It’s been a little over a year since my colleague, Check Warner, published our AI investing framework. That article sparked fascinating conversations, many of which guided our most exciting investments to date.

‍

Since then, we’ve backed twelve AI-first companies across the application and applied infrastructure layers. We continue to keep pace with developments in the broader AI world – of which GenAI is the latest wave – not only through interactions with exceptional founders but also through our own software and tooling efforts. The latter has been a significant part of my role since I joined Ada Ventures. It includes building a beta version of AdaGPT, a pitch deck feedback app used by over 200 founders per month.

‍

How has our thinking evolved? In this post, I’ll revisit our original framework and highlight how our views have grown. I’ll also share use-case areas we are excited about. This includes our views on reframing edtech, the overlooked consumer category, supporting frontline workers, how GenAI can enter the physical world, and the case for investing in services businesses. Along the way, I’ll touch on GenAI’s real-world traction, its current limitations, and some of the broader social implications.

‍

What We’ve Learned Since Our First Framework

‍

Our original framework started with a brief definition of AI-first companies. These are businesses where AI is at the core of what makes their products valuable. For example, Ada portfolio company Bilanc uses AI to measure and improve the productivity of both human and AI-powered engineering teams. Bilanc’s product could not exist without LLMs.

‍

We’ve seen this kind of framing before. Previous waves of innovation had categories such as cloud-native (e.g., Salesforce and Snowflake) and mobile-first (e.g., Instagram and Uber). We now have a new category of opportunity with GenAI that holds just as much promise.

‍

Winning Inclusively, at the Application Layer and Beyond

‍

As Check wrote, we see this new opportunity of GenAI through the lens of inclusive alpha. In other words, AI-first companies are compelling investment opportunities because they have the potential to deliver substantial financial returns while also increasing access to essential services.

‍

Our investments in fast-growth companies like Medly AI (equitable learning and education through AI tutors) and Valla (consumer access to AI-powered legal support) have reaffirmed our view that significant value will accrue at the application layer. We’ve also found that the scope and scale for positive impact are greater here. Additionally, we’ve noticed that the consumer category remains untapped. We plan to explore this opportunity more deeply.

‍

Meanwhile, we remain optimistic but measured about the applied infrastructure layer. In this category, we recently backed JigsawStack because it solves the messy backend of building AI-powered products. This, too, has an inclusive angle: Jigsawstack’s products enable more people – not just the technical elite – to build and participate in the AI software economy. 

‍

On the foundational model side, our stance has hardened. Unless a team is developing a new, capital-efficient, and defensible AI paradigm that captures value across the foundational model and application layers, we’re unlikely to participate as a pre-seed investor.

‍

DeepSeek models – one of the many ruptures to the expensive closed-source alternatives – and the steady stream of free models and data sets published on Hugging Face (there’s a new upload every 10 seconds) are eroding the defensibility of proprietary models. Generative intelligence is fast becoming a commodity.

‍

We continue to believe that the story of AI will mirror the historical journey of electricity. While the original invention was groundbreaking, it was the innovation built on top that transformed everyday lives.

‍

Where Gen AI Works Today and Where It Could Be Tomorrow

‍

Developers who got early access to GPT-3 in 2020 will remember how it felt like magic. My first use of the model was in its purest form. There was no instruction fine-tuning or reinforcement learning from human feedback, yet the often verbose and off-topic model responses didn’t overshadow a glimpse into astonishing capabilities. 

‍

Since that early magic, the GenAI narrative has swung wildly across extremes. We’ve seen existential fear (“we’re cooked”), overblown expectations (“this changes everything!”) and disillusioned disappointment (“GPT4.5 underwhelmed!”). There’s evidence to support parts of each perspective, and this makes navigating the AI world especially erratic. 

‍

For instance, the AI job crisis is arguably already here. Video generation has surpassed expectations and is remarkable. However, state-of-the-art models still hallucinate (OpenAI’s o3 model hallucinates more than older models), and real-world productivity gains from GenAI remain uneven: Some randomised control trials have found time savings of 15% or higher while a recent large-scale study of 25,000 workers found an average of 2.8% time savings and no impact on wages.

‍

Still, if you look past the mania, there’s now more clarity about where GenAI can genuinely deliver value. From code and content generation to knowledge retrieval and workflow automation with AI agents, founders are getting a better sense of the capabilities they can build with. We see a leap forward in what’s possible across many use cases. Here’s a sample of areas we are paying attention to.

‍

‍

‍

6 Areas of Potential

1. Applied Learning & Generative Grit

‍

Edtech has been fixated on the classroom format of learning, and technology has done little more than mask the flaws of a rigid factory-like model from the 1800s. But what if GenAI could go beyond that? There’s scope to impact not only existing modes of teaching abstract knowledge but also applied learning. Consider the following ways in which vocational education can be transformed:

‍

• Simulation-based education: Just as pilots use flight simulators, imagine a world where any student can step into a high-stakes simulation as a doctor, electrician, care worker or engineer and interact with AI-generated environments that dynamically respond to their decisions. This isn’t a far-fetched future. We recently ran a hackathon where the winning team built an MVP that does exactly this. Simulated learning brings knowledge to life, stimulating both mental agility and grit.

‍

• Elite mentorship for all: In many fields, rapid skills development depends on access to elite mentors. From athletes and business professionals all the way through to academics (e.g. the Oxford tutoring system) and craftspeople, having proximity to a knowledgeable mentor can accelerate learning. With GenAI, the opportunity to have an AI “master” tutor 24/7 anywhere in the world becomes possible and more accessible. 

‍

2. Interfaces to the Physical World

‍

GenAI doesn’t have to be restricted to bits in the digital realm. It can now interface with the physical world. Of course, we’ll need appropriate safety measures and guardrails (more on that later), but extending GenAI into the physical world unlocks even more capabilities. Take these two, for example:

‍

• Scientific lab agents: The open-source MCP framework (described by its creators as something akin to USB-C for AI) opens up a world of integrations that allows LLMs to communicate with robots, lab equipment and software at scale. This paves the way for a future where you can run thousands of lab experiments autonomously to help advance scientific discovery without the barriers of massive R&D budgets. 

‍

• Ambient monitoring: We already have cameras that monitor the real world. But GenAI adds new elements: multi-modal reasoning across audio, visual, and sensor data, plus interaction through natural language. So rather than use hardcoded rules or dashboards, an engineer could say, “Inspect all load-bearing joints for early signs of stress fatigue and prioritise based on risk levels.” Or at home, you might ask your fridge, “Do I have what I need to prepare a buttermilk chicken curry?” 

‍

3. Frontline Worker AI

‍

Tradespeople, healthcare employees, leisure and hospitality staff and other deskless workers make up most of the global workforce. However, only a fifth of frontline workers have the tech they need to be productive. They are chronically underserved by technology despite an estimated productivity boost of 22% on average if they were granted the right tools. For these reasons, we are on the lookout for innovation here. Examples?

‍

• Task-level intelligence: Imagine a bodycam that would watch and listen as you repair complex machinery. It could flag safety issues and corrective actions and auto-generate a compliance report at the end without adding laborious admin. Some version of this already exists today, as showcased by Google’s Project Astra demo. Similar tech could be applied in other settings and industries.

‍

• Shift-level support: Before a frontline worker starts their shift, an AI assistant could provide a briefing that synthesises a summary of recent hazards, known equipment faults, weather conditions, and other useful context. Then, as that person conducts their work, AI monitors mental state and environment cues, stepping in to suggest breaks, stress management tips and support, or a reordering of assigned tasks to better match a person’s capacity.

‍

4. Breakout Consumer Apps

‍

The bulk of GenAI activity has gone into enterprise software. Taking YC companies as an example (a group that provides a leading indicator on where technology talent is placing its bets), I found that only 10% of AI-first companies founded in the last 2 years were in the consumer category. In Europe, the fraction is halved. 

‍

‍

‍

We feel that this is a missed opportunity. Consumer is tough, but founders who break through can achieve incredible outcomes. For GenAI, this will require going beyond single-turn chat-based AI that’s common in enterprise software to high-value, fun interactions that span long time horizons and mixed modalities. To name a few:

‍

• Long-horizon AI Counsel: Imagine an AI assistant that ingested all your health and activity data on-device and securely. With this, the AI assistant wouldn’t just remind you to go to the gym or schedule a forgotten dentist appointment on your behalf. It could help you build and maintain healthier habits across months and years. We see a transformative opportunity here to build consumer assistants for health, finance, family life, and career development.

‍

• Agentic Representation: Every day, we’re faced with mundane personal life admin that’s usually reversible (if something goes wrong) but often impactful (it affects our way of life.) This could be a utility bill error you need to call your provider about or delivery that has to be rescheduled according to your availability. These tasks aren’t cognitively taxing but are laborious. What if you could delegate them to AI agents who act on your behalf but with clearly defined parameters? Unlike the AI counsel that helps you make better decisions over time, these agents would make small, low-stakes decisions based on your specified preferences.

‍

5. The Small-to-Medium Sized Business Stack

‍

Small and medium-sized companies make up 99% of the USA’s and UK’s total number of companies. They account for 46% and 60% of USA and UK employment, respectively. While many startups gravitate toward enterprise customers because of the higher contract values, SMBs still represent a massive opportunity, given their sheer volume and operational needs. This is why the likes of Shopify, Xero, Canva, and Toast are all billion-dollar companies. We believe GenAI will produce new SMB winners, too. Areas of potential include:

‍

• Enterprise-grade Ops Intelligence: Solopreneurs and small teams have to wear multiple hats. They often take on finance, payroll, marketing, legal, and other operational roles without the expertise of in-house specialists or expensive professional service firms. While GenAI may not immediately replace premium advisory services, it provides an affordable alternative. Imagine an AI CFO (e.g. Nume), a dedicated AI recruiter (e.g. Jack & Jill), or a mobile-first voice intelligence tool like Dukawalla – a research prototype for African SMBs that provides hands-free access to performance metrics and trends, helping entrepreneurs make smarter decisions without the need for spreadsheets or dashboards.

‍

• Agent-Powered Professional Services: Services businesses – whether it’s an accounting firm, a law practice, or a marketing agency – package human talent into billable hours. Small teams in these firms are already using GenAI to produce client deliverables more efficiently. A five-person firm can now do the work of fifty. Human review and accountability remain essential, but with AI, firms get better margins, faster turnaround times, and the ability to serve more clients. Investor Elad Gil is betting big on this trend: acquire traditional services firms, infuse them with GenAI automation, and scale through M&A and roll-ups. It’s a modern take on the classic M&A strategy of economies of scale, except this time we must account for something new: economies of GenAI.

‍

6. GenAI Safety & Guardrails

‍

We’ve had the world wide web for over three decades, and in that time, society has established robust mechanisms to subvert digital threats. Open standards like TLS encrypt data in transit. Payment authentication protocols like 3D Secure prevent credit card fraud. There are also simple ML classification algorithms that filter email spam. All these tools have helped the world create a resilient security system that enables trust on the web. A similar ecosystem for GenAI is embryonic today. And so we’re on the lookout for founders who are building the missing safety stack. A few promising areas:

‍

• Human authentication systems: It’s harder than ever to distinguish between AI and human content online. The world will need tools that identify AI bots and agents, as well as systems that can verify humans in real-time across all modalities (text, audio, video). This might require new cryptography, proof of genuineness (through time or some other value), and behaviour-based primitives. What these authentication systems will look like exactly – we don’t know. But we’re staying open to founders with creative ideas on how to build a trustworthy GenAI ecosystem. 

‍

• Guardrails and safety for autonomy: The more we delegate to AI agents the more we’ll need guardrails for autonomy. LLMs are brittle and minor changes in the context provided or even simply reordering or rephrasing a prompt can yield widely different and potentially dangerous outputs. There’s also the issue of prompt injection hacks (see an example of this GitHub MCP exploit that’s akin to SQL injections of previous eras.) This presents opportunities for founders to build applied infrastructure that can protect against biases, ethical blind spots, inaccurate world models, and reasoning drift. 

‍

What’s Next for GenAI?

‍

AI capabilities are notoriously difficult to predict. At Ada Ventures, our view is to keep an open mind about what might surprise us next while staying focussed on real value creation. We also recognise the limits of our knowledge. We have to ferociously learn from builders who are redefining the frontier of what GenAI can do. So if you have thoughts on this article, we’d love to hear them. And if you’re an AI-first founder building something that feels slightly crazy but is grounded in addressing real-world needs, get in touch!

‍

‍