Understanding Proprietary Data in Artificial Intelligence

Over the past several months, our team has broken a string of stories about AI agents — OpenAI's hiring of the founder of OpenClaw, Microsoft sales leadership emailing staff to pitch against OpenAI's new agent platform, Salesforce privately retreating from pure LLM reasoning in Agentforce, and a dozen startups founded by former OpenAI and DeepMind researchers raising at billion-dollar valuations to build agents for narrow verticals. I've been covering tech for more than two decades. I have not seen a gap between the public narrative and the business reality this wide since the early days of crypto. The public narrative says general-purpose AI agents that do everything are coming fast. The reporting tells a different story. Enterprises are losing trust in LLMs, not gaining it. The biggest platform companies are already fighting over who controls the layer between the model and the customer. (That's what those Microsoft internal emails were really about.) And the smartest money isn't going into do-everything agents. It's going into narrow ones trained on licensed data from specific industries — companies like Applied Compute (raising at a $1.3 billion valuation for legal AI) and Elorian (visual reasoning agents from ex-DeepMind researchers). That last part is the one I keep coming back to. OpenAI and Anthropic are actively courting biotech and financial firms to license proprietary data — genomics, tax records, code bases — to train agents that can do expert-level work in those fields. The companies making these deals right now are shaping competitive dynamics that will play out for years. The agent story isn't really a product story. It's a story about how specialized knowledge gets restructured across entire industries. And the details—who is selling what data, to whom, at what price, under what terms—matter enormously. That is the kind of reporting we built The Information to do. If you have ever considered subscribing to The Information, now is the time. https://lnkd.in/gNZJNJpM

If you use/subscribe to an AI system and you: 1) place your name or trademark on a high-risk AI system that has already been launched or is in use: or 2) make a substantial modification to a pre-existing high-risk AI system; or 3) make a significant alteration to an AI system which turns it into a high-risk system - then pursuant to the forthcoming AI Act - you’ll no longer be a user, but rather the provider of the system - with all the regulatory obligations that will bring. The substantial modification point is especially interesting. If you use a foundation model that’s directly integrated into a high risk system, and you make substantial modifications to the system, then you’ll be considered the provider of that model. This is very important because in a few years we will all be fine tuning models for our own business purposes on our own data, and that fine tuning in many cases will result in substantial modifications to the extent you may be considered the provider. It’s important to considerthe fine-tuning process through the lens of AI law, focusing on aspects such as data protection, intellectual property, and creators' rights. Creating a training and test set from an initial dataset requires meticulousness not just from a technical standpoint but also from a legal one. Any data used must respect privacy laws like GDPR in the European Union or CCPA in California. As models are fine-tuned iteratively, the collection of more data is inevitable. Here, the rights of data subjects and creators must be upheld. This includes the right to be forgotten, whereby data must be deleted upon request, and the right to data portability, allowing individuals to move their data between services. The token limits imposed during the training process are not merely technical constraints; they could also possess legal implications. For instance, truncating data might result in the loss of context, leading to biased or inaccurate models or data inaccuracy on unsuitability which could contravene the AI Act. This could also raise questions of liability if the model's outputs cause loss or harm. The formatting scripts and algorithms employed in fine-tuning are often proprietary. Therefore, it's crucial to have a clear understanding of intellectual property rights when using or modifying such algorithms. The code itself may be copyrighted, and any deviations for fine-tuning could potentially infringe on the creator's rights. Once a fine-tuned model is created and employed, allocating risk and responsibility becomes a complex legal issue. Does liability lie with the original model's creators, the team that fine-tuned it, or the end-user? The specific configurations used for fine-tuning, known as hyperparameters, could be considered trade secrets. Therefore, their disclosure may be legally restricted. Let me be clear - Fine tuning models is about to become a commercial necessity.

I passed on 473 AI startups last year who claimed proprietary models or data. None had actual moats. As someone investing in AI with data moats, I can spot the pretenders instantly. What founders claim: "We have a proprietary AI model" - Built on OpenAI API? "We have unique training data" - 1,000 data points they scraped? And the data points aren't directly relevant to their agentic use-cases. "We have technical advantages" - Using same tech as everyone else? Your RAG isn't a technical advantage. It's just your application's business logic. "We fine-tuned for our vertical" - and then in diligence I discover it’s just better RAG... fine-tuning requires really data science work. None of these are moats. The brutal truth about AI startups: 95% are thin wrappers on OpenAI/Anthropic APIs. When GPT-6 or Claude 5 comes out, your "advantage" disappears. When OpenAI launches native feature, you're dead. Your "proprietary" is replicable in 3 months. One founder: "We built a proprietary model for legal research." Me: "What's it built on?" Founder: "We fine-tuned Llama on legal documents." Me: "How many documents?" Founder: "About 500." Me: "What happens when OpenAI releases GPT-6 with better legal reasoning on 100k documents?" Founder: "Well... uh..." No moat. Another: "We have proprietary data for AI training." Me: "How much data? Is the data applicable to the use-cases you are solving for?" Founder: "We've collected 50,000 data points over 6 months." Me: "How long would it take a competitor to replicate? Is it available to everybody?" Founder: "Probably... 6 months? Yes, you can scrap the data online" That's not proprietary. That's a head start. Real AI moats are rare: Moat 1: Proprietary data that can't be replicated Years of data nobody else has today Legally protected or exclusive via partnership Moat 2: Network effects in data Each incremental user improves model for all users Each incremental agent launched and course correct by users builds more reinforcement learning data improving agents Competitor starting from zero has inferior product Moat 3: Vertical integration that's prohibitively expensive Own the full stack from data collection to deployment Data derived from proprietary Hardware Requires $50M+ to replicate. A great example is autonomous vehicle companies with proprietary sensor data + fleet Moat 4: Regulatory capture Exclusive rights through regulation Years to replicate approvals If you're building AI startup: Stop claiming your model is proprietary unless you truly can't be replicated for years. Focus on distribution. Build actual moat: Data, network effects, or regulatory. Or accept you're a feature, not a company, and build accordingly. The AI gold rush is creating hundreds of companies that won't exist in 3 years. Don't be one of them. #AI #StartupStrategy #VentureCapital

With the merger of xAI and X, Elon Musk has done something critical in today’s AI race: he’s secured a data moat. In an era where Ilya Sutskever warns that “we have but one internet,” access to proprietary, high-quality, real-world data is becoming the true differentiator—not just model size or architecture. While OpenAI and Anthropic often lead the narrative on innovation and model capabilities, ironically, their data moats are easier to cross. They rely heavily on publicly available data and curated datasets, which are nearing exhaustion. Meanwhile, Meta (with Facebook and Instagram), Google (with search, YouTube, and Gmail), and now Musk (with X) are locking down massive, natural-language data streams at scale. These aren’t just competitive advantages—they’re defensible ecosystems. In the AI arms race, the real question isn’t “Who has the best model?” It’s: Who owns the best data?

Is 𝗩𝗶𝗯𝗲 𝗖𝗼𝗱𝗶𝗻𝗴 𝗘𝘅𝗽𝗼𝘀𝗶𝗻𝗴 𝗬𝗼𝘂𝗿 𝗖𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝘃𝗲 𝗘𝗱𝗴𝗲 and IP? (Post 6 of ~27 in my Public AI Risk Series) Did you know that 𝗽𝗮𝘀𝘁𝗶𝗻𝗴 𝗽𝗿𝗼𝗽𝗿𝗶𝗲𝘁𝗮𝗿𝘆 𝘀𝗼𝘂𝗿𝗰𝗲 𝗰𝗼𝗱𝗲, 𝘂𝗻𝗶𝗾𝘂𝗲 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀, 𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗜𝗣 𝗶𝗻𝘁𝗼 𝗮 𝗽𝘂𝗯𝗹𝗶𝗰 𝗟𝗟𝗠 is a big risk. Bug fixing is, too. It sounds harmless: "Can you help me debug this?" But here's reality: 𝗬𝗼𝘂'𝗿𝗲 𝘁𝗿𝘂𝘀𝘁𝗶𝗻𝗴 𝘃𝗲𝗻𝗱𝗼𝗿 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝘁𝗼 𝗽𝗿𝗼𝘁𝗲𝗰𝘁 𝘆𝗼𝘂𝗿 𝗰𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝘃𝗲 𝗲𝗱𝗴𝗲 — 𝗮𝗻𝗱 𝘁𝗵𝗼𝘀𝗲 𝗽𝗼𝗹𝗶𝗰𝗶𝗲𝘀 𝗰𝗮𝗻 𝗰𝗵𝗮𝗻𝗴𝗲 𝗼𝘃𝗲𝗿𝗻𝗶𝗴𝗵𝘁. Once your code hits their logs, you no longer control where it lives or how it's used. Why it's a 𝗿𝗶𝘀𝗸: Public AI creates multiple exposure points: ⛔ Retention. Your code resides in vendor logs you don't control. ⛔ Training potential. Vendors can — and do — change training policies. ⛔ Model inversion. Rare, but proven: structural logic can be reconstructed through careful querying. And there's more (see the comments). Put simply: 𝗜𝗳 𝘆𝗼𝘂𝗿 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗼𝗿 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝘁𝗼𝘂𝗰𝗵 𝗮 𝗽𝘂𝗯𝗹𝗶𝗰 𝗺𝗼𝗱𝗲𝗹, 𝘁𝗵𝗲𝘆 𝗺𝗮𝘆 𝗹𝗲𝗮𝘃𝗲 𝗳𝗶𝗻𝗴𝗲𝗿𝗽𝗿𝗶𝗻𝘁𝘀 — 𝗶𝗻 𝗹𝗼𝗴𝘀, 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀, or 𝗳𝘂𝘁𝘂𝗿𝗲 𝗺𝗼𝗱𝗲𝗹 𝗯𝗲𝗵𝗮𝘃𝗶𝗼𝗿. This is how IP can leak without anyone noticing. Think of it like this: Using public AI to debug company code is like bringing a prototype engine into a crowded workshop and asking strangers for advice. They'll help… but now everyone knows how your engine works, what makes it special, and what makes it fast. And some of those people might be building engines, too. 𝗔 𝗯𝗶𝗴 𝗺𝗶𝘀𝗰𝗼𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻: People think: "I deleted the chat, so it's gone." Not even close. The deletion button clears your view — not the vendor's logs, not the retention systems, not the audit trails. It's a bit like deleting messages off your cell phone. 𝗧𝗵𝗲 𝗰𝗼𝗿𝗲 𝗶𝘀𝘀𝘂𝗲: 𝗣𝘂𝗯𝗹𝗶𝗰 𝗔𝗜 𝗮𝗻𝗱 𝗽𝗿𝗼𝗽𝗿𝗶𝗲𝘁𝗮𝗿𝘆 𝗰𝗼𝗱𝗲 𝗱𝗼 𝗻𝗼𝘁 𝗺𝗶𝘅. It's your secret sauce. Your differentiation. Your advantage. Why risk that on infrastructure where retention, access, and training decisions sit outside your control? 𝗧𝗵𝗲 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻? Use AI. But use Private AI. Private AI for proprietary work. Private AI lives in your environment, under your security, following your rules. No vendor logs. No policy surprises. Just sovereignty over your IP. 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆 𝗡𝗲𝘃𝗲𝗿 𝗽𝗮𝘀𝘁𝗲 𝘀𝗼𝘂𝗿𝗰𝗲 𝗰𝗼𝗱𝗲 𝗼𝗿 𝘂𝗻𝗶𝗾𝘂𝗲 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝗶𝗻𝘁𝗼 𝗽𝘂𝗯𝗹𝗶𝗰 𝗔𝗜. It's not just code — it's your competitive edge. 🌐 𝗣𝘂𝗯𝗹𝗶𝗰 𝗔𝗜 𝗰𝗿𝗲𝗮𝘁𝗲𝘀 𝗿𝗶𝘀𝗸. 𝗣𝗿𝗶𝘃𝗮𝘁𝗲 𝗔𝗜 𝗽𝗿𝗼𝘁𝗲𝗰𝘁𝘀 𝗶𝘁. I've added a simple explainer in the comments. #AI #Cybersecurity #PrivateAI #DataSecurity #DevSecOps

Here's a question: Why are so many businesses using the exact same off-the-shelf AI tools as their direct competitors and expecting to gain a unique advantage? A real, sustainable competitive edge doesn't come from a shared product. It comes from building your own intellectual property. This is the fundamental difference between 'renting' a generic AI and owning a bespoke one. When you build a custom AI, it’s trained on your most valuable asset: your proprietary data. Your internal process logs, your unique customer interaction history, your specific performance metrics. This is a goldmine that generic tools simply cannot access or understand. Let’s make this practical. Imagine a UK manufacturing firm struggling with machinery downtime. They try a generic predictive maintenance tool. It fails. Why? Because it can't integrate with their proprietary sensors or understand the unique operational stresses of their specific machinery. With a bespoke solution, you build an AI that: ✅ Integrates perfectly with their existing legacy SCADA systems. ✅ Is trained exclusively on their years of historical performance data (vibration patterns, temperature, etc.). ✅ Understands the specific failure signatures of their machines. The result isn't a generic dashboard. It's a pinpoint-accurate prediction that a critical component will fail in three days. Maintenance is scheduled, production isn't disrupted, and the business saves a fortune. That is an advantage your competitors cannot copy. That’s your secret weapon. Read more on our new blog: https://lnkd.in/eHk4tD42 If you could build an AI to solve just one unique, high-value problem in your business, what would it be? #BespokeSoftware #PredictiveMaintenance #AIforManufacturing

OpenAI just told #pharma exactly how it plans to extract #data and #IP. Most people missed it. With the launch of their new Frontier platform, they embed OpenAI employees—"Forward Deployed Engineers"—directly inside client teams. The announcement frames this as bespoke enterprise support. But just read the fine print carefully: ❝As you deploy agents, we learn not just how to improve your systems around the model. We also learn how the models themselves need to evolve to be more useful for your work.❞ https://lnkd.in/eRh5q2gd That is a training data pipeline described in consulting speak and covered in sheep's clothing. For #HEOR, #HTA, and #clinical research teams, this should raise red flags. These are environments where teams work daily with proprietary and confidential information. When an OpenAI engineer sits "side by side" with your team, the knowledge they absorb includes your: - proprietary models and workflows - pricing assumptions/rationales - regulatory strategies - patient-level clinical trial data - evidence packages, datasets, endpoints - cost-effectiveness modelling approaches That knowledge flows back to OpenAI Research. Their announcement says so explicitly. They call it a feature. But the improved models are then made available to every subsequent customer. Including your competitors. Three questions every biopharma organization should be asking before entering an arrangement like this: ❓ 𝗪𝗵𝗲𝗿𝗲 𝗱𝗼𝗲𝘀 𝘁𝗵𝗲 𝗱𝗮𝘁𝗮 𝗴𝗼? If proprietary information, know-how, or metadata reaches a vendor's research division, you are financing their data acquisition operation, not your deployment support. ❓ 𝗪𝗵𝗼 𝗯𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴? If the vendor's models are improve based on your domain expertise and proprietary data, then those models are made available to your competitors, you are subsidizing the commoditization of your own competitive advantage. ❓ 𝗪𝗵𝗮𝘁 𝗶𝘀 𝘁𝗵𝗲 𝘃𝗲𝗻𝗱𝗼𝗿'𝘀 𝘁𝗿𝗮𝗷𝗲𝗰𝘁𝗼𝗿𝘆? OpenAI has a documented pattern of observing what customers build on its platform then launching competing products. Your today's deployment partner may become your tomorrow's direct competitor. Of course the pharmaceutical industry needs AI in evidence research. But the question is whether the architecture of the AI partnership protects your data, your IP, and your competitive position... ...or it quietly transfers all three to a vendor whose business model depends on learning from you and selling what it learns to everyone else. The alternative? Purpose-built AI tools designed for HEOR, HTA, and clinical research, validated against established scientific methodologies, and structured so that your data stays yours. Not training meat for a foundation model that serves the entire industry. OpenAI described the economics of this arrangement in their own words. Better believe them. #PharmaceuticalResearch #AIinResearch #AIinHealthcare #DataGovernance #ClinicalResearch

Sovereign AI isn’t about owning everything. It’s about owning the thing that makes you irreplaceable and valuable – your organization's knowledge and IP. What is your knowledge layer? Semantics, models, policies, documents, IP, and data structure. This is the information that AI needs to work accurately inside your organization. Some enterprises are optimizing for speed and cost—they hand their advantage to vendors. They spin up chat, hard-code agents to proprietary frameworks, and pour semantics into someone else’s knowledge graph. That’s not an AI strategy; it’s technical debt with a countdown timer. Here’s the simple test: a compute, model, storage, or app provider changes terms tomorrow. What is the cost and time to adapt?  My view on how the vast majority of organizations ought to think about the layers of their AI stack: - Physical infrastructure: Rent it; avoid irreversible dependence via portability. - Models: Choose where you need control vs. speed; evaluate relentlessly.   - Data: Avoid the siren song of centralization. Move what you need to solve the business problem that you have. - Knowledge: Owning it is non-negotiable – this is your sovereign core. - Apps & tools: Stay nimble; swap freely when the roadmap shifts. Models will be replaced. Frameworks will churn. Your knowledge compounds, but only if you own it. 

Why data moats are the only true moat in AI. LLMs, or large Language Models, are quickly becoming commodities. While algorithms can be replicated and improved upon, the proprietary data that trains these models is emerging as a unique and invaluable asset. Meta, Google, and OpenAI have already either exhausted or found the limits of what they can scrape off the open web for AI training. This has kicked off a gold rush for suppliers who can help them acquire unique and original training data. And that market is projected to be big. According to Business Research Insights, they estimate the AI data training market is worth $2.5 billion and believe it could grow close to $30 billion in 10 years. Anyone who observed how the internet search market evolved won’t be surprised by any of this. Google’s dominance in the search space has always been attributed to the fact that they were able to amass a large library of searches and were much better at understanding what they wanted from a search than they were at crawling the searching the web. I am sure this insight will be key in how the AI market plays out. The platforms, tools, and apps that master the user experience and can provide the desired result, over just escalating the arms race by adding more processing power and more training data, will win. What does this mean for other AI startups without the deep pockets of a big tech company? → Prioritize Proprietary Data Collection Focus on gathering unique data that is difficult for competitors to replicate. → Develop Strategic Partnerships Collaborate with organizations that can provide access to exclusive datasets. → Enhance Data Quality and Variety Invest to ensure your data is high in quality and diverse in nature. → Invest in Data Acquisition Technologies Acquire data in innovative ways like from IoT devices, or other methods. It’s clear that the key to long-term success in AI lies in building a robust data moat. You can create a sustainable competitive advantage by focusing on acquiring, maintaining, and utilizing proprietary data.

DeepSeek AI didn’t upend the AI landscape. It just confirmed what we already knew.Large proprietary models were never going to be the value creator for enterprise AI. Your first-party data is. LLMs are raw material. Nothing more. If you don’t use your hands to mold it into something, it’s about as useful a lump of clay on a table. But mould it into shape with some curated, defined, and properly governed first-party data, and you’ve got the makings of some real business value. Model development will plateau. Innovation will stagnate. Access will commoditize. And the value will always settle where it started in the first place: with your data—and what you choose to do with it.