What is Data?

Ideas for a Better World · No. 67

Your data is now your most important strategic asset.

In the agentic era, this is not a slogan. It is a description of what will determine your company's competitive position over the next decade more than your product roadmap, your hiring strategy or your balance sheet. The companies that win will be the ones that played their data deliberately, offensively where they could and defensively where they had to. The companies that lose will be the ones that treated data as a back-office cost while their competitors treated it as the strategic asset class it has become.

This edition is about what changed, what the strategic choices now are, and why almost no one is making them well.

What changed

For most of its history, data was an analytical record. You collected it, stored it, queried it, used it to understand what had already happened. The job description has been rewritten three times in five years.

Data became a substrate. A foundation model does not consult data, it consumes it. The dataset becomes the model's behaviour. Your AI's capabilities are a direct function of the data it was trained on. Generic data produces generic models. Defensible AI positions require defensible data positions.

Data became scarce. The first foundation models trained on data that was effectively free: the open internet, scraped at scale. That era is ending. In June 2024, the research group Epoch AI published a peer-reviewed estimate that the stock of high-quality public text (around 300 trillion tokens) would be exhausted between 2026 and 2032. The licensing economy is already responding. Reddit's February 2024 deal with Google was reported at $60 million a year for training data access. A subsequent Reddit-OpenAI deal is estimated at around $70 million. News Corp's deal with OpenAI was reported at over $250 million across five years. The same content that was scraped for free in 2020 now has a price tag, and it is rising.

Data became contested. The New York Times sued OpenAI and Microsoft in December 2023; similar suits have followed from authors, music publishers, photographers and code developers. The question of who owns training data is being resolved in court right now, and the answer will shape the next decade of AI economics.

Generic data produces generic models. Defensible AI positions require defensible data positions.

These three shifts compound. Data is now scarce, valuable, contested substrate. None of those adjectives applied to data ten years ago.

Why the agentic era raises the stakes

Add agents to this picture and the stakes change again.

An agent is not a chatbot. A chatbot answers questions; an agent takes actions. It reads documents, writes code, makes purchases, runs workflows, files things, contacts customers, schedules things, decides things. The data quality that used to determine how good your dashboards were now determines how good your operations are.

Three consequences follow.

The cost of bad data goes from analytical to operational. A wrong figure in a report can be corrected. A wrong action by an agent has already happened. Provenance, accuracy and structure stop being housekeeping and start being first-order operational concerns.

The cost of missing data becomes immediately visible. An agent cannot act in a domain you have no data for. If your competitor has structured customer interaction data and you have unstructured emails, your competitor's agents do things yours cannot. The advantage is not subtle, and it is not easy to close once they have a head start.

Tacit knowledge becomes capturable in a way it has never been before. The expertise that used to walk out the door when an experienced employee retired can now be systematically captured (transcripts, decision logs, structured Q&A) in a form previously unavailable. The organisations that do this well build something competitors structurally cannot replicate.

This is the agentic asset class. It is the data the next decade of competition will actually run on. The strategic decisions about it are being taken now, mostly by default.

Defensive play

In an environment where your data is scarce, valuable and contested, defensive strategy has two parts.

The first is protection. Most organisations have not done an honest audit of what data they are currently making available to scrapers, vendors, AI integrations, partner APIs and the dozens of SaaS products their employees are pasting things into. Some of this is fine. Some of it is leaking exactly the proprietary information that constitutes the company's edge. The question every leadership team should be able to answer crisply: what data, specifically, would we be in trouble if a competitor got hold of, and what is currently preventing them from doing so?

The second is monetisation through control. Reddit's $60-million-a-year arrangement with Google did not happen because Reddit was generous. It happened because Reddit blocked unauthorised access first and then negotiated from strength. The same is true of News Corp, the Financial Times, Axel Springer, the Associated Press, and a growing list of others. The default state of the internet (free scraping) is ending. Organisations that have rights to valuable content are increasingly converting that into recurring revenue, on their terms.

Defensive play is the less glamorous half of data strategy. It is also the half that determines whether you continue to own the asset, or watch others extract it for free.

Offensive play

Offensive play is where the asset compounds.

There are four moves.

Manufacture data your competitors structurally cannot reproduce. This is the Clubcard play. Tesco overtook Sainsbury's as UK market leader within weeks of the national launch of Clubcard on 13 February 1995, with a programme that did one thing brilliantly: convert anonymous shopping baskets into customer-identified, longitudinal behavioural data. The discount was the inducement; the data was the product. Lord MacLaurin, Tesco's then-chairman, said it most plainly to dunnhumby's team after their first analysis:

What scares me about this is that you know more about my customers after three months than I know after 30 years. — Lord MacLaurin, then chairman of Tesco, on the dunnhumby findings

The data did not exist before Tesco built the system to manufacture it. That is the offensive move, and it remains the cleanest single example.

Capture tacit knowledge before it walks out the door. The expertise of your most experienced employees is, today, mostly stored in their heads. When they leave it leaves with them. Agents change this. Their interactions with experts (structured Q&A, decision-explanation, walkthroughs of edge cases) make tacit knowledge capturable in a form that previously did not exist. Done systematically, this is how organisations build internal capabilities that compound with every year of operation. Done not at all, it is the most expensive talent attrition you do not have a line item for.

Architect for the long game. Bloomberg L.P. is not Bloomberg because it sells better financial data than its competitors. Prices, holdings, filings and regulatory disclosures are widely available; LSEG (which acquired Refinitiv in 2021), FactSet and S&P Capital IQ have similar raw inputs. Bloomberg won by manufacturing the interpretation layer (the schema, the tickers, the workflow, the chat function that became the social network of finance) and refusing to unbundle it from the Terminal. Over four decades later, the Terminal still costs around $32,000 a year per seat, and finance still pays. Bloomberg had a data strategy. Competitors had data.

Design the agent-data flywheel deliberately. Every agent interaction is a new data point. The structured residue of agents operating inside your business (what worked, what failed, what users overrode) becomes the training material for the next generation of agents. Organisations that design this loop intentionally build compounding capability; organisations that do not simply produce expensive transcripts no one reads.

Three of these four moves are accessible to any organisation now. The fourth, architecture, takes longer and costs more. All four get harder the longer they are deferred.

The frontier lesson, in real time

The most expensive ongoing lesson in data strategy is unfolding in plain sight. The major AI labs built their early models on the assumption that internet data would always be available, free and legally uncontested. None of those assumptions has held. The result is a wave of copyright litigation, increasingly expensive licensing deals, an industrialising annotator economy (Scale AI, Surge AI, Mercor and others) and a growing strategic advantage to firms that anticipated this. The bills are arriving now, and they are large.

The lesson generalises. The data acquisition strategy of 2020 was a strategic decision. Few of the people making it at the time were aware they were making one. The decisions being made about agentic data infrastructure in 2026 are of the same character. Some of them will look obvious in retrospect. They are not obvious now.

Bad data strategy in the agentic era

The failure modes are easy to spot once you know what to look for.

The "data lake" that became a data swamp: a costly pool of half-structured information no strategic question shaped and no team owns. The cost is real; the value is theoretical.

The AI initiative with a budget line, a steering committee and no plan for the data the agents will need. The agents arrive on schedule. They are useless on schedule.

The unintentional give-away: proprietary information leaking through vendor agreements, default crawler permissions, SaaS tools and integrations no one mapped. Many organisations are simultaneously paying to acquire data and giving away their own for free.

And the newest pattern: companies whose AI roadmap quietly assumes that data acquisition in 2028 will be as free and uncomplicated as it was in 2020. It will not be. The competitors who were paying attention have already started buying or building.

The shared signature: none of these involve a hard choice. Data accumulation without a question is procrastination. Data investment without architecture is theatre. Strategy is where the trade-offs live.

Coda

The agentic era has produced a new asset class. Most organisations are not yet treating their data as one.

The strategic choices are three. Play it offensively (manufacture data others cannot reproduce, capture tacit knowledge before it leaves, architect for the long game, design the agent-data flywheel). Play it defensively (audit what you are leaking, decide what you license and on what terms, build the legal and technical protections). Or play it neither way, and discover over the next decade that you have slowly lost something you did not know you owned.

The previous edition argued that oil is not really fuel. This one argues something simpler. In the agentic era, data is the asset. The strategic question is not what data is. It is how you intend to play it.

Sources & further reading

1. Pablo Villalobos et al., "Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data", Epoch AI, June 2024.

2. Ilia Shumailov et al., "AI models collapse when trained on recursively generated data", Nature 631, 755–759, July 2024.

3. The New York Times Company v. Microsoft Corporation et al., S.D.N.Y., filed 27 December 2023.

4. Jaron Lanier, Who Owns the Future?, Simon & Schuster, 2013.

5. Lisa Gitelman (ed.), "Raw Data" Is an Oxymoron, MIT Press, 2013.

6. The Economist, "The world's most valuable resource is no longer oil, but data", 6 May 2017.

7. Clive Humby, ANA Senior Marketer's Summit, 2006 (origin of the "data is the new oil" formulation).