Here is how I see the AI world. There are three types of AI workflows and those are Rules Based AI, Generative AI and Agentic AI:
Rules Based AI is also sometimes called deterministic AI is AI based on a predefined set of inputs and a predefined set of corresponding outputs. These are the standard workflows in CRM, ERP and Accounting systems that follow the simple flow chart form triggered on a field value change that update a record or reach out through an API or reach out to send a simple email. A Lead comes in, it is synced to the CRM, and email is sent to the lead, the Lead is assigned and the new Lead owner is notified. When a proposal is sent, a date field is set. When the date field plus 5 days comes to pass, the Lead owner is notified and a follow up email is sent. All of these are rules based types of workflows.
Generative AI Workflows – Imagine a support chat workflow. It must know about language. It must know about intention and key business “domain” logic and what to do. It is a combination of the “large language model” (“LLM”) and “domain knowledge” (referred to as RAG) about your industry and your company. It needs to speak/reply and is often considered to be probabilistic AI since based on the inputs, it determines the highest probably appropriate response.
Agentic AI is to respond to entirely different things than it has ever specifically seen before. Think self-driving cars, fraud detection, cold calling, inbound telemarketing, more advanced chat bots. Agentic AI is the type of AI most commonly associated with AI “hallucinations” where AI produces totally wild outcomes. Often, these types of AI models need to have “guardrails” to keep the outputs within a certain range or certain set of responses. This is based on a model developed from looking at the data whether this is the LLM or the company’s actual data (RAG) or a combo of RAG and LLM. As the data underlying the model changes and as the AI system changes these workflows evolve and change. That is when you can get different results different times you run the workflow.
Now to the point. Agentic and Generative systems are non-deterministic workflow. The output isn’t necessarily known at the time of creation because both do not assume to know all possible inputs at the time of creation. Generative takes a more simplified approach and Agentic takes a more forward looking, more autonomous with simulated reason and adaptation. This implies a lack of predictability: Sometimes you may get a good result, sometimes not. It models it blindly, even with RAG, because the model doesn’t know all the domain knowledge. Sometimes this is advantageous, sometimes it is non-advantageous but it doesn’t know everything about the business that a human does and sometimes that bias is an advantage. Let’s say that error rate is 5-10%. The larger and more complex the workflow, the more random decisions are made the more the that error rate impacts. It is like rolling a dice 1 to 6. The workflow is great unless the die rolls a 1. The more rolls, the greater chance of landing on a 1.
On the positive side, the non-deterministic workflows runs until final disposition in all cases. By comparison, a complex human manual workflows, the gotcha is that the workflows get abandoned and are not run until final disposition plus there is an error rate in human execution of the workflows. Let us not forget that an automated workflow is going to be a lot cheaper to run than a human workflow. So, an AI workflow may have a higher error rate than a human but it is cheaper and it runs to final disposition in all cases. Often, this is a higher ROI but not always in your particular case.
In summary:
Positives
- AI workflows process to completion without abandonment in all cases which certainly cannot be said of human driven workflows.
- AI workflows surface insights and workflow optimization opportunities where rules based workflows would be difficult or impossible to find and very difficult to code
- AI workflows can handle workflow “branches” that are not definable (think chatbot example) in advance.
- More cost effective than humans and runs around the clock and can be optimized for best time of day without stressing staffing models
Negatives
- Stochastic error rate that compounds with greater complexity
- Inconsistent responses
- AI hallucinations
- An attitude of “hands off” abdication in management takes management away from the actual business activities
Atlantic CRM & Cloud Solutions is an expert in all types of business workflows automating a range of our clients processes. Let us know how we can help.
