.

As an example, inspired by the above ideas, it was easy to encapsulate business processes and practices and curate knowledge bases (e.g., an extensive database of organizational design and process examples) into a few tools that help with problem exploration, solution discovery, and solution refinement. Some examples are below, all built on the theOpenAI GPT Store and freely accessible (find them through the store's search menu).

• AIdea Collider - thoroughly explores business and organizational problems, then systematically combines diverse solutions for superior ideation.
• ACI Designer - generates uncommon ideas for complex problems by leveraging ACI (Augmented Collective Intelligence) principles to support the design of collective intelligence (supermind) architectures.
• Idea Hardener - strengthens and refines ideas and concepts by using AI to harness dozens of critique frameworks.
• Apta - identifies parts of an organizational process that Predictive and Generative AI can augment or automate.
• ACI UnBook - guides the reader to discover the most relevant concepts from a set of long-form content (a book) and other fact-bases

To an extent, these can be used as scaffoldings and exoskeletons, depending on users' preferences. Less-experienced people who rarely do those tasks can use the tools to perform at a "better-than-beginner" level. Those who want to permanently flex that muscle more can use the tools to learn faster. And skilled people use the tools to complement their work.

Illustratively highlighted below, their architecture is straightforward. Non-technical subject matter experts can now design and even build some of them.

This hints at the possibilities, especially when focusing on specific personas (e.g., the CFO in a board meeting or a salesperson preparing a proposal) and adding specialized knowledge that organizations routinely accumulate. Think of use cases like: 

• Strategy alignment. Translating company strategy documents into "Chief Strategy Officer's digital twins" tools enables strategy socialization, cascading, and team alignment. It also allows leaders and, in general, all employees to constantly query the strategy without needing access to the strategy team.
• Update of staff knowledge. Learning anything, not just new skills, in a more agile way - for instance, helping technology implementation teams understand how new products are better than old ones or bringing people up to speed on and contributing to new use cases for GenAI.
• Avoiding blind spots. Knowledge "whisperers" will be curation agents that incessantly scan within and outside the organization to find relevant and non-duplicative knowledge (compared to their human users' knowledge base) and summarize it for the user daily or whenever they perceive the human is trying to decide something related.
• Board prep. Assisting CXOs in preparing for board meetings by buttressing their thought process and preparing for objections.

Many others exist, both horizontal across roles and even more excitingly vertical, role-specific and domain-driven ones.

Like Excel forty years ago, which is now used by anyone (and much of today's organizations, willing or not, run on Excel and its macros), these AI-powered "mini-apps" could generate significant productivity improvements and alleviate employees' toil. And they might do something else: through widespread, bottom-up experimentation, they might surface use cases that don't exist today. And, of course, we will increasingly use more sophisticated tooling, all the way to AI agents or even networks (graphs) of them - with a human in the (very tight) loop.

The upshot can be more-effective people, enjoying their work more. But today, in most organizations, there is an imagination, not just a technical gap. That gap can be closed. It is time to start designing and building AI-powered exoskeletons and scaffoldings.

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.

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