Are we ready to let AI agents control money?
by•
Hot take: the next big UX shift won’t be another app.
It’ll be humans talking to agents, and agents quietly doing the work in the background.
The real question is: how much control are we actually willing to give them?
60 views
Replies
Where do you think the boundary will be between helpful automation and and loss of user control?
@craig_bennett1 I think boundary is: agent can suggest, prepare, and execute low-risk actions inside clear limits. Anything irreversible, large, or unusual should need human approval. Control dies when permissions become vague.
Trust isn't a dollar dial, it's an action-class permission tree: reversible operations (read balance, draft a transfer) versus irreversible ones (signed send, contract execute). Stripe Connect scopes (https://stripe.com/docs/connect/oauth-scopes) and Plaid auth tiers are how this gets shipped in production, not as a single "how much authority" slider.
@fabriziowexare Exactly. “AI controls money” is too vague. Real control has to be scoped by action type, amount, destination, frequency, and reversibility. That permission tree is the layer agents need before they can safely move capital.
memi
Agents should probably start with budgets, limits, and receipts before they get full control. The best UX may be "AI does the homework, human signs the check."
@sarveshsea Exactly. Start with budgets, limits, receipts, and revocation. “AI does the work, human signs the check” is probably the right bridge before agents get real spending authority.
Trust isn't a dollar dial, it's an action-class permission tree: reversible operations (read balance, draft a transfer) versus irreversible ones (signed send, contract execute). Stripe Connect scopes (https://stripe.com/docs/connect/oauth-scopes) and Plaid auth tiers are how this gets shipped in production, not as a single "how much authority" slider.
@fabriziowexare Agree. A single “trust level” is too blunt. Real agent money control needs scoped permissions by action type, amount, destination, frequency, and reversibility. More like permission trees than a slider.
The thread already has the right frame: agents do the homework, a human signs the check, money only moves on approval. The underrated part is the receipt, not "did it spend" but "did the output earn the spend." That's the bet I'm making with BotWork: the agent does the task, money sits in escrow, and you only pay if the result is actually good.
@theuniverseson This overlaps directly with what we’re building at Tychi. BotWork has the task/result layer; Tychi can power agent wallets, permissions, escrow, and settlement. Would you be open to collaborating on this?
@singhyash05 Appreciate that, and the split you describe is real: we sit on the task-and-result side, and settlement is its own hard problem. Happy to compare notes on where the escrow boundary should sit. Easiest is to reach me through botwork.network and we can find the overlap.