"AI swap" is one of the most overused phrases in crypto right now. At Cambio it means something specific — and it deliberately doesn't mean some of the things you might assume.
"AI swap" is one of the most overused phrases in crypto right now. Every product page claims it. Almost none agree on what it means. Some platforms call a chat support widget "AI." Others mean a black-box route picker that promises to "find you the best deal" without showing its work. A growing category of agentic wallets means a software agent that lives inside your wallet, holds session keys, and signs transactions on your behalf.
Cambio is none of those things. This post is the anchor for everything else we publish about how the product works. It defines what "AI" actually does at Cambio, and just as importantly, what it deliberately does not do.
What people usually mean by "AI swap"
Before defining what we do, it helps to name the three common framings — because if you have one of these in your head, the Cambio flow will surprise you.
Framing one: "AI picks the best route"
This is the DEX aggregator framing. Platforms like 1inch and Jupiter split a single swap across many liquidity pools and routing paths to extract a marginally better price. The "AI" framing is mostly marketing — the actual route picker is a deterministic search over a graph of pools, optimized for output amount net of gas. It works well, but the user is asked to trust an opaque optimization. You see a route diagram. You can't easily reproduce why one path won over another.
Framing two: "AI agent acts on your behalf"
This is the agentic-wallet framing. Products like Bankr, deBridge's MCP server, and a growing list of smart-contract wallets let you grant a software agent specific permissions ("session keys") and let it execute multi-step DeFi operations from natural language. The agent runs against your wallet. You retain a kill switch. The technology is genuinely impressive, but it pushes a lot of complexity onto the user: you need a smart-contract wallet, you need to understand session-key scopes, you need to evaluate the trust model of the agent provider.
Framing three: "AI = chatbot"
A small number of exchanges have added an LLM-backed chat widget to handle customer-support questions. This is useful, but it is not "AI swap." The chat does not place trades. It answers questions. Calling it "AI swap" is a category error.
Cambio is closer to the chatbot framing than the other two — but with a critical difference: the chat surface places the trade. It is the primary swap interface, not a help widget bolted onto a form.
What "AI swap" means at Cambio
Cambio's AI does two things, and only two things. It parses your natural-language intent into a structured swap. And once you have a quote in front of you, it explains the route, the price, and the competitor comparison in plain language. That is the entire scope of the AI surface.
Everything else — picking the route, fetching the rate, comparing competitors, settling on-chain, releasing your funds — is deterministic, auditable, non-AI code. The router is a state machine. The pricing engine is a guardrail calculation. The competitor strip is a pull from a public aggregator API on a 5-minute refresh. None of those decisions involve a language model.
What the parser actually does
When you type "swap 0.3 BNB to USDT on BSC" into the composer, the parser extracts five fields: from-currency (BNB), from-network (BSC), to-currency (USDT), to-network (BSC), and amount (0.3, with direction = "amount-in"). The parser is regex-based with confidence scoring. It runs on every keystroke for live feedback. It tells you when a field is missing ("which network?") and when it's ambiguous ("did you mean USDT on BSC, Ethereum, or Tron?").
The composer offers five preset categories that map to five different parse families: Swap (amount-in), Execute (amount-out — "give me enough X to get 500 USDT"), Suggest (multi-option — "I have $500 in crypto, what's cheapest to stables?"), Compare (cross-network or cross-exchange), and Help (canned educational answers).
Note what's not in that list: there is no 'AI mode' that secretly switches to a more powerful model, no opaque routing toggle, no hidden execution path. The five categories are the five things the parser knows how to do. Anything outside those falls through to a polite error or a help response.
What the explainer does
Once the parser produces a structured swap and the router produces a quote, the AI explains the result. This is the conversational text that appears alongside the quote card — 'Found you the best route via Cambio's partner liquidity, settles in ~15 seconds on BSC, beats the next-best aggregator by 47 basis points.' The model behind this is Gemini 2.5 Flash Lite.
The explainer never invents data. It receives a structured object with the route choice, the quoted amount, the competitor comparison rows, and the settlement window, and it produces a sentence or two of plain-English narration. If the explainer ever says "Cambio beats by X%," that X comes from the deterministic calculation, not from the model's opinion.
What Cambio's AI does NOT do
This list is more important than the previous section. The boundaries are the design.
The AI does not pick the route. The router is a deterministic state machine: validate that the pair is supported, check partner liquidity on the from-side and to-side, fetch the live rate from the liquidity pricing service, apply the competitor guardrail, return a quote. The same input produces the same output every time. If the AI never ran, the route would still be the same.
The AI does not custody your funds. You send your deposit to a one-time address for that swap, and the payout goes to your own wallet. The AI never sees a private key. It never signs a transaction. It never holds your assets even briefly.
The AI does not predict prices. The "live rate" you see in the quote is the current mid-market price plus a deterministic spread floor — not an AI forecast. If BTC is moving fast, the rate updates because the mid moved, not because a model made a guess about where it's going.
The AI does not make safety decisions. The hard rejects (unsupported pair, deposit below minimum, deposit above per-swap cap) are policy code, not model judgment. If the AI ever told you a pair was supported when it isn't, the deterministic layer would catch it before any funds moved.
The AI at Cambio is a UI affordance. It makes the product easier to talk to. It is not — and was never intended to be — the decision maker.
Why this design beats "AI picks the best route" black boxes
Most "AI swap" products fail one or both of two tests. The first test is auditability: can you reproduce, from the same inputs, why the AI made the decision it made? Most LLM-driven routing fails this because the model's reasoning isn't deterministic. The second test is failure-mode transparency: when the AI is wrong, do you see it, or does it silently route you to a worse outcome?
Cambio's design passes both tests by moving the AI out of the decision path entirely. The route the AI explains is the same route a thin curl wrapper around our public API would have returned. The competitor comparison strip lets you visually verify the claim — the recognizable named competitors, their live rates, the basis-point delta. If Cambio isn't winning, the 'Best Rate by Cambio' badge doesn't show — the strip falls back to a neutral 'rates compared' label. We don't pretend.
Compare this to 1inch's Fusion mode. Fusion is intent-based: you sign an off-chain order, and a competitive market of resolvers competes to fill it. It's a very good design, and the resolvers' competition is largely auditable on-chain. But the user has to sign a transaction with their wallet, has to understand intent semantics, and has to evaluate whether the resolver they got is the best one or merely the fastest. Cambio chose a simpler trust frontier: the deterministic router runs on our infrastructure, you watch the live competitor strip while you decide, you send a one-time deposit, you receive funds.
The non-AI parts that make this work
Once you accept that the AI is just the conversational surface, you can ask the better question: what are the non-AI parts that actually decide your swap quality? Three things matter most.
Liquidity pricing
A swap is routed to whichever of the exchange partners actually delivers more at your amount. There is no AMM pool to slip through and no bridge contract to fail. The comparison behind that choice is shown with the quote, including when the winner is not the one we would have picked.
The five-competitor live strip
Every quote on Cambio shows the same swap on up to six recognizable competitors — SimpleSwap and the other instant-swap brands Bestchange lists for that pair, recognizable names first. The data is pulled from the Bestchange aggregator API on a ~5-minute refresh — not from each competitor's API individually — so the comparison is uniform and apples-to-apples. The 'Best Rate by Cambio' badge shows only when Cambio genuinely leads; when it doesn't, the strip shows a neutral 'rates compared' label instead of a best-rate claim.
Same-chain settlement
For same-chain swaps (everything on BSC, everything on Ethereum, etc.) the settlement window is the chain's confirmation time plus the time it takes for our payout transaction to be picked up — typically about 15 seconds end to end. For cross-chain swaps where the deposit chain is slow (Bitcoin, in particular), the deposit-confirm wait dominates the total settlement time. The receive side is still instant from partner liquidity.
The competitive landscape, briefly
It helps to place Cambio against the three other categories we touched on earlier.
Instant-exchange platforms like ChangeNow, SimpleSwap, ChangeHero, and Exolix are form-driven. You pick a pair from dropdowns, enter an amount, paste an address. Their "chat" is human support. They are mature and reliable; they do not have a conversational primary surface.
DEX aggregators like 1inch and Jupiter are wallet-connect-first. They are excellent for users who already manage a wallet and want the absolute best rate on a chain they already hold gas on. They are not built for the user who has BNB and wants USDT on Tron without learning the bridge stack.
Agentic wallets like Bankr and deBridge MCP-server-driven flows are the closest thing to Cambio's conversational pitch — but they push the agent inside the user's wallet. That's the right answer for power users who are comfortable granting session keys to a software agent. It's the wrong answer for the long tail of users who just want to type "swap 0.3 BNB to USDT on BSC" without learning what a session key is.
Cambio sits in the gap: conversational composer (not a form), non-custodial settlement (not a wallet-connect agent), transparent competitor strip (not a black-box route picker).
Why we built it this way
Two reasons. The first is trust. AI is genuinely useful as a UI affordance and genuinely dangerous as a decision-maker. Putting the AI in the conversational surface and keeping all decisions in deterministic code means a user can verify our claims by reading the on-page competitor strip rather than having to trust the model. We also publish a public API at api.cambio.one that returns the exact same quote — without the chat — so any user, integration, or auditor can reproduce the route without going near a language model.
The second reason is reliability. LLMs are improving fast, but they still have a non-zero rate of hallucination, latency spikes, and provider outages. If our router depended on a language model to pick routes, every model outage would be a Cambio outage. By isolating the AI to the composer and the explainer, we can degrade gracefully: if the model service is slow, the parser falls back to keyword matching and quick-action chips, and the explainer falls back to a deterministic templated string. The trade still goes through.
A note on the model itself
We use Gemini 2.5 Flash Lite as the conversational backbone — not GPT-5 or a frontier reasoning model. The reasoning is simple: the AI's job at Cambio is parsing and narration, both of which are well within the capability frontier of any modern fast model. Latency matters more than reasoning depth, because the user is sitting in front of a composer expecting near-instant feedback. We have a dedicated post on the model-selection decision later in this series.
What to take from this
When you read "AI swap" in a product description from any crypto exchange — including ours — ask three questions. Does the AI place the trade, or just talk about it? Does the AI hold any keys, or just structured fields? Can you reproduce the route without the AI in the loop?
Cambio's answers, in order: yes (the conversational surface places the trade), no (the AI never sees a key), yes (the public API returns the same quote without the AI involved).
Every other post in this series builds on that frame. The next one walks through the five categories the composer recognizes and shows you what each one actually does under the hood.



