Judging many AI startups in quick succession is an unusual privilege. You get a compressed, cross-sectional view of where an industry is, not just the polished stories that make LinkedIn, but the full range. The companies that are genuinely onto something. The ones that were a great idea two years ago. The ones that will be absorbed by a feature update in the next twelve months. And those that are building something the major labs cannot easily replicate.

What follows is a set of observations about how AI businesses are evolving, what is working, what is being swallowed, and where I think the next real opportunity sits. I have drawn on the submissions directly without identifying specific companies. The AI for Good Innovation Factory Australia 2026 gave me this view.

1. The Window That Opened and Then Closed

The first wave of AI startups had a simple thesis: large language models are powerful but they do not know your data. The opportunity was to add proprietary data to generic models and charge for the resulting intelligence. This was, for a brief period, a genuine opportunity.

Then the models got better at retrieval. RAG (Retrieval-Augmented Generation) became a commodity. The major platforms built native document ingestion. The window for ‘we add your data to an LLM’ narrowed to almost nothing in roughly twelve months.

Several submissions in this cohort arrived at that window after it had already closed. The pattern is recognisable: a well-articulated problem, a plausible AI layer, and a value proposition that would have been defensible in 2023 but now relies on execution advantages that are very difficult to demonstrate against well-resourced incumbents. The lesson is not that these founders were wrong. They were right at the time. The lesson is about timing and speed of execution. Medvi, the US telehealth startup built in 8 weeks, is an external example of success in this space (it’s a fascinating case study that probably deserves an article of its own).

The question every AI founder should be able to answer clearly: what does your product do that a determined user with access to OpenAI, Anthropic, or Google’s tools could not approximate on their own? The answer must go beyond the interface. It points to one of three things: a proprietary dataset the models cannot access; a fine-tuned or bespoke model trained on your data that produces outputs the general models cannot replicate; or a workflow integration so deeply embedded that the AI layer is inseparable from the product itself.

The unstructured data play, and whether that window is also closing

A cluster of submissions was built on a second-order version of the same thesis: not ‘add your data to an LLM’ but ‘make your unstructured data AI-readable in the first place.’ The argument is that industrial knowledge, legal precedent, regulatory frameworks, scientific literature, and clinical records exist in formats that AI cannot parse: PDFs, scanned documents, legacy databases, fragmented websites.

Some are building AI agent fleets, agents that learn from each other rather than starting from scratch on every deployment.

The honest question to ask each of these companies is: how long before multimodal models and better document parsing make this generic? The most sophisticated answer in this group was that the value is not parsing but state reconstruction, engineering-grade rigour rather than probabilistic inference. The companies that will survive in this space are those that can articulate why their structuring layer is not replicable by a better prompt or a newer model.

The window is not closed yet. But it is narrowing. Rather than better parsing, the companies in this space need to be building network effects, proprietary datasets, or domain-specific validation layers that compound over time.

2. The Data Play That Still Exists

Those with the best chance of succeeding are those with deep domain specificity and proprietary data. The AI layer must be inseparable from a dataset or domain that cannot be replicated. A mineral exploration intelligence platform, built on geological data and validated by a major mining company’s accelerator, is not replicable by a better prompt. A medical diagnostics model trained on clinical data that took years to accumulate is not replicable from a public corpus. The data asset is the moat. The model gets refined for the optimum output.

These businesses are essentially doing work that was previously just not worth doing. They are not replacing a human task, but making economically viable something that was always theoretically possible but prohibitively labour-intensive. Reviewing every lease clause in a portfolio. Screening every satellite image for a geological anomaly. The man-hours were the blocker, not the knowledge. AI removes the blocker entirely. The companies in this space are not automating existing workflows. They are creating new categories of work that simply were not done before.

3. Removing the Regulation Burden

A clear theme across a number of submissions is the reduction in admin burden, the back-office hours that are required but add no revenue. The opportunity is existing regulation that AI can finally make navigable without specialists.

Tax returns. NDIS management. Licensing approvals. Complex compliance frameworks that have historically required expensive expertise and significant time. AI can absorb that complexity and build a usable interface between regulation and the people subject to it. This is less about the regulatory environment changing and more about AI finally being capable enough to remove the bureaucracy that has always existed.

The nuance that matters: if your product amounts to applying a widely available regulatory rulebook to an AI, can the same rulebook be uploaded to a general model and largely replicated with a $20 subscription and a well-crafted prompt? The moat is not knowing the rules. It is becoming the platform that practitioners rely on habitually, that holds their data, that their accountant or case manager is already using. The question every founder in this space needs to answer honestly: are you building the rails, or a well-designed access point to someone else’s rails?

This does not negate the opportunity. Not everyone will want or have the skills to do it themselves. But speed of execution and customer acquisition will matter more than timing alone. There is a real risk of being undercut in a race to the price bottom. If there is nothing truly unique and needed, the pathway may be short unless switching costs are high or additional services can be built and sold alongside.

4. Companies That Exist Because AI Exists

The Trust Cluster

An interesting category in this cohort is not companies using AI to solve old problems. It is companies solving new problems that AI itself created.

As AI-generated content, synthetic media, and automated decisions proliferate, a new set of infrastructure needs is emerging. Not ‘how do we use AI?’ but ‘how do we verify what AI produces, and who is accountable for it?’ These businesses did not exist before AI made them necessary.

The clearest example in this cohort is a deepfake detection company building truth infrastructure. An AI decision audit trail product sits in the same category. I have seen a few other players in this space and success will come down to model accuracy. If the numbers are there, the regulatory environment is moving toward them. The EU AI Act requires enterprise to demonstrate accountability for AI decisions. This is accountability infrastructure that simply did not need to exist before AI did. Regulation is one of the strongest adoption drivers there is.

One vulnerability in this cluster is the attempt to apply a classification or rating methodology to AI, human, or combined outputs; essentially self-service governance before any regulation exists. It is a noble intent, and there is real competition to win this segment. But the true winners will need to offer more than self-certification. A proprietary methodology with trusted independent certification will win this space. Demand is not driven by enthusiasm. It is driven by regulation, liability, and the simple fact that as AI output becomes ubiquitous, provenance and verification become valuable. The category will consolidate, but the underlying need is durable.

5. What Survives, and What Comes Next

The physical world — hardware using AI

The most underappreciated shift in AI right now is the move from software into hardware. Sensors, edge devices, industrial equipment, wearables; environments where latency, power draw, and offline capability create constraints that cloud models cannot address, or where AI is simply the tool to make something more meaningful happen in the physical world.

A neuromorphic processor company in this cohort is the clearest example, but I have had the privilege of working with other founders in this space and saw further evidence at last week’s Founders Factory event in Perth. Founders combining AI with physical infrastructure are not racing against a general model update. They are generally creating their own models. The constraint is physics, not compute, and that creates a different kind of moat.

Conclusion

What creates success across all of these categories is consistent. Proprietary data or physical constraint that a general model cannot replicate. A uniquely trained model working in a niche. A distribution, partnership, or workflow advantage that compounds over time. Speed of execution.

Some of the companies I have seen have those answers. Some are still working them out. The dot-com era is the closest reference point we have, not just as a warning but as a pattern to be aware of. Some companies will not survive. Some will have limited windows of opportunity. A handful will find genuine moats that look obvious in hindsight.

To everyone who submitted to AI for Good Innovate Australia, and to everyone else building seriously in this space: I wish you every success. The execution required to turn a good idea into a durable business is harder than the idea itself, and I have genuine respect for anyone attempting it. There has never been a better time to build.

These are my own thoughts and impressions, drawn from the judging process, the conversations around it, and my own experience working with companies in this space. Things will look different in six months. I look forward to updating my thinking when they do.

James Reeves

Judge, AI for Good Innovation Factory Australia 2026

If you are building in this space, feel free to reach out

BusinessWorx Group | June 2026