Everything Will Be Tokenised

Not Just Words on a Page

When people talk about tokens and AI, they usually mean one thing: the chunks of text a language model processes when predicting the next word. The Declaration of Independence is 1,627 tokens. GPT-4 processes roughly 10 trillion tokens a month. Pay per token. Think in tokens. Bill in tokens.

That framing is too narrow, and the organisations that mistake it for the full picture will pay for it.

Tokenisation is the process of rendering the world for computers. When you tokenise data, identity, money, or knowledge, you make what people know and do across human systems (finance, commerce, healthcare, logistics) safely accessible to software, including increasingly autonomous software. Tokens are the DNA of AI. They are its feedstock and its fuel.

Not all tokens are equal. There are seven distinct token categories emerging that will define which organisations win the next decade. Each has a different supply curve, a different competitive dynamic, and a radically different value trajectory. Understanding which ones your business generates, which ones it protects, and which ones it is quietly giving away is the most important strategic question in AI right now.

Seven Tokens That Define the Next Decade

Three broad groupings are worth holding in your head: tokens used for value transfer, tokens used for building expertise, and tokens used for personalisation. Within those three sits the taxonomy that matters in practice.

The Revolut Proof

You can talk about token taxonomy in the abstract for a long time. Revolut's April 2026 research paper, PRAGMA: Revolut Foundation Model, is what it looks like in practice when an organisation commits to its own token inventory rather than borrowing someone else's.

Revolut did not license a foundation model from OpenAI, Anthropic, or Google and fine-tune it on a subset of customer data. They built their own.

PRAGMA is a family of Transformer-based models pre-trained on 26 million user records spanning 111 countries, covering 24 billion behavioural events totalling 207 billion tokens. Those tokens are not words. They are card transactions, app navigation paths, trading activity, communications, account milestones, and profile states: every interaction a Revolut customer has ever had with the platform, encoded and made machine-readable.

The point of PRAGMA is not that Revolut built a big model. The point is what they built it from. Every time a Revolut customer makes a payment, opens the app, or buys a stock, they generate a context token. Aggregated across 50 million customers, those tokens form a representation of financial behaviour that no general-purpose model trained on internet text can replicate. You cannot buy that at an API endpoint.

Standard language models process financial data as text. They tokenise a transaction amount as digit characters. They process a merchant category code as subword fragments. They have no native representation for temporal distance between events, cyclical patterns in behaviour, or the structural relationship between a payment type and its fields. PRAGMA was built to reject this entirely. Its tokenisation scheme is native to banking data: numerical values become percentile buckets, categorical values map to single tokens, and time is encoded both as log-seconds to the last event and as cyclical calendar features. The model understands the difference between a card payment and a transfer not because it has read text about the difference, but because it has processed hundreds of millions of examples of each, structured as the actual events they are.

Revolut is not offering PRAGMA to anyone else. They are not sharing the architecture with competitors. This is a proprietary token factory built from proprietary context tokens, producing expert tokens that compound with every new customer interaction. The moat is not the model. The moat is the 24 billion events that trained it.

The Honest Questions Every Organisation Needs to Ask

Revolut had an unusual head start: a decade of behavioural data from tens of millions of customers, collected on a platform they own. Most organisations are not starting from that position. But every organisation is generating tokens right now, whether they know it or not. The question is whether those tokens are being captured, protected, and compounded, or whether they are flowing quietly to someone else.

What You Already Have That You Do Not Know About

The challenge for most businesses is not that they lack tokens. It is that their tokens are buried in formats machines cannot use, or held in systems that were never designed to generate a compounding advantage from them. The average large enterprise has petabytes of data and a clear AI strategy on the analyst call. In practice, that data sits in poorly structured databases and unstructured documents. The average knowledge worker still performs over a thousand copy-paste actions a week.

The gap between the data a business holds and the tokens it can deploy is where most of the value is being left. Not because the data does not exist, but because nobody has mapped it, structured it, or protected it as a strategic asset.

Why Picking the Best Model Is the Wrong Game

The cost of running large language models has dropped by a factor of one thousand over three years. Access to frontier model capability is approaching commodity pricing. Every organisation can afford GPT-4. Every organisation can afford Claude. The model is no longer the moat.

What Revolut built with PRAGMA demonstrates this cleanly. A standard language model fed Revolut's transaction data would treat it as text: amounts as digit sequences, merchant codes as subword fragments, event timing as a number with no structural meaning. PRAGMA treats the same data as what it is: a structured sequence of financial events with semantic types, value distributions, and temporal relationships. The underlying model architecture matters far less than the decision to build a representation of your data that only you can build.

This does not mean every organisation needs to build a billion-parameter foundation model. It means every organisation needs to understand which of its data assets are genuinely proprietary, which are being captured in a form that compounds, and which are being handed to third-party platforms that distribute the resulting intelligence to the entire market.

The Access Token Layer: Why Verification Is the Starting Point

There is one token category that sits underneath all the others: access tokens. Before context can be shared, before memory can be accessed, before a transaction can be initiated, something has to answer the question of whether the requesting agent is who it claims to be.

This is not a theoretical concern. An agent can claim to represent your brand today, query your systems, and return responses under your name, without any cryptographic proof that it is authorised to do so. As agent-to-agent (A2A) commerce scales from experimental to operational, the absence of a verified agent endpoint will be the single most exploitable gap in the commercial internet.

Fetch.ai's Agentverse and the Almanac registry are built around this problem. Every agent registered on the Almanac carries a cryptographically signed identity, verified on-chain. Before any token exchange happens, before context is shared or a transaction is initiated, both parties confirm with mathematical certainty that they are communicating with a verified, authenticated agent. This is what Know Your Agent (KYA) looks like in practice: not a policy, but a protocol.

In the same way that KYC verification became a compliance requirement as financial fraud scaled, KYA will become a commercial requirement as agentic interactions scale. The organisations that build this infrastructure now will not need to retrofit it when retrofitting means interrupting live commercial flows.

Where to Start

Revolut required a decade of platform data and a dedicated research organisation to build PRAGMA. Most businesses are not starting from that position, and they do not need to be.

The practical starting point is a token audit: an honest inventory of which tokens your organisation is already generating, which are being captured in machine-readable form, which are being given away to third-party platforms, and which represent a genuine proprietary advantage if protected and compounded.

Most organisations that do this exercise find the same three things. They are generating far more context tokens than they have made machine-readable. They are sharing more expert tokens with third-party platforms than they realised. And they have no verified access token infrastructure, which means any agent they deploy, or any agent that claims to represent them, is operating on assumption rather than verification.

The businesses that move from that audit to a strategy, even a narrow one focused on a single high-value token category, will be compounding while their competitors are still evaluating vendor proposals.

Revolut did not start by building a one-billion-parameter model. They started by capturing every banking event in a structured, machine-readable format. Everything else followed from that discipline. The same logic applies whether you are a retailer, a logistics operator, a professional services firm, or a manufacturer. Your context tokens already exist. The question is whether you own them.