AI agents have not superseded business rules. They both have a place, and can work together in a hybrid system where rules handle clear-cut cases and agents handle messy or context-dependent cases.
Automate routine, well-defined decisions with a business rules engine, then escalate ambiguous, unstructured, or context-dependent cases to an AI agent. Evaluate the agent's recommendation with deterministic guardrails and human oversight when the stakes exceed a defined threshold.
Behind this: 12 build steps · 6 tools and how each is used · how to validate demand · 6 more real examples · 6 things the video never answers.
Searchable transcript of AI Agents vs Business Rules: Which Should Make Decisions? — IBM Technology (10:25). Search for a phrase, then click its timestamp to jump straight to that moment in the video.
Captions sourced from the original video on YouTube, published by IBM Technology. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.
00:00 Both business rules and AI agents are ways to automate a decision. So maybe that's to approve a refund or maybe to flag a transaction, but they go about this in different ways. So my question to you is this. Have AI agents superseded business rules or do they both have their place? And what's the difference between these things anyway? Well, let's start with business rules.
00:29 And business rules run in a business rules engine. That has logic that a person wrote down explicitly, a person like me. So here's a business rule. We could say, if we have an order that is less than 30 days old, and we also have a condition here, which is the item. Was not marked as a final sale. So if and, and we've got a then, the then is we would actually approve the refund.
01:16 That is a business rule and that rule has conditions and an action. Now an AI agent doesn't necessarily work from rules that are written down. It works a bit differently because with an AI agent you give it a goal and you give it some context of the situation and you also give it access to some tools that it's allowed to call. And the agent works out the steps itself.
01:47 So instead of providing rules for a refund it's more like hey here's the customer and here's the order policy and here's the customer's order history. Now you sort it out Mr. AI agent, it's up to you. So which one should be making the decision? Well, one thing about a rules engine is you get a consistent and predictable answer. It checks against conditions, against facts.
02:17 So if the fact is that an order let's say is 12 days old, that's a fact. And if there is a fact that the order was not marked as a final sale, well then the condition comes from the rule. It's under 30 days, it was not a final sale, both are true. So the rule resolves as approve this condition, approve the refund. There's not any guesswork here because it's simple boolean logic.
02:49 The order is either under 30 days or it is not and that's what deterministic means, the output is a fixed function of the input and business rules are deterministic. Now AI agents don't work like this because they are built with large language models and we know that a large language model predicts next tokens. So given the order and the policy and the goal, the model works out a response from patterns that it's picked up in training and it's doesn't necessarily fix on one final answer.
03:29 It's working from a probability distribution over possible responses and picking from it. So for something like this, approve is still the most likely outcome, but deny as an outcome still has some probability as well. And that makes AI agents probabilistic. Whereas the rule engines evaluates a fixed condition. The agent picks from a range of likely answers, which is why you can run the same request twice and get different calls.
04:01 So that's the split. Rules engine, deterministic, the agent, probabilistic. So when to use each one? Well, business rules are used when the decision is well-defined, meaning you can define the logic ahead of time. So the refund policies are pretty good example of that. But then so is something like a loan eligibility check or an insurance pricing tier.
04:27 Anything where the inputs are structured and you already know what the answer should be. And then in regulated work, there's another reason as well. The rule that was processed is by its very definition, also the explanation. A rules engine points at the exact condition that triggered the decision. So that is great for audits. And there's also determinism which helps with testing.
04:54 Each branch can be unit tested and it will behave the same way in production. It also runs without a model call so there is no AI inference cost. So there's lots of advantages here to business rules. Now AI agents are for when the logic cannot be written down ahead of time because the result all depends upon understanding the context of a situation.
05:21 Now the input can be messy or it can be unstructured or maybe there's just a case that nobody anticipated. So should a refund actually be approved? If it turns out I sent back my order of dumbbells for being too heavy, which is arguably the point of dumbbells. That sort of thing is the wheelhouse of AI agents judgment, judging the merits of a reason for a return.
05:49 That's not really something that's easily captured in deterministic pre-defined rules. So look, that's the the trade-off. A rules engine only handles what was anticipated. It will struggle to handle an input that nobody has a rule for, whereas an agent will generalize because it's working from patterns instead of a fixed branch. So back to the question we started with.
06:17 Have AI agents superseded business rules? Well, in a lot of real systems, the answer is actually to run the two together. It's just a hybrid approach where you end up using both. So let's fill in this flow with a hybrid solution so we can see what it might look like. So at the top here we start with the request which has come In. And the first thing in the process flow is we're going to navigate to the rules engine.
06:52 So we'll come down to here. Now why rules first and not an agent first? Well, because rules are comparatively cheap and they're quick to run. So the rules engines checks the request against the policy and then it will determine an action, basically what we should do. So the order is 12 days old, it's not a final sale. That would mean it's under the auto approve limit.
07:19 And we are going to go down here to the final decision. And that final decision will be to approve that request. That's nice and easy. And if something is clearly outside of the policy, then instead of approving that request, we would decline the request. But that's the decision right there. We're basically done. And most refunds and decisions are gonna be clear cut like this.
07:49 It's only when the rules can't decide that the request instead will go down this pass where it will be escalated and it's escalated to an AI agent. And there's usually a couple of reasons for that kind of escalation. Now the first is that the data itself is going to need some additional work. So a rule needs clean structured facts to check like an order age or a final sale flag, but plenty of refund requests they're going to turn up with data that's not all that clean.
08:24 They're going have a bunch of unstructured input like a a paragraph of free text or maybe a picture. Rules engines they by and large work best with structured data, and the agent can also make use of tools as well. Now it might call a vision model to look at that photo of a defective blender or to to query the order system for the customer's purchase history for example, but the agent will work out which tools the situation calls for and in which order those tools should be called and it does this without being
08:59 scripted in advance. So at that point an agent can recommend a decision of its own, but it might not get the final word because its recommendation is then going to go down to a set of deterministic guardrails where it will then be evaluated. So a guardrail might specify where anything the agent green lights above a certain threshold should actually route somewhere else.
09:28 It should go first of all to this box here which is human in the loop before the final decision is ultimately reached. It's under a certain threshold it can go directly to the decision. So we have the potential to check the agent's judgment with a real person in situations where the stakes are highest. So no, agents have not superseded hard-coded business rules, but they can work pretty well with them particularly when the data is messy or when we need judgment or where tool calling can expand the context.
10:08 Ultimately, the two together can help make better decisions.