Where AI doesn't exist yet: the gaps MIT's research exposed

If Article 12 answered where AI is, this article answers where it isn't. The MIT mapping study showed that nearly all visible AI activity is piling into a tiny slice of work. The follow-up question for builders and investors is the one that matters more strategically: what sits in the rest of the map, and why has so little been built there?

The short answer is that most of the white space is not empty because it is irrelevant. It is empty because those activities are harder to understand, harder to productise, harder to sell, or harder to make people trust. That is exactly why the white space matters. In crowded categories, competition compresses quickly. In neglected categories, execution quality still creates real leverage.

The map is mostly empty

The defining asymmetry in the MIT study is not just that 92% of AI apps cluster in 6.8% of work activities. It is that the other 93.2% of activity space remains mostly untouched.

That flip in perspective matters. Most commentary reads the paper as a story about concentration. It is that, but it is also a story about absence. Once you map 39,603 work activities against the visible AI product landscape, the real strategic question becomes: what is in the near-empty part of the map?

The answer, according to Cai et al. (2026, arXiv:2603.20619), is not "nothing." The paper explicitly separates activities that are technically unfeasible with current AI from activities that are technically feasible but commercially unaddressed. That distinction is the heart of the opportunity.

It means the white space is not one uniform bucket. Some of it is still blocked by capability limits. Some of it is blocked by economics. Some of it is blocked by trust, workflow friction, regulation, or the simple fact that nobody has yet done the hard work of translating a raw model capability into a deployable product.

That distinction is what most AI market commentary gets wrong. It treats every missing category as proof that AI cannot go there. The MIT paper suggests a more useful interpretation: many gaps are not technical dead ends. They are unclaimed commercial ground.

For founders, that changes the frame entirely. The task is not to ask only where models are strong. It is to ask where model capability already exists, but product formation has not happened yet.

Three types of gap

The paper organizes unoccupied activity space into three types of opportunity. This is one of the most useful parts of the framework because it gives builders a practical way to sort gaps instead of treating them all as equally speculative.

That sorting discipline matters because white space can look deceptively similar from a distance. As Cai et al. (2026, arXiv:2603.20619) show, the important question is not simply whether an activity lacks products today. The more important question is why it lacks products: missing capability, weak economics, low social acceptance, or a market that has not yet translated feasible capability into an actual company.

Technical opportunities

Technical opportunities are the safest category in the paper's framework. These are activities where AI capability is already proven in adjacent tasks, but no one has built a product for that exact activity.

That is an important distinction. You are not waiting for a model breakthrough. You are not betting on an entirely new research frontier. The core capability already exists somewhere nearby. What is missing is application.

In practice, the gap is usually one of six things:

1. Workflow design
2. Data access
3. Domain packaging
4. Go-to-market clarity
5. Trust and compliance
6. Systems integration

This is where many outsiders misread AI strategy. They assume strong companies are built by inventing fundamentally new model capabilities. Sometimes that is true. Much more often, strong companies are built by taking an already-proven capability and wrapping it around an overlooked workflow that people actually need.

Take the pattern the spec points to. AI already writes legal prose. That does not automatically create a serious product for drafting a specific class of contract clauses in a narrow vertical with its own review norms, liability constraints, and approval process. AI already analyses medical images. That does not mean every radiology sub-task has a production-ready tool embedded into real workflows.

The gap is not "can the model do something vaguely related?" The gap is whether someone has built the end-to-end system that turns capability into dependable work.

That is why technical opportunities are attractive near-term bets. The uncertainty is narrower. The problem is usually not whether the core inference works at all. The problem is whether the product team can bridge the last mile between general capability and usable workflow.

Economic opportunities

Economic opportunities are more strategic. These are activities that are both technically feasible and economically attractive.

The paper describes this as the highest-value class of gap because it combines two properties that matter commercially: the work can likely be automated or augmented now, and there is real economic density behind doing so.

What makes an activity economically attractive in this framework is not a vague sense of importance. It is the combination of:

1. Large workforce size
2. Meaningful wage rates
3. High task repetition
4. Measurable output quality

That combination creates a real ROI case. If the work is expensive, repeated, and measurable, a buyer can justify paying to improve it.

This is where builders often make the wrong tradeoff. They chase the technically elegant use case instead of the economically painful one. But the most durable product categories are often not the ones with the flashiest demos. They are the ones where budget, pain, repetition, and measurable outcomes all line up.

The sectors named in the task spec point to that profile clearly: skilled trades coordination, insurance underwriting support, healthcare documentation, legal discovery, logistics exception handling, and field service operations. None of these are glamorous in the way image generation or consumer chat can be glamorous. But they are full of repeated work, expensive labor, operational delays, and accountability requirements.

That is exactly why they matter.

The catch is that these domains are rarely shallow. They require deep workflow knowledge, internal champions, system integration, and operational credibility. The risk is not lack of demand. The risk is shipping something too generic to survive contact with the real process.

Unrecognised opportunities

The third bucket is the most speculative and, in some cases, the most important: unrecognized opportunities.

These are activities where the possibility of AI has not been seriously explored at all. They often live in physical and social domains, and their absence from the product landscape is partly a failure of imagination.

This does not mean the gap is easy. It often means the opposite. There may be no strong analog, no obvious benchmark, and no accepted product category yet. But precisely because no established market narrative exists, the field can be much less crowded.

The examples in the spec show the range of this category well: grief counselling support tools, adaptive learning systems for vocational trades, AI-assisted negotiation in high-stakes settings, and coordination tools for field-based care workers.

What links these is not that they are all the same kind of product. It is that they sit in spaces where people often assume AI does not belong by default. The absence of products is therefore partly social and conceptual, not just technical.

That is why unrecognized opportunities are higher risk but potentially more transformational. If a team can prove that AI belongs in a category where the market had not yet taken the possibility seriously, it can open an entirely new wedge rather than fight for share in a mature one.

The physical gap in detail

The white space becomes much clearer when you isolate physical work.

The paper's robotics data shows 20.8 million deployed robots globally, which sounds enormous until you look at where they are actually concentrated. 76.7% of those robots are floor-cleaning units. Only about 23% handle genuinely complex physical tasks such as welding, assembly, surgical assistance, and material handling.

That distribution matters more than the headline deployment number. It tells you the Do category is not broadly automated in the way people often imagine. In the paper's framework, Do accounts for just 12% of total AI market value despite representing a huge share of real-world economic activity.

Why physical categories are still thinly served

Physical AI is hard for reasons that software builders often underestimate. The technical challenge is not only model quality. It is the full stack:

1. Sensors
2. Actuators
3. Messy environments
4. Safety certification
5. Long deployment cycles
6. Physical maintenance

The distance between a lab demo and a production system is therefore enormous. A pure software team can iterate daily. A physical AI team often has to think in hardware cycles, field testing, installation constraints, safety review, maintenance logistics, and system failure under real-world variation.

That slows iteration, raises capital requirements, and makes the go-to-market path harder.

Why that difficulty is also the moat

The same things that make physical AI difficult also make it defensible.

If a company gets a physical AI system working reliably in construction, agriculture, skilled manufacturing, elder care, field service, or mining, the barrier to fast imitation is much higher than in pure software categories. Competitors cannot just copy the prompt layer, rebrand an API wrapper, and buy ads.

They have to solve deployment. They have to understand the operational environment. They have to build confidence that the system works repeatedly in the real world.

That is why the physical gap matters strategically. It is underpenetrated not because it is small, but because it is hard. Hard categories often look unattractive early and defensible later.

The hardware-software integration problem

Most physical AI systems require both hardware and software to work together. That raises the bar in a way many general AI products do not face.

You are not only building intelligence. You are building the body, the sensing layer, the control logic, the integration path, and the operating procedure. That tends to slow product cycles compared with pure software, but it also means that teams with strong domain integration can create real separation.

For investors, this is one reason physical AI may be structurally underinvested relative to its eventual addressable market. It does not scale with the same apparent speed as software, but it may produce stronger moats once it works.

The social gap in detail

If the physical gap is about embodiment, the social gap is about legitimacy.

The paper places Interact activities at 48% of combined AI market value, making it the largest category in the combined software-plus-robotics view. Yet it is still structurally underserved.

That sounds contradictory until you look at what social work actually requires. In interaction-heavy settings, the system is not merely generating an answer. It is entering a relationship.

Why social categories are hard

Humans are much more tolerant of automated calculation than automated counsel. If a system calculates a route, sorts a queue, or summarizes a document, users mainly care about correctness and speed. If a system gives care guidance, coaching, conflict support, or negotiation assistance, users also care about legitimacy, sensitivity, accountability, and emotional fit.

That is why the acceptance threshold for "AI calculates" is completely different from the acceptance threshold for "AI counsels."

The model may be technically capable of producing useful language in these settings today. But technical capability is not the only barrier. Trust, liability, institutional policy, and social norms become binding constraints.

Care and support

The first large social gap is care and support: mental health triage, elder care companionship, and chronic disease coaching.

From a capability perspective, much of this is already adjacent to what current LLM systems can do well: structured conversation, follow-up guidance, summarisation, consistency, reminders, and adaptive explanation.

The hard part is not merely conversational quality. It is clinical integration, liability, escalation logic, and the question of how much human oversight is required before users and institutions will accept the system.

That makes care and support a classic white-space category: technically plausible, commercially meaningful, but hard to ship responsibly.

Education and coaching

The second social gap is education and coaching. There are already many tools here, but most are surface-level. The deeper opportunity is in adaptive tutoring beyond flashcard-style workflows, vocational skill coaching, and professional development feedback.

This domain has a large workforce implication, clear value creation, and repeated use patterns. But many products remain shallow because the hard part is not generating an answer to a question. It is creating a loop that understands skill progression, feedback timing, motivation, evaluation, and institutional context.

That makes education and coaching a category where technical feasibility is real, but serious product design remains early.

Workplace coordination

The third major gap is workplace coordination: team dynamics, conflict resolution support, feedback systems, and performance conversations.

This category has huge latent demand and almost no mature product category. Every company struggles with coordination quality. But very few organizations want to hand those moments to systems that feel generic, politically naive, or risky.

That means the winner here is unlikely to be the team with the flashiest general-purpose model demo. It is more likely to be the team that understands how trust, policy, escalation, and workflow all fit together inside real organizations.

The broader point is that social gaps may be among the largest addressable markets in the paper's framework, but they are not unlocked by capability alone. They require acceptance.

What this means for builders

For builders, the most important strategic message is simple: the most competitive part of AI is also the most crowded.

Writing, coding, customer support, and data analysis are valuable categories, but they are also the zones where competitive advantage compresses fastest. If you are operating in the same 6.8% of activity space as everyone else, you are likely competing on speed, packaging, distribution, and brand inside a market that already knows what it is.

The white space in the remaining 93.2% is not all equally accessible, which means you need a sorting framework. A useful one from the paper's logic is:

1. Technical feasibility
2. Economic attractiveness
3. Social acceptance
4. Your unfair advantage in the domain

That leads to three practical questions every builder should ask.

1. Is the core capability proven?

If the answer is yes, the gap may be a product gap or a distribution gap rather than a research gap. That is usually where technical opportunities live.

If the answer is no, you may still be right eventually, but you are taking on more fundamental risk than many founders realize.

2. Is there a buyer with budget and a measurable problem?

This is the difference between an economically attractive gap and a technically interesting one.

A lot of founders fall in love with automation possibilities that no organization will pay to improve. The existence of a feasible task does not guarantee a market. Someone has to care enough, financially and operationally, to buy the solution.

3. Will users accept AI in this context?

This is often the hardest question and the one technical teams underweight most.

If a workflow touches trust, status, care, conflict, or institutional responsibility, acceptance becomes a product variable in its own right. You cannot assume that because a system performs well in a benchmark it will be welcomed into the real activity.

The strongest teams in white space are therefore usually domain-first, not model-first. They understand the workflow, incentives, failure modes, and trust boundaries before they choose the exact AI stack.

That is also why being second in a white space can be better than being fifth in a saturated category. The white space gives you room to learn where others are not yet looking.

What this means for investors

The investment implication follows directly from the market map. Capital today is heavily concentrated in the same activity clusters the paper already identifies as saturated.

That is one reason valuation compression in AI writing, coding, and analysis tools is likely to continue. These categories are important, but they are also becoming more commodity-like as model access broadens and differentiation narrows.

The paper offers a better lens for durable investment theses. Look for gaps with:

1. A proven technical path
2. Economic density
3. Low current competition

That combination matters more than broad enthusiasm around "AI" as a label.

Physical AI companies deserve special attention in this framework because robotics, sensing, and actuation remain structurally underdeveloped relative to the size of the opportunity. These businesses may be harder to build, but that is often exactly what makes them durable if they work.

Social and Interact-focused AI is a different kind of bet: high-risk, potentially high-reward. The companies that crack adoption and trust in care, education, and coordination may be disproportionately large because so few serious product categories exist there today.

The core investor lesson is that the most obvious categories may be the least strategically interesting from here. The next durable winners are more likely to emerge where workflow difficulty and trust barriers have kept competition thin.

The honest caveat

Not every gap is a good opportunity.

Some activities are empty because the economics do not work. Some are empty because current systems are still not good enough. Some are empty because social acceptance is years away even if capability exists. The paper identifies feasibility gaps, not guaranteed markets.

That means builders still have to validate demand the hard way. A technically feasible automation can still fail if organizations do not care enough to change the workflow, cannot measure value cleanly, or do not trust the system enough to adopt it.

Speed also matters. As foundation model capability continues to improve, some categories that look technically infeasible today may flip within 12–24 months. That means white space is moving. The map is not static. Early position matters, but only if it is attached to real workflow insight rather than abstract excitement.

Conclusion

The finding that 92% of AI development concentrates in 6.8% of activity space is not only a description of the present. It is also a map of where competitive differentiation is likely eroding fastest.

The white space matters because it is harder. The workflows are messier, the trust bar is higher, and the operational details are less forgiving. That difficulty is precisely why strong positions there can be more durable.

If the MIT paper is right, the next major AI companies are more likely to emerge from the 93.2% than from the crowded 6.8%. Not because the white space is easy, but because hard markets with real unmet need are where durable products tend to get built.

Read the diagnosis first: MIT mapped every AI app against every job. Here's what they found.