Asher's Blog

Money is not merely a number inside a system. It represents time, work, obligation, security and possibility. When a company allows artificial intelligence to make or influence financial decisions, it is giving a machine access to consequences that can change a person’s life.

An AI system may identify fraud, approve credit, route payments, rebalance liquidity, recommend investments or decide which transaction deserves review. These uses can make financial services faster and more inclusive, particularly in markets where professional capacity is limited. They can also scale error, discrimination and opacity.

The central question is not whether machines can be trusted with money in the abstract. It is what design, governance and accountability make a particular level of trust justified.

Start with authority, not intelligence

Companies are impressed by what a model can infer and generate, but financial design should begin with what the system is authorised to do.

Can it recommend an action, prepare it for approval or execute it? What amount can it move? Which accounts can it access? Under what conditions must a human intervene? How quickly can the authority be withdrawn?

These boundaries should reflect consequence. An assistant that categorises expenses can operate with more autonomy than an agent that initiates payroll or invests customer funds. Capability should never automatically become permission.

The principle of least privilege is particularly important: give the system only the access required for the task, for only as long as necessary. An agent should not possess broad financial authority merely because broad authority makes the demonstration more impressive.

Every movement needs an explanation

When money moves, the organisation must be able to reconstruct why. The record should identify the instruction, relevant data, model or rule used, approvals, transaction reference and final outcome.

This protects both customers and the company. If a payment is disputed or an investment recommendation causes harm, a fluent summary generated afterwards is not enough. The organisation needs an audit trail created at the time of the decision.

Explainability does not mean every customer needs to understand the mathematics inside a model. It means the institution can provide a meaningful reason connected to evidence and policy. “The AI decided” is not an explanation.

Data can make exclusion look objective

Financial history reflects unequal access. If a model learns only from previous approvals, it may reproduce the preferences and exclusions embedded in those decisions. People without conventional records can appear risky simply because the system knows less about them.

This matters in African markets where many individuals and small businesses operate outside formal credit structures. AI can widen access by interpreting alternative evidence, but alternative data can also become intrusive. A company should not demand disproportionate visibility into a person’s private life merely because they lack a traditional credit file.

Fairness requires testing outcomes across relevant groups, examining which features drive decisions and creating a path for correction when data is wrong. Inclusion should not mean surveillance in exchange for access.

Human oversight must be real

Adding a human approval step does not automatically create meaningful oversight. If the employee cannot understand the recommendation, lacks time to review it or knows that disagreeing will be questioned, the human becomes a ceremonial signature.

The reviewer needs relevant evidence, authority to stop the action and training to recognise model failure. The system should highlight uncertainty and unusual conditions rather than presenting every recommendation with identical confidence.

Human review should concentrate on high-risk, novel and contested cases. Requiring manual approval for every ordinary transaction may defeat the value of automation while encouraging employees to approve mechanically.

Design for failure before granting trust

Financial systems must assume that models, infrastructure and third-party providers will sometimes fail. What happens when the AI is unavailable, produces contradictory recommendations or attempts an action outside policy?

Safe design includes transaction limits, separation of duties, reconciliation, anomaly detection, rollback where possible and clear escalation. The company needs a manual route for essential services and must know how to suspend automated authority without losing visibility into pending work.

Regulators and central banks increasingly focus on model risk, data governance, operational resilience and third-party concentration because an institution remains responsible even when the intelligence comes from an external provider. Outsourcing the model does not outsource accountability.

Customer consent should match customer understanding

Financial companies often rely on broad terms and conditions to authorise complex data use. Legal permission may be necessary, but trust requires more.

Customers should know when AI materially influences a decision, what type of information is used and how they can challenge an outcome. The explanation should be proportional to the consequence. A small categorisation feature needs less disclosure than an automated credit refusal or investment decision.

Consent must also avoid manipulation. People in urgent need of money may accept terms they do not understand. The institution has a responsibility not to hide significant risks behind convenience.

The machine should serve the customer’s objective

An AI agent handling money may be asked to optimise returns, liquidity, fees, repayment or engagement. These objectives can conflict.

A savings platform could increase short-term revenue by encouraging more transactions, while the customer’s goal is to preserve money for the future. A lending system could optimise approval volume while increasing the probability of distress. The objective encoded into the machine is therefore an ethical decision.

Companies must define whose interest the system serves and where commercial incentives are constrained by duty. Intelligent optimisation without moral boundaries can become highly efficient exploitation.

Trust should be earned in stages

Do not begin by allowing an unproven system to move significant customer funds. Start with observation and recommendation. Compare its decisions with experienced people, test unusual conditions and examine not only average accuracy but the consequences of the worst errors.

Authority can expand when evidence supports it. Limits can increase, more cases can become automatic and human review can focus on exceptions. Trust becomes the result of demonstrated reliability rather than enthusiasm for the technology.

The company should also continue testing after deployment. Financial behaviour changes, fraudsters adapt and data distributions move. A system that performed well last year may become unsafe without obvious failure.

Responsibility remains human

Machines will handle more money because they can monitor patterns and execute processes at a speed humans cannot match. That can reduce errors, expand access and make complex services affordable.

Yet money requires more than intelligence. It requires duty. Someone must decide the objective, approve the authority, protect the customer and answer when the outcome is wrong.

We can trust machines with increasingly important financial work when their power is bounded, their actions are visible and accountable people remain responsible. The machine may execute the decision. It must never become the excuse that allows the institution to avoid owning it.