When crypto and AI is actually usable
The market is not short of models that can talk about a wallet. It is short of surfaces where a person can see the input and refuse the trade.
Sample note, written to show the format. The table is a product map, not a survey, and it has no market-size figure.
The market is not short of models that can talk about a wallet. It is short of surfaces where a person can see what the model used, disagree, and still end up in a position they understand.
Three shapes are doing the rounds.
Chat agents. You connect a wallet, ask for a plan, and the agent proposes swaps, bridges or vaults. The interface is a transcript. It is fast to try, and the reasoning is in prose. It is a poor place to compare two risks, and “execute” is one button at the bottom of a confident paragraph.
Canvases and copilots. Positions are laid out, some notion of risk is computed from on-chain features, and the model annotates the picture rather than replacing it. The person can see the inputs. It is harder to build, and less magical in a screenshot.
Autonomous vaults. The user delegates a mandate, and an agent rebalances inside it. This matches how people say they want automation. Delegation dies the first time the vault does something the user cannot explain. A repeated theme in this market is that people will not hand over the keys if they cannot replay the decision.
Explain, then execute
The usable version is probably the middle shape, with a narrow door into the third. Show the features. Show the band. Show the action. Let execution be deterministic where the action is a well-defined call. Use the model to rank and explain. Use code to move funds.
That split is less exciting than an agent with a wallet of its own. It is also the split a neobank, a wallet or a custody product can stand behind, because a compliance conversation needs a reason, not a vibe.
Who buys it depends on the seat. A retail wallet wants something that cuts support tickets: why this yield is not the one the user saw somewhere else. A fintech wants an intelligence layer it does not have the team to build. A protocol wants distribution and will sponsor an integration. Those are three sales motions. One demo does not cover them.
| Surface | Can the person audit the input? | Can they refuse execution? | Fits a regulated app? |
|---|---|---|---|
| Chat agent | Weakly. The input is a thread. | Only by not clicking. | Rarely. |
| Canvas or copilot | Yes, if the features are on the page. | Yes. The action is separate. | Yes. |
| Autonomous vault | Only if the mandate is explicit. | Only by leaving. | Sometimes, if the mandate is narrow. |
“Usable”, for a product rather than a paper, means a new user can answer three things on one screen: what they hold, what risk band that implies, and what would change if they accepted a suggestion. If any of the three requires reading a thread, it is not usable yet. It is a demo that happens to be online.
No dollar figure belongs in this file. A precise market size here would be fiction. The useful cut is the one between explain and execute.