AI in DeFi: what DeFAI does with real money, and the controls that keep people in charge

Product design

Imran KhanNovember 20, 202616 min read

Illustration of an agent control panel with a limit slider, approve and stop buttons and an activity log, beside cards labelled AI in DeFi and limits

Search "AI in DeFi" and you'll meet a new word: DeFAI. It's DeFi plus AI, and it covers everything from a chat box that explains a lending market to an agent that moves your stablecoins between pools while you sleep.

The second kind is where the money is, and where the trouble is. A tool that suggests a trade is something you can ignore. A tool that signs the trade is handling your money, and in DeFi a signed transaction usually can't be taken back.

This guide sorts out what AI really does in DeFi today, what's hype, how it has gone wrong already, and what a designer has to put on screen so a person stays in charge. We're a design studio, so we look at the controls and the screens more than the models. Every outside fact is linked to the page it came from, read on 20 November 2026 (last updated 20 November 2026). Much of the controls material comes from one guide, Alchemy's, and Alchemy ships tooling for building agents, so we credit it each time and add the screens and permissions it leaves out. None of this is investment advice, and we don't cover AI tokens as investments.

The short version

  • DeFAI is mostly agents and assistants sitting on top of ordinary DeFi. The protocols underneath (swaps, lending, staking, voting) are the same ones you already know.
  • There are three levels. An assistant that explains, a copilot that prepares a transaction for you to approve, and an agent that signs by itself within limits.
  • The risk follows the signature. A wrong explanation is cheap. A wrong signature isn't.
  • The safe pattern is the same everywhere. The AI proposes. A separate check, enforced by the wallet and not by the prompt, decides whether it can go through.
  • Two real incidents show where it breaks. In one, a clever message talked an agent out of a prize pool. In the other, per the developer, an attacker got into an agent's dashboard.
  • Design the permission, not just the chat. What the agent may touch, how much, for how long, and how to stop it should all be visible, in plain words.

What DeFAI means

"DeFAI" has no settled definition. BNB Chain's own blog says as much: the definition will keep evolving. Most guides use it for AI agents, copilots and automation that help people trade, lend, rebalance or manage assets across DeFi. Alchemy puts the sharp line at execution: "A DeFi AI agent is software that reasons, signs, and settles onchain."

That gives us three levels worth keeping apart.

LevelWhat it doesWho signs
AssistantExplains, summarises, answers questionsYou
CopilotFinds a route or a strategy and prepares the transactionYou, after a preview
AgentDecides and acts within limitsThe agent, under a permission you gave

There's a fourth word you'll trip over: intents. On CoW Protocol, a user signs "an intent to trade", which its docs describe as "not an executable user transaction" but "a signed expression of trading constraints", and solvers compete to fill it. BNB Chain draws a line between intent-centric chains, which carry out a goal you set, and DeFAI, where the AI suggests and you keep control of execution. In practice the two blur. An agent that turns "put my idle USDC to work" into a signed intent is both.

Why DeFi is a harder place for AI than most

Lots of products add AI. DeFi is unusual for four reasons, and each one changes the design.

  • Actions move money. Alchemy's guide says it plainly: "The difference in DeFi is that the action can move money."
  • Transactions are usually final. There's no support desk to reverse one. That's our reading of how the chains work, and it's why the preview step matters.
  • Everything is composable. One instruction can touch a swap, a bridge and a lending market in a row, so one mistake can travel.
  • Anyone can talk to the agent. If the agent reads chat messages, posts or token names, so can an attacker.

What AI actually does in DeFi, job by job

Swaps and routing

The most visible job is the chat box. Coin Bureau gives the example of telling an assistant to "Stake my USDC in the highest-yielding pool with the lowest risk". Alchemy's version is "swap 500 USDC to SOL on the cheapest route and stake it", and it notes how much is hidden behind that sentence: the agent has to find a route, cap slippage, check balances, simulate the transaction and ask for approval when the request goes past policy. "The user sees one sentence. The system sees a workflow."

That hidden workflow is what the interface has to show back. We come to it below.

Yield and rebalancing

The job most guides list first. An agent watches lending markets and moves funds when another one pays enough more to cover the cost of moving. Alchemy describes a stablecoin agent that estimates net returns after gas and bridge costs, and a lending agent that works inside a protocol such as Morpho. Ledger adds the factors an agent weighs: APY, gas fees and risks like impermanent loss.

The safety model in Alchemy's guide is worth repeating. It isn't "let the agent chase the highest number". It's "let the agent move only between pools an expert has already approved".

Lending and borrowing

Here a number matters more than any other: the health factor. Aave's docs say positions are "always over-collateralized", that risk is tracked with a Health Factor, and that below 1 a position becomes eligible for liquidation, when a liquidator repays part of the debt and takes collateral at a discount.

An agent that manages a loan has one clear job: keep that number above a floor you chose, and tell you when it acts. That's an easy rule to show on a screen. It's also a good test of whether you can trust an agent at all, because you can check it.

You'll also see "AI credit scoring" on lists of uses. Coinbound's guide says agents can assess wallet behaviour, collateral ratios and repayment patterns to produce decentralised credit scores, enabling undercollateralised lending, though it admits the approach faces problems with data quality, privacy and regulation. We found no live example in the sources we read, so we'd treat it as an idea.

Governance and treasuries

Ledger says agents can analyse DAO proposals and even vote for a user on criteria set in advance. Coinbound calls the trend hybrid governance, with human and AI oversight together. Alchemy's guide calls treasury rebalancing "the hardest job to hand an agent because the amounts can be large and the downside can be immediate", and suggests routine moves can be automatic while anything large or unusual needs a person.

Risk checks and contract monitoring

The least controversial job. Ledger describes AI monitoring transactions and contracts for anomalies and raising an alert, and BNB Chain describes AI fraud prevention that detects unusual activity. An alert doesn't sign anything, so a wrong one costs a little attention and not funds. It's also the same technology cutting the other way. Anthropic's research tested AI agents on 2,849 recently deployed contracts in simulation and found two zero-day flaws, with exploits worth $3,694 in total, one agent doing it at an API cost of $3,476. Only simulators were used, and the authors argue for adopting AI for defence. If agents can find holes that cheaply, monitoring by agents stops being optional for big protocols.

Staking, and agents you own

BNB Chain's blog describes "smart staking", where AI evaluates validator performance to help a user pick one, and goal-based trading toward a target you set. Both are a recommendation followed by a choice about who acts on it. A few projects go further and let people run agents they own. Olas says its Pearl app lets you control your funds while your agents work, and its site shows 21,193,122 on-chain transactions by its agents as of 12 November 2026, which it calls the last confirmed value because the live source was unavailable. That's the company's own number, and we haven't checked it independently.

Names you'll see

We didn't test any of these, and we don't rate them. They're the ones the guides above name, with what each guide says. Coin Bureau lists Orbit (multi-agent DeFi tasks across chains), AIXBT (market intelligence from on-chain and social data), HeyAnon (a natural-language "super app" for DeFi actions) and Bittensor (a network that rewards machine learning models, which Coin Bureau says DeFAI protocols could draw on). Ledger lists Virtuals Protocol, ChainGPT, Bankr, Kvants and Defi App. Alchemy names Bankr as an agent you instruct from Telegram, Discord or X. On the wallet side, Coinbase's agent docs describe a Wallet MCP that lets an agent interact with DeFi, check balances, send funds and swap tokens, using individual approvals or a dedicated agent wallet with spending controls. Read each product's own page and permission screen before you trust it. That's the point of the rest of this guide.

What's real and what's hype

A fair question for any DeFAI product is whether the AI does anything on chain. Coin Bureau is blunt that many so-called AI agent tokens are "nothing more than memecoins", with AI features as basic as posting on social media or running a chatbot with no on-chain use. Ledger lists honest limits too: it's fragmented (an agent working on Ethereum can't act on Solana events), the data can be poor, the models are hard to explain, people don't yet trust them with funds, and similar strategies could cause herd behaviour.

Our rule of thumb for judging a DeFAI pitch:

  • Can you name the on-chain action the AI performs, and who signs it?
  • Can you see its limits before you give it anything?
  • Can you stop it, and can you see what it did?

If the answer to the first is "it posts to X", it's marketing.

How it has already gone wrong

Two incidents, both public, teach different lessons.

Freysa (November 2024). Freysa was a game built around an AI agent told never to release a prize pool. After 482 attempts by 195 players, one of them won about $47,000 in ETH. The Decoder reports that the winner pretended to have admin access and redefined what the agent's "approveTransfer" function meant, then announced a fake $100 deposit, so the agent thought it was handling an incoming payment and sent its whole balance. Alchemy's guide uses the story as its warning about prompt injection. Freysa wasn't a DeFi protocol, but it shows the pattern: text from an untrusted source got to the part of the system that moves money.

AIXBT (March 2025). Decrypt reported that 55.5 ETH, around $105,000 at the time, left the AI agent AIXBT's Simulacrum wallet after someone queued malicious replies into its autonomous response system. The developer said it was a dashboard access hack and "not a result of agent manipulation", and moved servers and swapped access keys. We have only that account and the reporting of it. Alchemy puts the lesson neatly: "The AI did not fail. The locks around it did."

The research side agrees. A research paper on arXiv describes "context manipulation", an attack on the input channels, memory and external data feeds of agents that work with financial protocols. Alchemy lists a third failure that needs no attacker: a bad objective. "Maximize yield" can push funds into fragile pools, and "rebalance every hour" can cost more in fees than it earns.

Designing it so people stay in charge

Most of the safety work is engineering: keys held outside the agent, limits enforced by the wallet. Alchemy's rule is a good summary of the engineering: "The model can propose an action. A separate policy check decides whether that action can be signed." Our job as designers is to make that check visible and to make the limits something a person can read and change. Our AI in crypto apps guide covers the general principles; here are the DeFi-specific ones.

1. Make the permission the main screen

In DeFi the unit of trust is the permission, more than the chat. Coinbase's Spend Permissions let you name a trusted spender and set limits by token, time period and amount, and the docs list "agentic payments", "algorithmic trading" and "dollar-cost averaging" among the uses. The draft standard ERC-7715 describes a way for an app to ask a wallet for permission to act for a user, with rules such as an expiry, and a call to revoke. Its security notes say apps "should only request the permissions they need, with a reasonable expiration time", and that wallets "MUST clearly display the permissions to users".

Show it as a sentence first ("This agent may move up to 500 USDC a week, only between these three markets, until 20 December") and the technical detail second.

On Fiteats, an AI meal planner we designed, everything the AI needed was asked once, as taps with examples, not typed. A mandate for a money agent can work the same way: tap the markets it may use, pick a limit from a short list, choose when it expires.

2. Limits first, then the goal

Alchemy suggests the safest agents turn broad goals into narrow rules: allowed protocols, maximum position size, slippage caps, daily spend limits and escalation thresholds. That's a good order for a screen too. Ask for the limits before the goal, because "make me the best yield" is the instruction most likely to be read the wrong way.

3. Give the health factor a floor

For a lending agent, let the user set the health factor it must hold and say what it will do when it nears the floor: add collateral, repay, or alert. Our DeFi app design guide goes through how to draw the health factor so it means something to a normal person. An agent should use the same scale, so what it reports matches the screen the user already knows.

4. Preview in plain words, then sign

Before anything is signed, say what will change in terms a person can check: what leaves, what arrives, which protocol, the worst case after fees and slippage. Alchemy lists simulating transactions before signing as one of its policy controls. For an agent acting alone, the preview becomes an entry in the log, and for an action over the limit, it becomes an approval request.

5. Show what's coming, not only what happened

An agent that acts on a schedule should show its next moves. A calendar or a queue works, with each future action labelled and a way to cancel it.

This is a stand-in screen from Finsaro, a finance website template we designed, so the events are made up. The pattern is the useful part: future money movements laid out by day, each one a block you could tap to change. An agent's queue should read the same way.

6. Keep a log a human can read

Every action needs a line: what, when, why, and whether it worked. The statuses matter. A "Failed" row is as important as a "Completed" one, and an agent that quietly retries is worse than one that tells you.

Finsaro's dashboard, again with stand-in data, shows the kind of status tags that make a log scannable. On Dino, a group payments app we redesigned, every change to a bill goes into an edit history with who made it and when. An agent should appear in that kind of history as itself, not as the user.

7. Let people undo, remove and stop

Make every part of the mandate removable with one tap: a market, a limit, the whole agent.

Fiteats lets a person delete a single item without regenerating everything, and rebuild a single meal without losing the week. The smallest change should be the easiest. Add one big "pause the agent" button that works at once and says what it stops. Revoke should be one action, not a settings hunt, and both ERC-7715 and the Coinbase docs describe a way to revoke.

8. Show the evidence, not a verdict

If an agent says a pool is "risky", the reason should be one tap away. On Reputable, a Web3 reputation product we designed, every figure is one click from the commits and protocols behind it. A claim that comes with its working is one a person can check.

9. Say when the agent can't act

When the agent can't act, because a limit was hit, a price moved or a check failed, it should stop and say so, in words. Design that state with the same care as the success one.

A one-page mandate checklist

Our view, not a standard. Before any agent gets access to funds, write down:

  • The job, in one sentence.
  • The tokens it may touch.
  • The protocols and contracts it may call, as a short allowed list.
  • The most it may spend per action and per day.
  • The date the permission expires.
  • The health factor floor, for anything that borrows.
  • What needs a human to approve.
  • Where the key lives, and that it is not inside the agent.
  • How to pause it and how to revoke it.
  • Where the log is.

If any line is blank, the agent isn't ready.

Which jobs to hand to an agent first

Our view. The order follows how bad a mistake is and how easily you can check the work.

JobCost of a mistakeOur advice
Explain a position, watch for alertsLowStart here
Prepare a swap or a deposit for approvalMediumNext, with a clear preview
Rebalance between approved pools within limitsMedium to highOnly with limits enforced by the wallet
Manage a loan against a health factor floorHighOnly after the floor and the alerts are proven
Move a treasury or vote on large proposalsHighestPeople approve, the agent prepares

What to ask a team before you hand over funds

  • Who holds the key, and can the agent ever see it?
  • Are the limits enforced by the wallet or only written in the prompt?
  • What can the agent read, and could a stranger put text in front of it?
  • What happens if the AI service goes down in the middle of a move?
  • Can I see every action, including the failed ones?
  • Which contracts is it allowed to call, and who chose them?
  • How do I revoke access, and how fast does it work?

What we've done and what we haven't

We haven't designed a DeFi protocol or a DeFi agent, so the pictures here are from other work, and we'd rather show real work than draw fake DeFi screens. What we have done is Fiteats, an AI planner where the user changes the plan in a sentence, Dino, a money app where every balance had to say whose it was, Reputable, a Web3 product where every number leads to its evidence, and Finsaro, a finance dashboard and website template with roles, approval rules and an audit trail among its features. Those are the habits we'd bring to an agent that handles money. If you need a studio that has shipped a live DeFi agent, ask for that exact work, ours included. Our web3 UX design guide has more on the wallet and onboarding side.

FAQ

It's the mix of DeFi and AI: assistants, copilots and agents that help people trade, lend, stake or vote across DeFi protocols. There's no settled definition, and BNB Chain says its own will keep evolving.

Mostly for natural-language commands, swap and route finding, yield and rebalancing, loan monitoring, governance analysis and risk alerts. The common thread is that the AI proposes and something else, ideally the wallet, decides whether it goes through.

Alchemy describes them as software that reasons, signs and settles onchain under predefined policies. A classic bot follows fixed rules. An agent reads context and chooses, which is why it needs limits a person can see.

They can be, within narrow limits. The two public incidents we cite came from untrusted text reaching the agent and, per its developer, from access to an agent's dashboard. Keys outside the agent, limits enforced by the wallet and a human approval for large moves reduce both.

Yes, if you give it a permission to. Coinbase's agent docs describe both: its Coinbase MCP page says an agent can operate autonomously or with your approval, and you choose, and its Wallet MCP offers individual approvals or a dedicated agent wallet with spending controls. We'd start with approval.

Coin Bureau's test is whether the AI does anything on chain. If the "AI" mainly posts to social media or runs a chatbot, it's branding.

Some guides say so. We found no live example in what we read.

We can design the permission screens, the log and the states, and prototype them. We haven't designed a live DeFi agent, so ask any studio for the exact work you need.


If you're planning an AI feature that touches money and want the permission screen designed and tested before anything is built, book a call. Our UI/UX design services cover the flows, the states and the design system your developers build from.

About the author

Imran Khan, Founder & Creative Head

15 years of industry experienceFormer Head of Design at Day TranslationsFormer Creative Lead at SupersideWork at Superside for brands including Meta, Red Bull and Ray-Ban

Find him on-

Imran Khan, founder of Usarion

Book a call

Pick a time that suits you

Get in touch

Let’s build something amazing together

Project details

or write directly to hello@usarion.com

We use your details only to reply to you. Privacy policy

For product, interface and web-system work that needs clear direction and careful execution.

// CookiesWe use Google Analytics and Microsoft Clarity to see which pages help people. They only run if you say yes. Privacy policy