Welcome to the Agentic AI Organization!

ariverse
Nov 30, 2024
7 min read

Updated: Dec 16, 2024

USE CASE:

Peter, a Purchasing Manager in a corporation, logs into his PC and starts his Purchasing Agent.

Agent: “Good morning, Peter. What tasks would you like me to perform today?”

P: “Please provide a list of all invoices that are overdue by two days.”

A: “Here is the list of overdue invoices: 25 from the IT department totaling X million euros, 10 from the Facilities department amounting to Y thousand euros, and 30 on your desk totaling Z million euros.”

P: “Who is the business owner for the IT department?”

A: “The owner is z@email.com. Would you like me to send them a reminder?”

P: “Yes, please.”

*The Agent creates a list of the invoices with their links and sends a reminder to the owner at z@email.com.*

P: “Show me the list of my invoices.”

*The Agent retrieves the actual invoices from the database, compares them to their purchase orders in SAP, and generates a summary with links to the invoices.*

A: “Here is your list: 10 invoices are over 1 million euros, 15 are recurring and identical to last month, and 5 invoices with changed payment terms were automatically sent to Business Owners for their review.

P: “Approve the 15 recurring invoices.”

*Peter reviews the details of the 10 invoices that are over 1 million euros as they necessitate human oversight.*

*The Agent approves the 15 recurring invoices on behalf of Peter and sends notifications to the 5 Business Approvers.*

If software eats the world, AI will eliminate mediocracy. Agentic AI has the potential to fundamentally transform how we work and organize ourselves, both professionally and personally. Agentic AI goes beyond Generic AI and is not merely a “co-pilot". It is the way that organizations will work in the coming years and for many years to come. It’s the most revolutionary technology to build and operate organizations (profit or non-profit). And because of that it’s not a technology to be left to technology departments but to be owned by the C-Suite.

What is Agentic AI?

Agentic AI refers to the technology in which agents execute tasks using generative AI, machine learning, and other technologies, with or without human intervention (Human in the Loop, HITL). This "intelligent" service is capable of obtaining data from predefined system sources, applying reasoning to interpret the data, presenting it in an intuitive manner, communicating with the user through natural language, and either responding to user commands or autonomously performing actions on other systems or agents.

Agentic AI Characteristics

Data Access: Agents can access predetermined databases or source systems according to specific credentials and criteria, such as document libraries, data lakes, and ERP systems.

Reasoning: Agents are capable of responding to user or other agent inputs and making logical decisions. This is where large language models (LLMs) or small language models (SLMs) play a significant role.

Prompting: Agents should be "programmed or trained" with explicit guidelines on dos and don'ts. They should possess the ability to take deterministic actions in addition to AI-based actions.

Security and Agent Rights: Agents function similarly to users; they should have defined agent permissions, and the users interacting with them should also have certain rights.

Integration: Agents should have the capability to integrate with other applications through graphical user interfaces (GUI), interaction with other agents, Robotic Process Automation or by calling APIs.

Planning: Agents should be aware of the sequence of upcoming events within the process.

Memory: Agents should possess both short-term and long-term memory. Short-term memory is utilized for interactions with users, such as chatting. Long-term memory involves understanding past right/wrong decisions to inform future choices and the ability to retrieve previous user conversations. For example, an agent can remind the user of a previous discussion on a specific topic.

UI in Natural Language (Optional): Agents should be able to comprehend input from users in natural language, respond accordingly, and act upon their commands. This interface can operate across various channels such as Slack, Teams, apps, etc.

Why is Agentic AI so Important?

Software has significantly contributed to the automation of many repeatable tasks. Organizations employ software to manage tasks and collect data, enabling humans to make more informed decisions, act swiftly, enhance productivity, and reduce costs. However, in most instances, it is still humans who input data, request data, interpret data, make decisions, and take actions.

AI Agents are now emerging to realize the true potential of artificial intelligence for organizations (and our personal lives as well). These agents possess the capability to input, retrieve, interpret, predict, make or suggest decisions, and act on both unstructured and structured data sources. Importantly, users can interact with AI Agents using natural language AI Agents will start as intelligent assistants (similar to personal assistants) and will eventually evolve into a system of agents managing self-regulated processes.

The future state of agent-self-regulated processes is set to transform the operations of service-driven organizations, ultimately impacting the economy and society. Agentic AI is anticipated to introduce a fundamental shift comparable to the advent of the internet and mobile phones in our professional and personal lives.

Potential Agentic AI technology Advantages

Reasoning: Agents are capable of making logical suggestions, conclusions, and decisions within the parameters defined in their job description (see AR blog).

Natural Language Interface: This is the most user-friendly method of interfacing with technology. Over time, agents will become the primary interface between users and various applications.

Orchestration / Integration: They have the ability to integrate with systems or orchestrate actions in collaboration with other agents.

Action: They can execute actions on applications and systems in a deterministic manner.

Learning: They can be trained, learn from user inputs, and achieve self-regulation (cybernetics).

Prediction: They can make predictions based on data and input from agents or users.

Planning: Once they understand the subsequent steps in a process, they will be able to plan these steps and request input from users or other agents.

USE CASE:

Maria, the Manager, contacts the Forecast Agent to prepare a forecast presentation for an urgent sales meeting.

Maria: "Please build me a forecast for Q3 on Products XYZ, including all deals weighted above 40%."

*The Agent accesses the CRM and retrieves the 100 deals above 40% that are due in Q3.*

Agent: "The forecast is X million dollars."

M: "Which are the top deals?"

A: "The top 5 deals are these" — *provides a list of deals along with their respective owners.*

M: "Please ask the owners for their comments and confirmation that these deals are accurate and will close in Q3."

*The Agent opens an MS Teams session and communicates with each owner, requesting they enter their comments on the Q3 deals in the CRM.*

A: "Comments have been updated by the owners. All have confirmed the deals except John, who has moved deal Y of $200k to Q4."

M: "Alright, please create a presentation with the forecast and the top deals, including the comments."

*The Agent utilizes Office Co-pilot to prepare the presentation using a standard forecast format PowerPoint, adhering to Maria's instructions. It also includes a risk analysis based on historical data that another Agent processed with Machine Learning*

The above advantages will provide users and organizations the following capabilities:

Increased Productivity: This is evident and represents the core benefit that all tools have historically provided. For example, an electric screwdriver allows you to fasten bolts more efficiently, though it does not inherently make one a carpenter.

Instant Access to Information: AI Agents will have access to vast amounts of data, enabling them to locate, retrieve, and summarize information intelligently. However, simply reading instructions on building a wooden desk does not qualify one as a carpenter.

Augmentation/Enhancement of Capabilities: Agentic AI will fundamentally transform our work processes by enhancing human ability to perform tasks in various ways. It will also enable organizations to operate more intelligently. The standardization of the production of wooden desks on an assembly line is replacing traditional craftsmanship with mass production.

In our example, just as mass production replaced the average carpenter and improved the efficiency and quality of producing wooden desks, Agentic AI will enable users to run processes more effectively and economically. Craftsmen will focus on handling difficult cases, managers will oversee the process and verify outcomes, engineers will develop the next Agentic Process, and office employees who previously performed repetitive tasks on a PC will likely transition to roles that require deep human (human 2 human) interaction.

NOTE: OUR VIEW OF CO- PILOTS VS. AI AGENTS…

Co – Pilots are smart crutches to the user of an application…

Co-Pilots are AI-based software integrated within applications (e.g., Office) designed to enhance user experience by making it smarter and more efficient. These tools significantly benefit users by increasing productivity, fostering creativity, and improving skills such as writing competence.

This series of blogs was written in MS Office with the assistance of Co-Pilot. The Co-Pilot feature expedited our work with auto-fill, improved our English, and restructured original documents according to the authors' preferred style – so thank you for this!

However, AI Agents, as previously defined, offer capabilities beyond mere information retrieval or application productivity enhancement. They can intelligently manage processes across multiple applications, enabling fully automated workflows, which we refer to as Intelligent Process Automations (IPA). Eventually Agentic AI driven Processes will combine several technologies, including co-pilots, as to create as much as possible, or allowed, Agentic-Regulated processes.

Note: This series of blogs were written before the Microsoft Ignite 24. In that event Microsoft repositioned Co-Pilots as the UI of the Agents. So, in this scenario Co-Pilots and AI Agents co-exist and complement each other.

USE CASE:

Upon receipt of a response to an RFI, an Agent processes the email accordingly. The Agent then forwards the response to a Document AI Agent to extract all needed information. This data is subsequently aggregated into a DataLake, enabling the Agent to compare responses for each RFI, retrieve user-requested information, and summarize the responses. The Purchasing Officer reviews the summary and initiates a series of in-depth negotiation meetings with the potential vendors.