It’s Time to Understand Your Data’s True Value

A question worth asking

Around two years ago, at the start of the generative AI boom, I wrote that many organisations were missing an opportunity to create new revenue from the data they were already cleaning up for AI. The market has moved a long way since then, and I think the question deserves a fresh look, because the value that AI companies now place on high quality, proprietary data has changed considerably.

When we launched WeBuild-AI, we spoke a great deal about helping customers reframe how they use their data for commercial gain. In truth, the idea did not gain much traction at the time, as most organisations were understandably focused on getting their first AI use cases into production. With the market for data now developing quickly, I think that conversation is finally starting to change.

Most enterprises are understandably focused on adopting frontier models and putting them to work across their operations, and that is where a great deal of value is being created. Alongside that work, it is worth leaders asking whether their data strategy reflects what their data is now worth, both to their own AI ambitions and to the wider market that is forming around it.

What has changed since 2024

The frontier model providers have now trained on most of the useful text available on the public internet, and the search for new, high quality material has turned into a commercial market. Model developers are paying content owners for access, in part to broaden their training data beyond large scrapes of the web.

Two examples show how far this has moved. In 2024 Reddit agreed a licensing arrangement with Google reported to be worth around $60 million a year, giving Google access to its content for training. In 2025 Anthropic agreed to pay $1.5 billion to settle a class action brought by authors whose books had been taken from pirate sites, which works out at roughly $3,000 per book. The court in that case found that training on lawfully acquired books was permissible, which has encouraged providers to buy and digitise printed material directly.

The common thread is that well written, trustworthy and hard to obtain content now commands a real price. Enterprise data, much of which has never been published anywhere, sits firmly in that category.

A new market for enterprise data

A growing number of firms are now approaching enterprises with offers to buy their data, repackage it or broker it on their behalf. In many cases the likely end use is the training of frontier models, even where that is not stated plainly in the commercial terms.

I have seen this first hand in recent months, having received an unsolicited email offering £300,000 for our company’s data, while my LinkedIn feed has filled with adverts telling me that my company’s data is worth £2.5 million and that I should sell it now. Neither approach showed any real understanding of what our data contains or who might value it, and the gap between the two figures suggests how loosely many of these valuations are being formed.

Some of these offers will be perfectly reasonable, but leaders should be cautious about accepting the first price put in front of them. An intermediary needs to resell at a margin, so the amount offered is unlikely to reflect what the data is worth to the organisation that eventually uses it. Before entering any arrangement, you should understand what you hold, who would value it and on what terms you would be comfortable sharing it.

Why proprietary data matters

The data that frontier models still lack is the detailed, operational knowledge held inside businesses, which has never been published and cannot be scraped. This includes transaction histories, asset and maintenance records, risk assessments, regulatory correspondence, customer interactions and the many years of expert judgement captured in documents and systems.

Regulated sectors such as financial services and energy hold particularly rich examples. A bank's view of credit behaviour across economic cycles, or an electricity network's record of how assets perform under different weather and load conditions, is very difficult for anyone else to reproduce. That scarcity is what gives the data its value, both to the organisation that holds it and to anyone building AI for that sector.

Fixing data at source

None of this value can be realised while data remains fragmented, poorly described and of uncertain quality. Many organisations still struggle to say where their data is, who owns it and what it actually contains, and that uncertainty weakens any position they might take, whether internally or with a third party.

The practical starting point is to address problems where the data is created, rather than repeatedly cleaning it further downstream. You should begin with an audit of the data your organisation holds and how it is used, which will show which sets are in most demand and which are underused. From there, assess quality against the familiar measures of accuracy, completeness, timeliness, consistency and relevance, and assign clear ownership for each important data set.

It also helps to treat your most valuable data as a product. Give each data set a clear name, a stated purpose and a description that someone outside the team could understand, supported by good metadata so that every field has an agreed meaning. Data prepared in this way is easier to use for your own AI and far easier to value if you later decide to share it.

Putting your own data to work

Once data is in good order and organised around a clear model of how the business works, it becomes much more useful to the organisation itself. Frontier models can draw on it to reason across complex questions, while smaller models can be trained or tuned on it to carry out specific, repetitive tasks within defined business processes.

Those smaller models are well suited to much of the routine work found in large organisations, such as classifying documents, extracting information from forms, checking cases against policy, triaging requests and drafting standard correspondence. Because they are focused on a narrow task and trained on your own data, they can be cheaper to run, quicker to respond and easier to govern than relying on a general purpose model for everything. They can also be hosted within your own environment, which matters a great deal where data is sensitive or regulated.

A sensible approach is to reserve frontier models for the work that needs broad reasoning and use smaller models for the high volume tasks around it. You should start by identifying a handful of processes where the work is well understood and the data is already in reasonable shape, prove the approach there and then extend it.

Sovereignty and control

Sovereign AI is often discussed at the level of nations, but the same idea applies to individual organisations. It describes the ability to run AI on your own terms, using data and models that you control, within boundaries that you set.

When data is structured around an ontology of your business, meaning a shared definition of the entities, relationships and rules that describe how it operates, it becomes considerably more valuable to your own AI. It also becomes harder to replace, because that structure reflects knowledge only your organisation holds. Once such data has been handed to a third party and absorbed into a general model, much of that advantage is difficult to recover.

The practical benefits are significant. In our experience, grounding AI in an ontology noticeably reduces hallucinations, because models are working from defined entities and verified relationships rather than inferring connections from loosely related text. When an answer can be traced back through the ontology to the underlying records, it is also far easier to check and explain, which matters greatly in regulated industries.

An ontology also makes data much easier to find. When every data set is described in the same business terms, people and AI agents can locate the right information without needing to know which system holds it or what an individual table happens to be called. Better discovery leads to more consistent answers, and consistent, explainable answers are what build trust in AI across an organisation. You should therefore treat the ontology as part of your core data foundation, and design it with the people who understand how the business operates.

This does not mean data should never be shared. It does mean that any decision to share it should be taken deliberately, with a clear view of what is being given up, what is being received in return and how the arrangement supports your own longer term AI capability.

Commercial options worth considering

Where there is a case for sharing data externally, licensing generally offers more control than an outright sale. A licence can set out how the data may be used, for how long, whether it may be used to train models and what happens to it when the agreement ends. That allows the organisation to earn revenue while retaining ownership.

Sector partnerships are another option. Organisations facing similar problems can pool anonymised data under agreed rules to build shared models or benchmarks, which can benefit every participant without any one of them giving up its position.

Finally, as the examples from Reddit and the publishing industry show, model providers have demonstrated that they will pay for high quality content they cannot obtain elsewhere. For organisations holding genuinely scarce data, a direct agreement with a model provider on clear terms may be worth exploring, and it is reasonable to expect any such arrangement to reflect the full value of what is being provided.

Legal and ethical considerations

Any use of data beyond its original purpose needs a firm legal footing. Where personal data is involved, you should check whether the original basis for processing covers the new use, and whether anonymisation is robust enough that individuals cannot be identified once data is combined with other sources. UK GDPR and the Data (Use and Access) Act 2025 should be the starting points for UK organisations, alongside any sector rules set by regulators such as the FCA or Ofgem.

Reputation matters as much as compliance. Customers and employees are increasingly aware of how their information is used to train AI, and an arrangement that is lawful but poorly explained can still damage trust. Being open about what is shared, with whom and for what purpose is the most reliable way to protect both.

Closing reflection

I promise this isn’t an “I told you so” moment, although, if I’m honest, I did tell you so. A few years ago I made a similar case about how quickly AI would speed up app development and was more or less laughed out of the room. These days you can build a working app over a cup of tea, and the world keeps moving faster than most of us expected.

So I’ll take my chances again. Your data is worth far more than the figure in that unsolicited email, and getting it sorted at source means you can put frontier and smaller models to work on your own processes, keep hold of something valuable and, if the right offer does come along, negotiate from a position of strength. If I turn out to be right this time too, I’ll try very hard not to say it twice.

Which leaves one question to sit with. If someone offered you a cheque today for your organisation’s data, would you sell it to them forever, or would you rather license it on a pay as you go basis and charge them every time their model comes back for more?