“Innovation is the market introduction of a technical or organisational novelty, not just its invention”. In corporate treasury, that distinction has become a bottleneck. AI can already analyse cash flows and plan transactions in milliseconds, yet B2B payment execution remains bound to legacy infrastructure built for human sign-offs.
Treasury leaders face a clear divide – software can now reason and plan transactions instantly, but global clearing rails lack the deterministic policies and identity models needed to let software move capital safely. The solution? Moving beyond automated API triggers toward built-for-purpose governance, a true “trust layer” that dictates how, when, and why autonomous software can transact.
On this note, in the second article of “Voices from the frontier founders,” our Editor-in-Chief, Deepesh Patel, speaks with Ghali Bennani Laafiret, Co-founder and CEO of Ralio. This is our new series, where we profile some of the leaders and founders who are rebuilding liquidity and risk management. Deepesh and Ghali explore how Ralio is engineering the policy, identity, and audit infrastructure required to make agent-initiated B2B payments secure, predictable, and fully controllable in this exclusive interview.
Deepesh Patel: Ralio is building a “trust layer” so AI agents can make business payments safely. For a treasurer or payments banker who hasn’t yet let software move money autonomously, what does the near-term reality of agentic payments actually look like?
Ghali Bennani Laafiret: For a treasurer or payments banker who hasn’t yet let software move money autonomously, the near-term reality is more constrained and more familiar than the “agents moving money” framing suggests. Autonomy arrives payment type by payment type, starting where the risk is lowest and the patterns are most predictable.
In practice, the first payments to go autonomous are the ones you already trust to be routine, like recurring supplier invoices, subscription renewals, or payroll runs that match an approved schedule. These are repeated counterparties transactions with known amounts and predictable timing, so an agent executing them is closer to running a rule-based script you already follow than to applying agentic reasoning. Higher-risk payments can require a human in the loop for sign-off, such as a first payment to a new supplier, anything above a set threshold, or cross-border transactions.
The near-term reality therefore looks like one where payment types get dedicated agents with clear boundaries, and the treasurer decides where the human-in-the-loop checkpoint sits for each one. Ralio’s role in this reality is to make that enforceable by holding the policy on what an agent can and cannot execute, and keeping a record of what was approved and why. As trust grows, treasurers and payments bankers can delegate more, starting with the more standard payments and expanding as confidence builds.
Deepesh Patel: What are the specific risks — fraud, runaway spend, identity, auditability — that make agentic payments different from today’s automated or API-based B2B payments, and how does Ralio address them?
Ghali Bennani Laafiret: Today’s automated and API-based B2B payments run on fixed rules, where a scheduled payment file executes exactly what it was told, with counterparty, amount, and timing set in advance by a person. Agentic payments are different because the agent decides what to pay, how much, and when, based on reasoning rather than pre-set instructions. That shift is where the new risks come from.
An attacker no longer has to steal a credential to commit fraud, but can feed the agent bad inputs, a poisoned invoice or a spoofed instruction, which it may act on as legitimate. Ralio forces every agent to record what the user asked for and why before it touches a rail, then has a separate AI score recipient accuracy, amount, and consent, so a manipulated input has to survive an independent second opinion.
When it comes to runaway spend, an agent making its own decisions can pay the same thing twice or send far more than intended. Here, Ralio’s deterministic rulebook comes in, with pure code and no language model involved. If a spend limit is exceeded or a field is missing, the payment is blocked or escalated. Calls also accept an idempotency key, so a retried task never pays twice.
Identity also gets harder because existing controls authenticate the payer once rather than each payment. If an agent triggers a payment, current controls wouldn’t record whom nor why. Ralio gives each agent its own identity tied to a fixed purpose set at creation that can’t be renegotiated in a chat, so every payment records which agent acted, and anything outside its mandate is blocked. Ralio stores that reasoning for auditability, and flags any case where the independent review disagrees with the agent’s own reasoning.
In all four cases, Ralio makes the agent’s intent visible, its actions enforceable, and its decisions reviewable before it acts.
Deepesh Patel: You support everything from FPS and SEPA to stablecoins in a single flow. How should businesses think about choosing rails when an agent, rather than a person, is initiating the payment?
Ghali Bennani Laafiret: When a person initiates a payment, they pick their rail of habit or the one their bank defaults to. When an agent initiates it, the rail should be chosen by the payment’s requirements, and the agent can make that choice at the moment it pays.
Therefore, businesses should think about it depending on a few points.
- Speed and timing. Some payments need to settle immediately, like a supplier releasing goods on confirmation, for which a faster rail like FPS clears in seconds. For others which aren’t as time-sensitive, a slower and cheaper rail is fine. The agent can match the rail to how quickly the money actually needs to arrive.
- Cost. Rails differ a lot in what they charge, and the right choice depends on the size of the payment. A small domestic payment doesn’t justify an expensive rail, and a large one may be worth a pricier route if it settles faster or more reliably. In this case, the agent can weigh the fee against the value and pick accordingly.
- Destination and currency. A domestic payment in the local currency has obvious rail options. But a cross-border or cross-currency payment is where stablecoins can be faster and cheaper than a traditional route, or where SEPA comes into play inside the euro area. The agent can choose to route to whatever reaches the counterparty in the currency they want.
The point of supporting all of them in one flow is that the business doesn’t pre-commit to a rail. You set the rules, and the agent picks the rail that fits each payment, within the rails your bank and providers actually give you access to, rather than forcing every payment down one route.
Deepesh Patel: Treasury and finance functions are built around human approval and controls. How do those governance frameworks need to change for a world where agents transact, and what should finance leaders be doing now?
Ghali Bennani Laafiret: When it comes to governance frameworks for agentic payments, most of what treasury already does (approval limits, segregation of duties, audit trails…) carries over. What changes is that the controls have to apply to an agent as the payment initiator, and that some of them have to run automatically before the payment does, rather than as a manual check a human performs.
Approval limits become rules the agent is checked against automatically, with a human stepping in only when a payment falls outside them. Segregation of duties then has to extend to the agents themselves, so the model reviewing a payment isn’t the same one that proposed it. Identity has to record every payment which an agent acted on and on whose authority. And the audit trail now has to capture the agent’s reasoning as well, so you can always reconstruct why a payment happened.
For finance leaders, the useful work now is familiarising themselves with the technology and delegating small amounts of payment to start. That means understanding how agentic payments work, and the value they unlock, before real money is involved, then handing an agent a narrow set of payments, while potentially keeping humans in the loop. As the controls prove themselves and confidence builds, they can delegate more.
The real shift is that governance moves from reviewing payments after they’re proposed to encoding the rules before the agent acts.
Deepesh Patel: Looking three to five years out, how large a share of B2B payments do you expect to be agent-initiated — and what’s most likely to slow that down: technology, trust, or regulation?
Ghali Bennani Laafiret: Given how fast the agentic payments space is moving, three to five years is a very long time, and putting a share on it is hard to say. What’s already realistic, however, is that this is how agent-native companies will run their payments. These companies already delegate operations like Sales and Marketing to AI agents, and in the near future, agent-initiated payments will be the default way money moves for them.
As for what slows it down, the technology is the least of it, as the rails already exist, and the hard part of making an agent’s intent visible, enforceable, and reviewable is an engineering problem that is solvable now. Trust, however, is what’s most likely to hold things back. A finance team will need to see the controls work on their own payments first, which only happens with time and use. That’s why teams start by handing over a small, low-risk set of payments and widen it as they get comfortable, and why the pace depends on them rather than on the technology.
Regulation is also a hard one to predict, as much of the existing framework was written assuming a human initiates the payment. Agent-initiated payments raise questions that the rules don’t yet cleanly answer, such as who’s liable when an agent pays incorrectly, or what counts as valid authorisation. Regulatory clarity there could accelerate adoption, while uncertainty could hold it back regardless of how good the technology or the track record is.
