Banks, brokers, insurers and corporate treasuries are putting machine learning into credit scoring, market risk, margin forecasting and fraud detection. The technology gets the attention. The harder problem is staffing: who builds it, who validates it, and who has the authority to say no.
Most mid-sized finance employers do not need a quant lab. They need a small group of people who can run models responsibly. Here is how we would think about building that group.
Separate the people who build from the people who check
Regulators in both countries expect model risk to be managed independently. In the US, the Federal Reserve and OCC guidance on model risk management (SR 11-7) applies to machine learning models as much as to older ones. In Canada, OSFI's guideline on model risk management sets similar expectations for federally regulated institutions.
That means two distinct hires, not one:
- Model developers who build and tune the models.
- Independent validators who test them, challenge assumptions and document limits.
If one person does both, you have a control weakness. If you only hire developers, you will have models you cannot defend in an audit.
The roles that matter most
In our experience filling finance and data roles, these are the positions that decide whether an AI risk programme works:
- Data engineers who clean and connect data from older trading, risk and core systems. Model quality depends on this work more than on the algorithm.
- Quantitative or risk analysts who understand both the maths and the business question, such as exposure, margin or credit loss.
- Model validation specialists with enough statistics to test a gradient boosting model, not just a regression.
- Risk and compliance leads who can read a model's documentation and explain it to a regulator.
- Operations and collateral staff who can review AI suggestions on margin and funding before anyone acts on them.
What to screen for
A strong résumé in machine learning is not enough in a regulated setting. Ask candidates to show:
- Explainability work. Have they produced reason codes or factor-level explanations that a credit officer or underwriter actually used?
- Drift monitoring. How did they detect that a model had gone stale, and what did they do?
- Failure stories. A candidate who has seen a model misfire and can describe the controls that followed is worth more than one who has only seen it work.
- Documentation habits. Validation lives or dies on written records.
- Plain-language communication. Can they explain a result to a controller or a board member without jargon?
Keep a human in the decision
The sensible pattern in most firms is that the model recommends and a person decides, at least for high-stakes calls such as credit limits, trading limits and collateral substitution. Many teams also run a new model in parallel with the old process for a period before relying on it.
That design has a staffing consequence. You need people with enough judgment and standing to override the model. A junior analyst who is afraid to question the output is not a control. When you write the job description, state the authority the role carries.
Data quality is a hiring decision
Bad inputs produce bad outputs, and legacy systems often hold the data in silos. Before you hire modelling talent, check whether you have anyone who owns data governance. Teams that skip this step end up paying expensive modellers to do data cleanup.
For smaller employers, a contract data engineer or analyst can be a sensible first move. It lets you fix the pipeline, prove a use case and then decide whether a permanent hire is justified.
Expect a tight market
People who combine finance domain knowledge, machine learning skill and regulatory awareness are scarce in both Canada and the US. Many candidates have two of the three. Decide which one you can train. In most cases it is easier to teach a strong data scientist the business than to teach a risk veteran modern modelling, but that depends on your team. Pay for the scarce combination, and be clear about which gap you are willing to carry.
Also think about retention. People who build models that regulators scrutinize want a clear mandate, decent tools and a say in governance. Employers who treat the work as a side project lose these hires within a year or two.
What to do next
List the AI-assisted decisions your finance or risk team already makes or plans to make. For each one, name who builds the model, who validates it, who can override it and who owns the data feeding it. Wherever you find a blank, that is your next hire or contract engagement.
A Traverse recruiter can help you scope those roles, decide between contract and permanent, and find candidates who have worked in controlled environments.