AI Agents Need Allowances, Not Credit Cards

Every conversation about agentic payments right now is about agents buying things. Checkout flows, wallets, settlement rails, agents negotiating with agents. I read Kahlil Lalji's agentic payments memo like everyone else in fintech did, and it's the rare piece of category writing that actually earns the traffic.

But I run a company that sits on the other side of the transaction, and from where I sit the discourse is missing half the market. Agents don't just need to buy things. Agents need to give money to people. And nobody wants to talk about that half, because giving an AI the ability to send money sounds like the setup to a horror story.

Here's the thing though: we've been running this experiment in production. Hoppier shipped an MCP server this summer, which means Claude, ChatGPT, and any MCP-capable assistant can create reward programs, add recipients, and send top-ups through a Hoppier account. Real money, real people, real agents in the loop. And what we've learned is that the problem everyone is afraid of has had a solution for years. It just wasn't built for agents. It was built for interns.

Nobody hands the new hire a corporate card

Think about how companies actually delegate spending to humans they don't fully trust yet. You don't give the intern the company Amex. You give them a budget: $500, for the offsite, at these vendors, spent by Friday, receipts required. The guardrails aren't in the employee handbook. They're on the money itself.

That is exactly the right mental model for agents, and it's the opposite of how most of the agentic payments stack is being designed. The industry is building open rails and then bolting review flows on top: human approval over $100, then over $500, then over $2,500 as trust grows. Kahlil describes this trust curve well, and I think he's right that it's the adoption path. But approval thresholds are guardrails in the workflow. When a prompt injection or a bad tool call slips past the workflow, the money is still sitting there, unrestricted.

An allowance flips that. The constraint lives on the money, not in the prompt. A Hoppier card an agent sends can be restricted to food merchants, valid only during your event window, capped per person, and anything unspent comes back to your balance. If the agent screws up, the blast radius is a $25 lunch card that only works at restaurants this week. Not a wire. Not a line of credit. There is no prompt injection against a merchant category restriction.

What the data says about giving money well

We looked at 2,090 gift and incentive programs run by 350 companies on our platform: 30,986 cards, $1.18 million sent. Three numbers changed how I think about agentic disbursements.

First: 37% of gift value goes unspent. More than a third of the money companies send as rewards is never used. In a payout model, that money is simply gone, priced in as breakage. In an allowance model it comes back. The moment you put an agent in charge of sending hundreds of rewards a month, that difference stops being rounding error and becomes the budget.

Second: small beats big. Programs with 25 or fewer cards see 70% redemption. Programs over 500 cards see 40%. Personal, well-timed sends outperform mass blasts by 30 points. This is precisely what agents are good at: catching the individual trigger, the closed deal, the work anniversary, the finished survey, and sending one right-sized reward at the right moment instead of a quarterly spray.

Third: the average card we see is $38. Agentic disbursements aren't wire transfers. They're small, frequent, and personal. Which means the risk conversation everyone is having about agents moving money is, for this half of the market, mostly a conversation about lunch.

Agent-initiated is enough to be useful

The maturity framing I'd steal from the payments world: agent-initiated first, agent-executed later. Today, a human sets the rules - the budget, the merchants, the window - and the agent operates freely inside them. That's not a compromise while we wait for the real thing. For disbursements, it might just be the right permanent architecture, because the person setting guardrails and the agent doing the repetitive sending is a better division of labor than either one alone.

What we see in production: an ops lead connects their assistant to Hoppier once, then delegates the parts of the job nobody loves. Birthday sends. Monthly wellness stipends. Webinar thank-yous. The agent watches the calendar and the CRM; the human approves the shape of the program, not every individual $30 card. Trust widens the same way it did for code review: you check everything, then spot-check, then stop checking the small stuff entirely.

What we haven't solved

Honesty about the gaps, because category essays that pretend everything works are marketing, not thinking.

Agent identity is unsolved, for us and for everyone. Today an agent acts through your authenticated account, so accountability is clean but coarse: the audit trail says what happened and under whose login, not which agent, which version, which prompt. As teams run fleets of agents, per-agent identity and reputation will matter, and I don't think anyone has cracked it yet.

And our own REST API is still in closed beta. The MCP server is live and it's the fastest path into this today, but developers who want raw programmatic access still have to talk to us. We know. We're working on it.

Where this goes

I think agentic payments splits into two markets that get conflated today. Agents buying things is a rails problem, and the companies building for it are building settlement infrastructure. Agents giving money to people is a controls problem, and it needs a different primitive: budgets that enforce themselves.

The first market will be enormous. But the second one is deployable now, with technology that already exists, at a risk level a CFO can sign off on this quarter. Give the agent an allowance, not a card. Put the guardrails on the money. Let the humans decide the rules and the agents do the sending.

That's the bet we're making. If you want to try it, the Hoppier MCP takes about two minutes to connect, and the agentic payments page covers how the controls work. If you're building in this space and think I'm wrong about any of it, I genuinely want to hear why.