• Written by Azhar Aliyar, Director

A typical operations dashboard has a dozen tabs, each with its own set of metrics and alerts, but the real challenge lies in taking action based on that data, rather than just displaying it. This is where agentic AI workflows come in, allowing organisations to automate complex processes that involve multiple steps and systems.

What separates an agent from a chatbot

An agent is distinct from a chatbot in that it uses tools, maintains state, and has a goal. A chatbot, on the other hand, is designed to answer a single question or provide a simple response. In an enterprise setting, an agent might use a combination of systems, such as ERP and CRM, to complete a task, whereas a chatbot would simply provide information from one of those systems.

For example, a shift supervisor checks the maintenance log and sees that a piece of equipment is due for repair. An agentic workflow could then use the ERP system to generate a work order, the CRM system to notify the maintenance team, and the scheduling system to assign a technician. This level of automation and integration is what sets an agent apart from a chatbot.

Where a human approval step belongs

Most people get wrong the idea that human approval steps are always unnecessary and should be removed to increase efficiency. However, this is not always the case. In many situations, a human approval step is still necessary to ensure that the agent is taking the correct action. Removing this step prematurely can lead to errors and unintended consequences.

A human approval step belongs in situations where the agent is making a decision that has significant consequences, such as cancelling an order or changing a customer's account information. In these cases, it is still important to have a human review the decision to ensure that it is correct and appropriate.

How an agent fails differently from a model

An agent fails differently from a model in that it has a greater potential for cascading failures. If an agent fails, it can cause a chain reaction of errors that affect multiple systems and processes. This is because an agent is designed to take action and make decisions, rather than simply providing information.

When an agent fails, it is important to log the error and the circumstances surrounding it, so that the issue can be diagnosed and corrected. This might involve logging the input data, the decisions made by the agent, and the actions taken. By doing so, organisations can identify the root cause of the failure and take steps to prevent it from happening again.

The cost of an agent loop that retries silently

An agent loop that retries silently can be costly, as it can cause the agent to become stuck in an infinite loop, repeatedly attempting to complete a task that is failing. This can lead to wasted resources, such as CPU time and network bandwidth, and can also cause other processes to be delayed or interrupted.

For example, if an agent is designed to send a notification to a customer, but the customer's email address is invalid, the agent may retry the notification repeatedly, causing unnecessary resource usage and potentially causing other notifications to be delayed. To avoid this, organisations should implement logging and monitoring to detect and prevent silent retries.

Logging and monitoring

Logging and monitoring are critical components of an agentic workflow, as they allow organisations to detect and diagnose errors, as well as monitor the performance of the agent. By logging key events and metrics, organisations can identify areas for improvement and optimise the agent's performance.

Organisations can use tools such as our platform to log and monitor their agentic workflows, and to identify areas for improvement. By doing so, they can ensure that their agents are running efficiently and effectively, and that they are achieving their goals.

When to use agentic AI workflows

Agentic AI workflows are not always the right choice for every situation. In some cases, a simpler approach, such as a chatbot or a scripted process, may be more appropriate. Organisations should carefully consider their needs and goals before implementing an agentic workflow.

For example, if an organisation is looking to automate a simple process, such as answering frequently asked questions, a chatbot may be a better choice. However, if the organisation is looking to automate a complex process that involves multiple steps and systems, an agentic workflow may be more suitable. To discuss your specific needs and determine the best approach, contact us at our contact page.

What this does not do

Agentic AI workflows do not replace human judgement and decision-making. While they can automate complex processes and make decisions based on data, they are not a substitute for human oversight and approval. Organisations should always ensure that their agentic workflows are designed and implemented with this in mind.