Today every company, including your competitors, can use the same world-class AI models through an API. That is good news for adoption. It also means the model alone is no longer much of an advantage. If you and your competitor both build on the same model, what makes your product better?
Increasingly, the answer is data that only you have. We learned this first-hand building HAI, our health AI platform, and AIPOS.
Why off-the-shelf data was not enough for HAI
HAI turns photos of the eye into a wellness check-in. When we started, no public dataset had the iris images we needed at the quality, consistency and volume required, captured on ordinary smartphones in real lighting, from a population that looks like our users.
So we built one, and that was most of the work. The models came later, and were the comparatively easy part.
What building a proprietary dataset actually involves
1. Design capture into the product
The best datasets are a by-product of a useful product. In HAI, guided capture in the app checks focus, glare and framing in real time, so every image is usable for the user's report and also for training. Practitioners using HAI BOSS add images captured with consistent equipment.
2. Make labelling part of someone's workflow
Raw images aren't training data until they're labelled. Instead of a separate labelling project, practitioner notes and zone overlays recorded in HAI BOSS during normal sessions become structured annotations. Experts label as they work, and the dataset grows with every client seen.
3. Treat consent and retention as features
Health-adjacent data is sensitive. Consent, purpose limitation and retention rules are built into the capture flow from the start, not added before an audit. It's also simply the right way to treat people's data, and it makes the dataset defensible.
4. Version everything
Every model we train is linked to a versioned dataset snapshot. When a model behaves unexpectedly, we can trace exactly what it learned from.
The flywheel
Once capture, labelling and governance run as part of everyday operations, you get a flywheel:
- A useful product attracts users.
- Users generate high-quality, consented data.
- The data improves the models.
- Better models make the product more useful.
A competitor can license the same foundation model tomorrow. They can't license several years of your flywheel.
You probably already have proprietary data
Most enterprises have more than they realise:
- Transactions: years of sales by item, outlet and hour. This is the foundation of AIPOS forecasting.
- Service history: tickets, chats and call notes with outcomes attached.
- Documents: claims, contracts, inspection reports and approvals, each with a decision.
- Operations: sensor and device data from buildings, rooms and equipment.
The problem is rarely having no data. It's data that is scattered, inconsistently labelled and has no clear consent basis for reuse. Fixing that is a data engineering job, and it's the best AI investment most companies can make this year.
Where to start
- List the decisions your business makes thousands of times a year.
- For each, find where the inputs and the outcomes are recorded today.
- Check the consent and purpose under which that data was collected.
- Pick one where data is plentiful and outcomes are recorded. That's your first model.
Not sure what your data is worth?
Our data readiness assessment maps your data sources against high-value AI use cases and gives you a practical plan to close the gaps. Talk to us →
