Decision Engineer: role, skills, and impact in automated decision making
A scoring model can include an attribute that is not supposed to be there because it has a strong correlation and looks cute. It has no lever behind it, so it stops working when something moves, and you do not know when.
The decision engineer engineers the decision by looking first at the outcome, then at the situation, and then at the levers within that situation that lead to the outcome. They look for the data only after that, then check whether they can automate the decision through situational awareness, situational understanding and situational logic. The situation has to be known.
How is a decision engineer different from a data scientist?
A good decision engineer gathers situational awareness, then situational understanding, and then situational logic. They look at where the data came from, what the data represent in the business, and what the data can predict about future behaviour.
A data scientist can take collected data and produce a strong statistical model. That model may have no connection to where the data came from or what the data represent in the business. In credit, someone may say they can predict who will be good or bad. When you examine how the data was collected and what the attributes represent, those attributes may be unable to support that prediction. A decision engineer makes these connections before relying on the model.
What does a decision engineer do with the product team?
A good decision engineer often talks to the whole product team. They work with the team to improve situational awareness, situational understanding, and situational logic. They ask how the team collects the data, what it represents, and what the customers look like. When there are edge cases, they consider what the data would look like.
They validate situational awareness by testing the decision models and comparing the outcomes with expectations. They adjust and refine the models as needed.
How does a decision engineer connect underwriting and collections?
They might go to the collections team and ask about patterns in the customers they cannot find. They ask what data could improve collections when they do find a customer.
This can lead them back to underwriting. A group of customers may look good in scoring, yet be impossible to find later in collections. The decision engineer looks for data available at the time of application that could identify this group earlier.
For example, the application may show a new phone number that has just been established. The phone may not be connected to anyone else. It may have no connection to WhatsApp or Telegram. The decision engineer considers how these details connect the application to the later collections outcome.
This way of thinking connects the whole process. It follows the causal chain of events and shapes the decision so the outcomes are delivered well. That is the difference between a classical data analyst and a decision engineer.
What goes wrong when a lender has only data science or AI?
A lender can analyze the data already in its systems and use the technical capabilities available. It can also put the data into an LLM and get a model. The model may be technically capable and still have no relation to reality.
Finding that connection requires someone to poke and probe, stay present in the product, talk with the product team, and think about how more data can be collected. This work involves gaining situational awareness and thinking about what the data would look like if something were true.
If an outcome is supposed to change, there should be a lever. The work is to find the data that indicates the lever and decide how to collect it. If there is no primary data, proxy data may provide a source that is reasonably correlated with the practical data.
How can a decision engineer check whether someone really lives there?
You can ask someone whether they live at an address. They may provide an electricity bill as proof of address. That shows they have a connection to the address, but it does not show whether they really live there in person.
You can look at activity connected to the address. Do they get food delivered there? Do they get packages delivered there? This activity can indicate that someone spends enough time there to have a reason to receive deliveries there.
This is proxy data for someone living somewhere. It may not be the official data given to a bank or employer, but it can indicate what is happening in the world. The decision engineer asks what data can show that someone really lives at the address, then considers how to collect that data.
What skills does a decision engineer need as technology changes?
LLMs, Claude, and OpenAI ChatGPT make it possible to analyze data much faster. A person still has to think about the product, probe the product, and ask what the data means in reality. That thinking remains the most important part.
Improvement comes from someone poking around, asking questions, and understanding the answers. It requires systems thinking and the ability to communicate with other people. Technology comes after that work.
In the past, a decision engineer needed statistical skills to run R, then Python, and wrangle data as a data scientist. Those hard skills are no longer necessary. What remains necessary is clear thinking about the levers in a situation, how they shape the final outcome, and how decisioning can improve it. The most important work is philosophical probing: connecting assumptions and hypotheses with reality, and doing good tests.
- Questioning the product and data: Probe the product and ask what the data means in reality before using technology.
- Systems thinking: See how levers shape outcomes and how decisions can improve them in context.
- Clear thinking: Hold conflicting hypotheses together, explore each one, and avoid becoming attached to any single idea.
- Communication across teams: Explain technical concepts clearly so non-technical stakeholders can understand them and act on them.
- Situational awareness: Understand where decisions run, who they affect, and the operating context that keeps models relevant.
- Problem solving: Turn business problems into clear decision logic by working across teams.
- Presenting decision models: Present models, assumptions, and results clearly so teams can implement them correctly.
- AI and machine learning: Understand core principles and when to apply them to decision models.
What should you look for when hiring a decision engineer?
A good decision engineer keeps testing the model against reality. A consultant, product manager, or marketing person may try to present something that looks nice and make it appear true. A decision engineer keeps asking what it would look like if they were wrong, and what it would look like if they were correct.
This matters with scoring models. Some attributes have high correlations and look useful, but they are not connected to a lever that explains credit behaviour or repayment behaviour. Those attributes may stop working when conditions move, and you do not know when that will happen. A good decision engineer looks for the lever behind the correlation and tests whether the connection holds.
A good statistician may say they are quite sure because the statistics show that the model works. A decision engineer can show what is connected to the lever, what will happen if the lever moves, and how the data connects to the phenomenon.
How do you test a decision engineer candidate in an interview?
Give the candidate a problem with a strong, clean correlation built into the data, but do not tell them that the correlation is false. The cause should be in how the data was produced. One example is leakage, such as a “contacted by collections” flag that is filled in after the outcome and predicts default perfectly. Another is a process artifact, such as applications submitted at night defaulting more often because one broker sends its batch at 2am.
Someone who wants to give a nice answer will show that there is a nice correlation. A good decision engineer will say they do not understand how the correlation can be so strong because it makes no sense in reality. The interviewer should also include one correlation that is real and has a sensible mechanism. This shows whether the candidate can tell the two apart instead of doubting everything.
There are people who want to show that they did the job even though there is no real correlation. The task is to let the candidate explain how they think about the problem and whether they can find the lever behind the data.
Questions about a decision engineer
How do you become a decision engineer?
Start by learning to connect assumptions and hypotheses with reality, then test those connections. You need clear thinking about the levers in a situation, how they shape the outcome, and how decisioning can improve it.
Does a decision engineer need to know how to code?
In the past, decision engineers needed statistical skills to run R, then Python, and wrangle data like a data scientist. Those hard skills are no longer necessary, because technology comes after thinking about the product, the data, and the levers behind the outcome.
What does a decision engineer work on each day?
A decision engineer talks with the product team, asks what the data means in reality, and probes edge cases. They test decision models against expectations, adjust them as needed, and think about how more data can be collected.
What makes someone a good decision engineer?
A good decision engineer keeps testing a model against reality and asks what it would look like if the model were wrong. They look for the lever behind a correlation, test whether the connection holds, and explain the model clearly so other people can act on it.
