How Traditional Finance and DeFi Will Integrate Over the Next Five Years

Listen to this article

How Traditional Finance and DeFi Will Integrate Over the Next Five Years

0:00

The argument, in ten slides

Save it, or send it to whoever owns digital assets and AI risk in your organisation.

Swipe to preview · 10 slidesDownload the PDF

For most of the last decade, bank boards treated decentralised finance as somebody else’s market: volatile, lightly regulated and easy to ignore. That position stopped being tenable in 2026. The largest US settlement utility now runs tokenised collateral in production, the SEC has allowed tokenised equities to trade through liquidity pools, and MAS is legislating its stablecoin framework. This post sets out where the two systems stand today, how I expect them to integrate by 2031, how AI changes the risk on every layer, and what I would ask boards and regulators to do now.

Where the two systems stand in 2026

The building blocks are already live, and several arrived in the last four months.

Sources listed at the end of the post. Figures are as published or as read from the data provider on 26 September 2026.

Two things stand out. The first is that the institutions leading this move are incumbents: DTCC, JPMorgan, BlackRock, Franklin Templeton, a consortium of European banks including ING, UniCredit and BNP Paribas building a euro stablecoin, and the US bank owners of Zelle’s operator, which is extending the network across borders using stablecoins. The second is that the market is cooling even as the plumbing is laid. Stablecoin supply has grown only about six per cent in a year, tokenised Treasuries fell almost seven per cent in the thirty days to this week, and DeFi deposits have not recovered from the spring exploits. The infrastructure is running ahead of demand. That gives boards time to prepare. It does not remove the need to.

Three layers where the systems meet

Integration is easier to reason about when it is split into three layers, each moving at a different speed.

Money. Regulated stablecoins and tokenised bank deposits become the cash leg of on-chain transactions. The BIS argued in 2025 that stablecoins fail its tests of singleness, elasticity and integrity, and proposed a “unified ledger” holding central bank reserves, commercial bank deposits and government bonds together. MAS takes a similar line, favouring tokenised bank liabilities. I expect both forms to coexist: deposit tokens for wholesale and inter-bank settlement, regulated stablecoins for payments, cross-border transfers and machine-to-machine commerce.

Assets. Treasuries, money market funds and listed equities move onto shared ledgers, first as collateral and then as tradable instruments. DTCC’s first live use cases were collateral pledges, repo delivery-versus-payment and margin at central counterparties. That sequence matters. Collateral mobility is where tokenisation saves money today, because a tokenised Treasury can move between counterparties around the clock and settle in minutes.

Market mechanisms. This is the layer where DeFi contributes most. Automated market makers, liquidity pools, over-collateralised lending and composable smart contracts are the ideas that make DeFi distinct. The SEC exemption lets regulated venues use them for tokenised stocks, with access standards and caps. BlackRock made its BUIDL fund tradable on UniswapX for whitelisted investors in February. Aave built a separate institutional market for real-world assets. The pattern is consistent: the mechanism is public and programmable, and the participants are identified and permissioned.

The next five years

The dates below are my forecast, built on the regulatory calendar and the pilots already announced.

The size forecasts vary widely. Citi projects $1.9 trillion of stablecoins by 2030 in its base case. Standard Chartered and the US Treasury Borrowing Advisory Committee both put the figure near $2 trillion by 2028. McKinsey expects about $2 trillion of tokenised assets by 2030, within a range of $1 trillion to $4 trillion, while the older BCG and ADDX estimate is $16 trillion. Each of these needs the market to grow six- to eightfold from its current, nearly flat base. I would plan for the lower end of every range and treat the infrastructure dates as firmer than the volume forecasts.

How AI changes the picture

AI arrives in tokenised finance in four roles at once. Each one changes a different part of the risk.

AI agents as customers

Agent payment protocols have multiplied in the past eighteen months. Coinbase launched x402 in May 2025 so that software can pay for a service within a single web request, usually in USDC. Google’s Agent Payments Protocol, announced in September 2025 with more than 60 partners, uses signed “intent” and “cart” mandates that give a non-repudiable record of what a user authorised. Visa, Mastercard and the Stripe-backed Tempo network have each launched their own frameworks. In June 2026 Mastercard extended its Agent Pay product to cover stablecoins, recording agent permissions on public blockchains.

Stablecoins suit machines because they settle in seconds, need no merchant account and can carry payments of a fraction of a cent, which card minimums make uneconomic. The volumes are still small and often inflated. Chainalysis counted well over 100 million cumulative x402 transactions by the first quarter of 2026, much of it driven by meme-coin farming, and CoinDesk, citing Artemis data, put real payment value in March at about $28,000 a day. The protocols are ready before the demand. That is the right order for building controls.

AI as attacker

In December 2025 Anthropic published SCONE-bench, a test of AI agents against 405 smart contracts that had been exploited in the real world between 2020 and 2025. The agents produced working exploits for 207 of them, worth $550m in simulated proceeds. Against contracts exploited after the models’ training cut-off, three models produced exploits worth $4.6m. A scan of 2,849 newly deployed contracts found two previously unknown flaws, at a cost of about $1.22 per contract. Simulated exploit revenue doubled roughly every 1.3 months over the year. OpenAI and Paradigm’s EVMbench, published in February 2026, found the same trend, with the strongest model’s exploit success rising from 32 to 72 per cent in a single model generation.

The FATF’s July 2026 update warns that AI coding assistants and autonomous agents may help attackers find flaws, write exploit code and speed up attacks. None of the 2025 or 2026 hack records I checked attributes a major real-world theft to an AI agent yet. Measured AI harm so far sits in fraud. Chainalysis found that scam operations with on-chain links to AI vendors earned $3.2m each, against $719,000 for those without, and the FBI’s 2025 report logged more than 22,000 AI-related complaints with losses above $893m. The capability curve in the lab is steep enough that I would not wait for the first large AI-driven exploit before acting.

AI agents as targets

An AI agent that holds funds is a new kind of account: one that can be talked into a transfer. Princeton researchers showed in 2025 that planting false memories in an agent built on the ElizaOS framework could trigger unauthorised crypto transfers, and that the models they tested were more vulnerable to memory injection than to direct prompt injection. The incidents have followed. In May 2026 a prompt injection hidden in Morse code in a post on X led an AI agent on the Base network to transfer tokens worth an estimated $150,000 to $200,000, according to the OECD AI Incidents Monitor. The pattern resembles social engineering against a junior employee with signing authority, repeated at machine speed and at scale.

AI as supervisor

The same technology is becoming the only practical way to supervise a market that settles in seconds. In the BIS and Bank of England’s Project Hertha, analytics across a payment system helped banks find 12 per cent more illicit accounts, and 26 per cent more where the behaviour had not been seen before. Earlier, the BIS Project Aurora found graph-based machine learning caught about twice as many money-laundering networks as rules. Programmable compliance, where eligibility and transfer rules are written into the asset’s own code, is part of MAS’s BLOOM initiative and the Global Layer One work. The FSB’s June 2026 consultation on AI in finance goes further, noting that real-time human monitoring of agent decisions becomes impractical as their use grows and that monitoring may itself need another AI agent.

Where the risk concentrates

Three features of tokenised finance interact badly with AI.

Atomic settlement removes the buffer. In today’s markets, a trade agreed on Monday settles a day or two later. That gap is inefficient, and it is also the time in which operations staff, risk teams and counterparties catch errors, reverse fraudulent instructions and stop a runaway algorithm. When settlement is atomic, the transfer and the payment are final in the same instant. An AI agent that misreads its mandate, or one that has been manipulated, completes the damage before any person sees it. The controls have to sit inside the code and the agent’s permissions, because there is no later step in which to apply them.

Composability spreads failure. DeFi protocols stack on one another. When KelpDAO lost about $292m in April 2026, the stolen tokens were posted as collateral on Aave, leaving an estimated $123m to $230m of bad debt in a separate protocol. AI agents that rebalance across pools and lenders will move that contagion faster, because they react to the same prices in the same way. The IMF warned in 2024 that AI trading models responding similarly to a shock could raise market speed and volatility under stress, and its 2026 note on tokenised finance says stress events in tokenised markets move faster.

Concentration hides in the dependencies. A tokenised market depends on a small number of stablecoin issuers, blockchain networks, price oracles, cross-chain bridges and, increasingly, AI model providers. The FSB lists third-party concentration as the first of its AI vulnerabilities. In the TRM Labs count for the first half of 2026, infrastructure compromises made up 15 per cent of hacks but about 76 per cent of losses.

What to do this year

None of this requires a firm to issue its own stablecoin or join a DeFi protocol. It requires a decision on each layer and controls that match the speed of the new settlement model.

1. Decide your position on each layer. For money, assets and market mechanisms, record whether the firm will issue, use, accept as collateral, or stay out, and review it yearly. Staying out is a legitimate choice, as long as it is a decision.
2. Give every AI agent an identity, a mandate and a limit. An agent that can move value needs a named owner, a signed statement of what it may do, spending and counterparty limits enforced outside the model, and a log that ties each transaction to a mandate. The IMF’s 2026 note on agentic payments puts the principle well: keep the probabilistic reasoning upstream and the authorisation and settlement deterministic.
3. Test smart contracts with the tools attackers use. Run AI-assisted review before deployment and again whenever a stronger model is released, because a contract that passed last quarter may fail against this quarter’s model. Budget for it as recurring cost.
4. Build circuit breakers into the code. Transfer caps, time delays on large movements, pause functions with clear authority to use them, and rate limits on agent activity replace the settlement buffer that atomic settlement removes.
5. Map the dependencies. List every stablecoin issuer, network, oracle, bridge, custodian and AI model provider your tokenised activity relies on, and set concentration limits the way you would for any critical outsourcing.
6. Upgrade financial crime controls for stablecoins and synthetic identities. The FATF reported that stablecoins made up most of illicit virtual-asset volume in 2025, and deepfakes are defeating remote onboarding. Blockchain analytics and liveness checks belong in the same control set as card and wire monitoring.
7. Build one team across treasury, risk, technology and legal. Tokenised finance falls between existing functions. The firms moving fastest have a single accountable owner and a standing group that meets on it.

For regulators and government agencies, the parallel priorities are agent identity standards (the IMF calls this “Know Your Agent”), supervisory access to on-chain data in real time, and minimum circuit-breaker requirements for tokenised venues. MAS’s Project Guardian and BLOOM give Singapore a head start on testing these before they are needed at scale.

Questions for the board

Before the next strategy review, I would want written answers to five questions:

1. What is our position on stablecoins, tokenised assets and DeFi market mechanisms, and who owns each decision?
2. Which of our AI agents can move money or assets today, and what limits sit outside the model?
3. How are our smart contracts and those of our counterparties tested against AI-driven attacks, and how often?
4. What stops an erroneous or manipulated transaction when settlement is final in seconds?
5. Which stablecoin issuers, networks, oracles and AI providers are we dependent on, and what happens if one fails?

On the cryptography under digital assets

What Quantum Computing Means for the Encryption Behind Your AI

Every stablecoin, deposit token and agent credential rests on the signature schemes a quantum computer would break.

On MAS expectations for banks

How Financial Institutions Should Use AI to Keep Up With Regulatory Change

Stablecoin, tokenisation and AI rules are arriving at the same time, and someone has to track all three.

On board oversight of technology risk

What the DBS Outages Teach Boards About Technology Risk

Why oversight for a large infrastructure change has to be in place before the change starts.

On supervising agents at machine speed

From Human-in-the-Loop to AI-on-the-Loop: Redesigning Oversight Architectures

The oversight model an organisation needs once settlement is too fast for a person to approve each step.

My thanks to Dr. Christina Chua for her contribution to this post. Her perspective from finance and banking shaped the argument, and the post is better for it.

If your organisation is deciding how far to go on stablecoins, tokenised assets or AI agents that move money, the position-by-layer decision in step one is where I usually start. My consulting work covers AI and technology risk governance for boards of regulated firms and public agencies.