There are two unhelpful positions on AI in financial services. One holds that it changes everything immediately. The other holds that it is a passing enthusiasm best waited out.
Both are ways of avoiding the actual work, which is unglamorous: figuring out specifically where this technology creates value in your specific operation, and being disciplined about everywhere it does not.
The filter I use
Before approving any AI initiative, I want clear answers to four questions.
1. What is the task, precisely? Not “improve customer service.” Something like: “draft a first response to routine account queries so an agent edits rather than composes.” Vague scope produces vague results and unmeasurable outcomes.
2. What does failure look like, and who absorbs it? This is the question that matters most in a regulated industry. If the system is wrong, is the consequence a slightly awkward draft that a human fixes, or a customer wrongly declined? Tolerance for error should scale inversely with consequence.
3. Where is the human, and do they have real authority? “Human in the loop” is meaningless if the human is approving two hundred recommendations an hour. That is not oversight; it is a signature. Genuine oversight requires time, context, and the practical freedom to disagree.
4. Can we explain the outcome to a customer? If a decision affects someone’s access to a financial service, “the model determined it” is not an acceptable explanation. This is not only a regulatory consideration — it is a basic matter of treating people properly.
Where the value tends to be
In my observation, the reliable wins are internal and unglamorous:
Drafting and summarising. Documents, summaries, first-pass responses. Low risk, immediate time savings, human review built in.
Search across internal knowledge. Staff spend a surprising share of their day looking for information that already exists somewhere. Making institutional knowledge genuinely findable is quietly transformative.
Triage. Sorting, routing, and prioritising queues so that human attention lands where it matters. The model is not deciding — it is ordering the work.
Pattern surfacing in monitoring. Not autonomous decisions, but better prioritisation of what a human investigator examines first.
Notice the shape: the technology is drafting, sorting, and surfacing. A person is deciding.
Where I would be cautious
Anywhere the output is customer-facing without review, and anywhere the decision materially affects a customer’s access to services, the bar rises sharply — and the honest answer is often that the current tooling is not ready for unsupervised use.
I would also be wary of initiatives whose primary justification is that competitors are doing it. That is not a business case. It is anxiety with a budget.
On the workforce question
People ask whether this reduces headcount. The more useful question is what the freed time is used for.
If it is simply removed as cost, you get a short-term saving and a demoralised team. If it is redeployed to the things that were always underserved — genuinely difficult customer cases, quality of analysis, improvements nobody had time for — you get a better business.
That is a leadership choice, not a technological outcome. The technology does not decide it. We do.
Where this goes
The institutions that do well with AI will not be the ones that adopted earliest or spent most. They will be the ones that were honest about what worked, ruthless about retiring what did not, and disciplined enough to keep humans accountable for decisions that affect people’s money.
The technology is genuinely useful. The discipline is what makes it valuable.
Dr. Mohamed Mousa writes about financial services, technology, and leadership.