There is one question we need to ask before anything else when a team wants to bring AI into a process. It is simple, it takes minutes, inspires good conversation, and it prevents some of the most expensive mistakes with AI.
For this step, do we need exactly the same answer every time given the same input, or do we need good judgment in a range of situations?
That is the difference between deterministic and probabilistic work. Not discussing these expectations and analyzing this is where AI teams struggle on the details of where to implement AI.
What the Terms Mean in Plain Language
Deterministic work has one right answer that follows from clear rules. Is the form complete? Does the amount exceed the limit? Is the date within thirty days? Traditional software and simple automation handle this well. You want it to behave identically every time.
Probabilistic work involves interpretation. Which permit type does this unusual request fall under? What is this customer really asking? Is this exception reasonable? AI models excel here, because they can weigh context and ambiguity. They also produce varied answers and can be confidently wrong.
Most real business processes mix both, often within a single step.
Why This Matters So Much Right Now
When AI tools became easy to use, many teams started applying them everywhere. That creates two failure modes.
The first is using AI for deterministic work. A model checking whether an expense is under a policy limit might get it right 97% of the time. For a rule with a clear answer, 97% is a failure rate nobody would accept from a simple calculation.
The second is forcing deterministic rules onto probabilistic work. Teams write hundreds of brittle rules to cover situations that really need judgment, and the system breaks on the first case nobody anticipated.
BAs are perfectly positioned to prevent both, because we already break processes into steps and understand the business rules behind each one.
Tip 1: Walk the Process Step by Step
Take any process and list its steps. For each, ask:
- Is there a clear rule that always produces the right answer?
- Does it require interpreting language, context, or intent?
- What happens if the answer is wrong, and can it be reversed?
You will often find a single step needs splitting. “Classify the permit type” might be deterministic for standard requests and probabilistic for unusual ones. That split is where good design lives.
I use this exact exercise with a municipal permit process in my courses, and the conversations it creates are some of the richest of the whole cohort. There are many right answers depending on assumptions. The thinking is the point.
Tip 2: Pair the Type With the Right Controls
Once you know the type, the design choices follow.
- Deterministic steps: encode rules explicitly, test them exhaustively, and let AI call them as tools.
- Probabilistic steps: define confidence thresholds, sample outputs for review, and design escalation for low-confidence cases.
- High-stakes probabilistic steps: AI recommends, a human decides.
This gives engineers a clear brief and gives business owners confidence that the design matches the risk.
Tip 3: Make the Classification a Shared Conversation
Do not do this analysis alone. The classification of each step is a joint call. Business owners understand the risk appetite. Technical teams know what the AI can do reliably. The BA holds it all to the outcome.
Run it as a working session. Put the steps on a board. Let people disagree. The disagreements are exactly the assumptions that would otherwise surface in production.
A Small Question With Big Leverage
This one question reshapes how teams think about AI. It moves the conversation from “where can we use AI?” to “where should we use AI, and how?” That is a much better conversation, and BAs can lead it starting this week.
Practice It With Real Scenarios
My AI/BA course series includes hands-on deterministic versus probabilistic analysis on real process scenarios, plus how to turn that analysis into agent design.
Visit www.maven.com/angela-wick to explore current courses and upcoming cohorts.
Ask the first question first. Everything else gets easier.