The product
Artificial intelligence in this product
The job description names AI and machine learning under Technology savvy, and calls out agentic payments in wholesale transaction banking as a strong advantage. This page is the answer to 'where does AI actually fit, and what would you refuse to let it do?'
The one line to lead with: artificial intelligence prepares the evidence; a human or a pre-authorised mandate approves the movement of money. Say this before describing any use case and most risk objections are answered before they are raised.
Where it genuinely helps
Ordered from safest to most speculative, because the credible position in a bank is to start where a mistake is recoverable. Each one states what a human still owns — that boundary is the product decision, not a technical detail.
Reducing false alarms in sanctions screening
Screening systems flag any name resembling one on a government watchlist. The overwhelming majority are innocent coincidences, and each one costs an analyst time. Machine learning ranks alerts by genuine likelihood so analysts open the dangerous ones first.
Concrete example
A payment to 'Ming Liu Trading' is flagged because a sanctioned individual shares a similar name. The model weighs country, payment history, counterparty age and context, scores it as very low risk, and pushes it to the bottom of the queue instead of stopping a client's cash for four hours.
The value
Faster release of legitimate payments and analyst attention spent where the real risk is. This is the clearest near-term financial case in the whole list.
The risk, and how it is contained
If the ranking is wrong in the dangerous direction, a genuine hit sits unexamined at the bottom of a queue. Mitigated by never letting the score close an alert, and by sampling low-ranked alerts for audit.
What a human still owns: The model re-orders the queue. It never clears an alert by itself, and it can never un-flag a true watchlist hit.
Explaining exceptions to clients in plain language
When a transfer is held, a language model turns the internal case record into a clear client-facing explanation of what happened and what comes next — without disclosing anything confidential about the screening logic.
Concrete example
Instead of 'TXN-4471 STATUS: HELD — RULE SC_0142', the client sees: 'Your transfer is under a standard compliance review. Reference 4471. We expect an update by 14:00 today and no funds have left your account.'
The value
Fewer inbound support calls and materially better client experience at the exact moment trust is most fragile.
The risk, and how it is contained
A model that over-explains could leak how detection rules work, which is a financial-crime risk. Mitigated by constraining output to pre-approved templates rather than free generation.
What a human still owns: Templates and disclosure rules are fixed by compliance. The model fills an approved shape; it does not decide what may be revealed.
Predicting liquidity needs before the client asks
Models learn a corporate group's cash rhythms — payroll cycles, tax dates, seasonal supplier runs — and forecast which entity will be short before it happens.
Concrete example
The system notices the Singapore entity has run short every 27th for six months, and prompts the treasurer on the 25th: 'Singapore is likely to need approximately US$2m by Thursday. Prepare a transfer?' The treasurer decides.
The value
Turns the product from a tool the client must remember to use into one that earns its place daily. Strong retention argument.
The risk, and how it is contained
Over-confident forecasts could encourage a client to move cash they actually need where it is. Mitigated by always showing the forecast's confidence and the reasoning behind it.
What a human still owns: It proposes. The treasurer approves. A forecast never triggers a movement of money on its own.
Drafting deal and proposal responses
Retrieval-based assistance for responding to a Request for Proposal: pulling approved product answers, prior responses and control documentation into a first draft for a specialist to correct.
Concrete example
A 200-question proposal from a multinational arrives. The assistant drafts answers from previously approved material and flags the twelve questions with no approved precedent, so humans spend their time only on genuinely new ground.
The value
Directly addresses the deal-solutioning responsibility in the job description, and shortens the slowest part of the commercial cycle.
The risk, and how it is contained
A fabricated capability claim in a contractual document is a serious commercial and legal exposure. Mitigated by retrieval from approved sources only, never open-ended generation.
What a human still owns: Nothing reaches a client without named human review. The assistant is explicitly a drafting tool, not an approver.
Agentic payments
Software agents that initiate payments on a client's behalf within limits the client set in advance — the specific capability the job description names as a strong advantage. The agent holds a delegated mandate: what it may pay, to whom, up to how much, and how often.
Concrete example
A company authorises an agent to top up its Singapore entity automatically whenever the balance falls below US$500,000, capped at US$2m per day, only to its own pre-approved entities. The agent acts inside that envelope and cannot step outside it.
The value
This is where digital money becomes genuinely different from a faster wire — money that can move under programmatic rules, safely, with an auditable reason for every movement.
The risk, and how it is contained
The hardest question in the whole domain: who is liable when an agent pays the wrong party? A compromised agent could drain accounts at machine speed. This is why mandates, velocity caps and an instant revocation path matter more than the agent's intelligence.
What a human still owns: The mandate is the control. A human sets the limits, the counterparties and the kill switch; the agent operates strictly inside them and every action is attributed to the mandate that authorised it.
Detecting reconciliation breaks earlier
Anomaly detection watching the token ledger against the core banking record, catching divergence patterns before a scheduled end-of-day reconciliation would surface them.
Concrete example
Balances match at every checkpoint, but the model notices settlement confirmations arriving 40% slower than normal for one corridor — an early symptom of an upstream problem, spotted hours before it would become a break.
The value
Shortens the window in which a problem compounds. In settlement, the cost of a break grows with how long it goes unnoticed.
The risk, and how it is contained
Alert fatigue if tuned too sensitively, causing real anomalies to be ignored. Mitigated by tracking the ratio of actioned to ignored alerts as a product metric.
What a human still owns: It raises an operational alert for investigation. It never adjusts a balance or writes a correcting entry.
The control framework that makes it acceptable
Any of the above is only viable inside a bank if these hold. Being able to recite this framework matters more than naming the models — it is what a committee will actually ask about.
Artificial intelligence never has payment authority
It prepares, ranks, drafts and predicts. A human or a pre-authorised mandate approves. This single line answers most risk objections before they are raised, and it is the position to state first.
Every model output is attributable and reproducible
If a model influenced a decision, the record must show which model, which version, and what evidence it saw. Without that, you cannot answer a regulator asking why a payment was treated the way it was.
Explainability scales with consequence
Ranking a queue needs less justification than holding a client's money. The more severe the consequence, the more the reasoning must be inspectable by a person.
A tested path to switch it off
Every model needs a proven fallback to the manual process, rehearsed rather than assumed. Resilience testing is a roadmap gate for a reason.
Bias and drift are monitored as live risks
A screening model that quietly becomes harsher on payments from one region creates fair-treatment exposure. Models degrade as the world changes, so monitoring is ongoing, not a launch checkbox.
Client data boundaries are explicit
Corporate treasury data is highly sensitive. Whether it may train shared models, and whether it can leave a jurisdiction, are product decisions with legal consequences — not technical details to delegate.
Questions you are likely to be asked
Why is artificial intelligence relevant to a settlement product at all?
Because settlement is mostly checking, not moving. The movement is nearly instant; the time is spent on screening, funds verification, exception handling and reconciliation. That is exactly the shape of work models are good at — high volume, pattern-heavy, with a human needed only at the point of consequence.
What is the difference between automation and an agent?
Automation follows a fixed rule you wrote. An agent pursues a goal you set and chooses its own steps. That difference is the entire risk conversation: with automation you review the rule, with an agent you must constrain the space it may act in, because you cannot enumerate what it will do in advance.
Where would you refuse to use it?
I would not let a model make the final sanctions decision, and I would not let it approve a payment outside a pre-agreed mandate. Both create accountability I cannot explain to a regulator. Refusing specific uses is what makes the proposed uses credible.
How does this connect to tokenised money specifically?
Programmable money and software agents are complements. An agent needs money that can move under rules, around the clock, with instant confirmation — which is what a tokenised deposit provides and a batch-based legacy rail does not. That is the real long-term argument for tokenisation, beyond speed.
Why this pairs with tokenised money rather than sitting beside it
Software agents need money that can move under rules, around the clock, with instant confirmation of whether it worked. A batch-based legacy rail with daily cannot serve an agent acting at 3am. That is the strongest long-term argument for tokenisation — not that it is faster, but that it is the only settlement layer software can actually operate on its own. If you want one forward-looking point to close on, this is it.