AI agents are beginning to take on defined tasks across treasury, wealth management, and onchain finance. The models are becoming increasingly capable, and the financial opportunity is becoming clear. The main constraint on adoption is now the data infrastructure these agents depend on.
The Largest Capital Deployment in History Is About to Be Automated
Financial institutions are beginning to automate activities that have historically required people to make and execute individual decisions.
McKinsey projects that agentic commerce could orchestrate $3 trillion to $5 trillion in transactions globally by 2030. Finance executives expect the share of their functions operating with meaningful agentic AI to increase from 20% today to 66% in the same timeframe.
These systems will do more than improve analytics or generate recommendations. Agents can discover yield opportunities, assess risk across thousands of variables, construct transactions, and manage portfolios continuously without requiring a person to approve every step. BCG found that agentic workflows can increase operational capacity by over 50%, while reducing costs by approximately 40%. Nearly half of finance teams surveyed expect to deploy agentic AI in 2026, a 600% increase year over year.
The remaining challenge is execution, and much of that challenge comes down to the quality and structure of the underlying data.
Financial Agents Are Moving Into Operator Roles
The automation of financial decisions is happening across several parts of the financial system at the same time.
In treasury, AI agents are beginning to handle parts of the operational core of institutional finance. They can monitor cash positions in real time, rebalance liquidity across accounts, execute FX hedges within defined policies, and route payments without requiring human approval at every step. These systems increasingly participate in the process itself rather than simply producing information for humans to act on.
In wealth management, agents can continuously monitor portfolio performance against client mandates, identify rebalancing opportunities, generate recommendations in response to changes in a client's circumstances, and automate compliance and audit workflows that have historically required significant analyst time. Parts of the advisory process that depend on repeated monitoring and defined rules can increasingly operate through agents working within governed parameters.
Onchain finance pushes these requirements further. Agents operate across thousands of protocols on dozens of networks, where they can stake assets, lend, provide liquidity, and manage tokenized real-world assets. There is no T+2 settlement window or overnight batch process to catch an error before execution. An action based on incorrect data can reach the chain immediately and, in many cases, cannot be reversed.
These areas of finance share the same underlying requirement. The data an agent uses must be live, normalized, structured for execution, and auditable when the agent acts.
Bad Data Leads to Execution Risks
There is an important difference between a traditional AI system and an autonomous agent.
A conventional machine learning model makes a prediction and stops. If the prediction is wrong, a person can intervene before anything else happens. An autonomous agent can act on its output, and that action can trigger another action with little or no human involvement. As Dataiku describes it, a stale record can produce a bad decision that feeds into subsequent decisions before a person has an opportunity to intervene.
With an agent, bad data can therefore lead directly to bad actions.
The consequences are particularly clear in finance. An agent might construct a transaction using stale market data, deploy capital into a protocol that has paused withdrawals, or calculate net yield without including every applicable fee. In each case, the problem does not end with an inaccurate recommendation. The agent can execute the transaction, and the resulting loss can be immediate and difficult or impossible to reverse.
The Anatomy of Agentic Finance Data
Data that is sufficient for a dashboard is not necessarily sufficient for an agent making and executing financial decisions. Analytical systems generally need accurate historical data. Financial agents need data that is live, normalized, structured for execution, calculable, and auditable.
1. Live
The data must be current when the agent acts.
A yield rate from six hours ago, a price from yesterday's batch process, or a liquidity figure that has not been updated since the previous indexing cycle can change the transaction an agent constructs. Stale data is no longer just an analytics problem once the system can act on it.
2. Normalized
The data must be normalized across the systems the agent compares.
Financial agents operate across systems that use different schemas and calculation conventions. One protocol might express APY as an annualized rate while another uses a weekly compounding rate. If those values are compared without normalization, the resulting calculation can be wrong even when the model itself behaves correctly.
3. Structured for Execution
Agent data needs to support execution rather than simply provide information.
A human analyst can read a table, interpret its contents, apply judgment, and decide what to do next. An agent needs typed, schema-defined outputs with enough metadata to pass information directly into an action or transaction layer.
Much of today's financial data infrastructure was built for the first use case.
4. Calculable and Fully Auditable
Financial agents perform calculations. They compare risk-adjusted returns, estimate yields after fees and gas costs, and monitor variables such as the health factor of borrowing positions.
Those calculations need to use consistent logic and produce reproducible results across different contexts. The system also needs to record the inputs, intermediate calculations, and resulting actions so institutions can inspect how an automated financial decision was made.
McKinsey's research on agentic AI reaches a similar conclusion: agents depend on the quality and accessibility of enterprise data, and organizations need to move from use-case-specific pipelines toward reusable data products.
Onchain Finance Puts Data Infrastructure Under Pressure
Onchain finance puts these requirements under particularly demanding conditions. Thousands of protocols operate across distinct networks. Yield rates can change hourly, fee structures can differ even between versions of the same protocol on different chains, and there is no single authoritative source of truth. There is also no T+2 settlement window or overnight batch process available to correct an error before execution.
This is the environment in which Yield.xyz built its infrastructure, and many of the underlying data problems apply beyond crypto. Onchain finance concentrates the same issues that other parts of finance will face as autonomous agents begin to make more decisions and execute more transactions.
Yield.xyz provides B2B infrastructure for wallets and custodians through a single API that abstracts staking, DeFi lending, liquidity provision, and tokenized real-world assets across onchain markets. It powers yield products for Ledger, Trust Wallet, Trezor, and Anchorage Digital among more than 200 companies. The infrastructure serves 4 million users and covers more than 3,300 opportunities across 80 chains. Yield.xyz is SOC 2 Type II certified and fully non-custodial.
What Yield.xyz Agentkit Built and Why the Schema Matters
Yield.xyz Agentkit designed its infrastructure around the assumption that reliable data is part of the product itself. Its schema supports systems that need to act on financial data rather than simply display it to analysts.
Each layer addresses one of the requirements that autonomous financial systems introduce.
1. The Core Object: Every Yield Opportunity, Normalized
The core object provides a normalized representation of each yield opportunity across more than 3,300 eligible positions. The same structure applies whether the opportunity is Ethereum native staking, an Aave lending pool, a Yearn vault, or a tokenized real-world asset.
The system separates yield by source rather than presenting only a headline APY. Base staking rewards, incentive token emissions, MEV revenue, and points programs are represented separately. An agent can therefore distinguish between a 12% APY generated by structural staking rewards and a 12% APY that depends primarily on temporary liquidity incentives. The data layer resolves that distinction before the model begins reasoning about the opportunity.
The schema also represents timing constraints such as cooldown periods, withdrawal windows, and warmup requirements. An agent considering a position with a 28-day unbonding period receives that information before constructing the transaction.
2. The Lifecycle Object: Capital Tracked Through Its Full Arc
Once capital has been deployed, the data requirement changes from discovering opportunities to managing existing positions.
Many position-tracking systems tell an agent what assets it holds. Yield.xyz also represents the current state of each position, including whether it is active, entering, exiting, withdrawable, claimable, or locked. The system then provides the operations currently available for that position.
Those operations can include claiming rewards, restaking, redelegating, withdrawing, or claiming unstaked assets. The server determines which actions are currently possible, while the agent decides which action to take.
3. The Execution Object: Unsigned, Simulated, and Auditable
When an agent decides to act, Yield.xyz returns a fully constructed, chain-correct unsigned transaction payload. The payload includes human-readable amounts alongside base-unit values, a decoded annotation, and a gas estimate calculated when the transaction is constructed.
Before the transaction reaches the chain, Yield.xyz's Shield library pre-simulates and validates it against a library of known-good operation patterns. The validation can identify issues such as an incorrect contract address, mismatched delegate authority, or a malformed transaction type before execution.
The system also logs each action with timestamped status transitions from creation through onchain confirmation. Institutions can inspect what was constructed, how its status changed, and whether it ultimately reached the chain.
4. The Intelligence Layer: Risk Ratings Agents Can Use
Yield.xyz uses two independent risk methodologies that map into a common schema.
Credora applies a credit-risk framework to lending markets and produces a letter grade and a Probability of Significant Loss (PSL) that is updated daily. An agent or institution can use these fields to enforce requirements such as a minimum credit grade or maximum PSL threshold.
Staking Rewards uses a three-pillar framework covering operations, security, and strategy. Its granular sub-scores provide information that an agent can use when comparing validators.
Both providers return information through the same structure, including the grade, score, sub-assessments, and update timestamp. An agent can therefore apply a common risk framework across lending and staking without implementing separate parsing logic for each provider.
Considerations Before Deploying Financial Agents
For financial institutions considering autonomous agents, the infrastructure test can be reduced to four questions:
- Is the data current when the agent acts?
- Is it normalized across everything the agent compares?
- Are the outputs structured so the agent can execute directly?
- Can every calculation and action be audited?
A system that cannot answer these questions clearly may work for analysis or experimentation, but it is not yet designed for autonomous financial execution.
Data Infrastructure Is Part of the Financial Strategy
Autonomous financial agents need more than capable models. They need live, normalized data that is structured for execution and auditable from decision through transaction.
As more financial activity moves from human-assisted workflows to autonomous systems, that data layer becomes part of the financial infrastructure itself. Institutions that solve the data problem can expand the financial activities that agents can safely manage rather than spending time compensating for unreliable inputs.
About Yield.xyz
Yield.xyz provides a unified API for onchain yield infrastructure and powers Ledger, Trust Wallet, Trezor, Anchorage Digital, and more than 200 institutions serving over 4 million users. Its infrastructure covers more than 3,300 yield opportunities across 80+ networks. Yield.xyz is SOC 2 Type II certified and non-custodial by design.