Pillar
AI and blockchain: where the two fields actually meet
For years these were two separate hype cycles that borrowed each other's vocabulary. They now have a genuine intersection, and it is narrower and more specific than either community claims: identity, payment, provenance and proof.
Start with the honest version
Most claimed combinations of AI and blockchain do not survive contact with a straight question: what does the ledger do here that a database could not? Training on-chain is absurdly expensive. Inference on-chain is worse. "Decentralised AI" as a marketplace for GPU time is a real business but has nothing to do with the ledger beyond payment. Tokens attached to model access are usually a funding mechanism wearing a technical costume.
What remains after that filter is a short list - and everything on it shares one property: a claim has to be believed by someone who has no reason to trust whoever is making it. That is the only problem a ledger is uniquely good at, and it turns out autonomous agents generate it constantly.
Ask: who is the untrusting party, and what would they otherwise have to take on faith? If the answer is "nobody, we control both sides", a database is correct and cheaper. If the answer names a counterparty, a regulator or a future court, the calculus changes.
Four intersections that hold up
1. Identity for non-human actors
An agent needs an identifier its operator cannot silently swap, and a way to prove whose authority it carries. Decentralised identifiers give a subject a persistent identifier that no single registrar controls; verifiable credentials let a party assert something about that subject in a form anyone can check offline. Neither strictly requires a chain - but an anchor that no participant can rewrite is exactly what stops a compromised issuer from rewriting history. The full mechanics.
2. Payments at machine granularity
Agents transact in amounts and at frequencies conventional rails handle badly: fractions of a cent, thousands of times a day, between parties with no prior relationship. Stablecoin settlement and HTTP-native payment schemes both address this, and the second is notable for treating payment as a property of a request rather than a separate checkout flow. What works today.
3. Provenance and content authenticity
When most media can be generated, the useful question stops being "is this fake" and becomes "where did this come from, and who stands behind that claim". Content credentials - signed manifests travelling with a file, recording capture device, edits and generation steps - are the mechanism. A ledger is not required to sign a manifest, but it is a good place to anchor revocation and to prove a signature existed before a given moment. Why detection loses and provenance scales.
4. Verifiable execution
The hardest and least mature of the four. If an agent claims it ran a particular model, on a particular input, under a particular policy, can it prove that to someone who was not there? Approaches range from trusted execution environments producing signed attestations to zero-knowledge proofs of inference - the former practical and dependent on trusting a hardware vendor, the latter elegant and, for models of interesting size, still far too expensive. This is where the field's most serious work is happening and where nobody should claim a solved problem.
Smart contracts, read literally
A smart contract is code that executes on agreed rules with an outcome no party can unilaterally revise. For agents, that is interesting for one narrow reason: it makes conditional commitment possible without an intermediary. An agent can place funds under a condition - released on delivery, refunded on timeout - and the counterparty can verify the condition without trusting the agent's operator or a payment processor's dispute team.
This is not a replacement for contract law, and treating it as one has produced most of the failures in the field. It is an escrow mechanism that works between parties with no relationship, which is precisely the situation agents keep creating.
Blockchains spent a decade solving trust between parties who cannot vouch for each other, then waited for a use case with enough of them. Autonomous agents are that use case.
Tokenisation, in the narrow sense that matters here
Tokenisation as an institutional finance topic - securities, custody, regulatory treatment - is a different subject with its own literature. The narrower question relevant to agents is simply: what can a piece of software hold and transfer on its own? A licence, a ticket, a compute allocation, a data-access right. If an entitlement exists only as a row in a vendor's database, an agent can use it but cannot prove it holds it, cannot transfer it, and cannot compose it with anything else. If it exists as a bearer claim, it can.
That is the whole of the argument, and it is enough. The interesting consequence is not financial but structural: entitlements that agents can hold directly are entitlements that can be traded between agents, which is a precondition for autonomous commerce reaching past retail.
What this does not need a chain for
- Model training and inference. Centralised, and correctly so.
- Agent memory and retrieval. A database. Always was.
- Tool access. An interface description and an auth token.
- Coordination between an organisation's own agents. Internal infrastructure; there is no untrusting party.
Keeping this list honest is what separates a useful analysis from a pitch. The intersection is real, it is load-bearing, and it is about a quarter the size the more enthusiastic version claims.
Frequently asked
Do AI systems need a blockchain?
Not for training, inference, memory or tool use - those are ordinary infrastructure problems. A ledger becomes relevant only where a claim must be believed by a party with no reason to trust the claimant: agent identity, machine payments, content provenance and proofs about execution.
What is verifiable AI?
The ability to prove after the fact what a system actually did - which model ran, on what input, under which policy. In practice this means either hardware attestations from a trusted execution environment or cryptographic proofs of inference. The first is deployable and requires trusting a chip vendor; the second is not yet practical at useful model sizes.
What are content credentials?
A signed manifest that travels with a media file recording how it was produced: capture device, editing steps, and whether a generative model was involved. Anyone can verify the signature and see who is standing behind the claim. It answers provenance rather than attempting detection.
Why do agent payments keep pointing at stablecoins?
Because of granularity and counterparty reach, not ideology. Agents need to pay fractions of a cent, thousands of times a day, often to parties they have no relationship with. Card economics make that impossible below a certain ticket size; account-to-account settlement in a stable unit does not.