IBM has published a position arguing that the next wave of tokenization adoption will be driven not by human users but by autonomous AI agents — and that the financial system is structurally unprepared for it. The argument is straightforward: agents need to negotiate, settle, and reconcile transactions in real time, and the existing rails were built around human workflows, business hours, and manual reconciliation. Tokenization, IBM contends, is the only settlement layer that can serve a machine-native economy. The question is whether that thesis holds up under technical scrutiny or whether it collapses into another enterprise blockchain narrative repackaged for the AI cycle.
The Core Argument: Agents Cannot Use Human Rails
IBM’s position rests on a structural mismatch. Traditional financial infrastructure assumes a human initiates a transaction, approves it, and reconciles it after the fact. Settlement takes hours or days. Identity is tied to a person. Operating hours are bounded. An autonomous AI agent breaks every one of those assumptions. It may need to purchase compute, pay for data access, and settle with a counterparty in milliseconds, thousands of times per hour. No card network, no ACH rail, and no existing correspondent banking workflow can accommodate that pattern. IBM argues tokenization solves this because on-chain settlement is atomic, programmable, and operates continuously. The agent holds a wallet, signs a transaction, and the transfer is final in the same block. There is no clearinghouse, no reconciliation lag, and no business-hour dependency.
Identity, Settlement Finality, and the Missing Layers
The more technical part of IBM’s argument concerns what tokenization must additionally provide beyond a payment rail. An AI agent needs an identity that is machine-readable, verifiable, and revocable — not a KYC document tied to a passport. It needs settlement finality that is deterministic rather than probabilistic, which rules out some Layer 2 constructions where reorg risk or sequencer failure creates ambiguity. And it needs programmatic access to liquidity: the ability to draw on a credit facility, swap into a settlement asset, or route through a bridge without human intervention. IBM identifies these as open problems. Stablecoins like USDC and USDT solve the settlement asset question but not the identity or credit layers. Projects like Coinbase Agent Payments and Skyfire are building the wallet and spending-control infrastructure, but none of this is standardized. The gap between what agents need and what exists today is wider than the marketing suggests.
What the Thesis Gets Right and Where It Overreaches
IBM is correct that on-chain settlement is the only existing rail that satisfies the constraints of autonomous machine-to-machine payment. Card networks cannot do sub-second, high-frequency micropayments. ACH cannot operate outside business hours. SWIFT cannot settle atomically. These are facts. Where the argument overreaches is in assuming that tokenization alone closes the gap. The harder problems are not settlement but trust, dispute resolution, and accountability. If an agent overspends, who is liable? If an agent enters a bad transaction, can it be reversed? On-chain finality cuts both ways: it removes reconciliation friction but also removes the safety nets that traditional finance provides. IBM gestures at these questions but does not answer them. The projects actually building agentic payment infrastructure — x402, Skyfire, Payman — are grappling with these issues directly, and their solutions involve spending limits, policy engines, and escrow mechanisms layered on top of raw settlement.