Say ‘hi’ to your new virtual team members: AI agents
Dr Philipp Herzig, chief technology officer at SAP, explains why agentic AI is set to take us to the next era of business productivity
5 min read

Breaking down a complex task into components, each one assigned to different specialists or teams, is nothing new. This clever division of labour has supercharged industries since the start of the 20th century, from Ford’s automotive manufacturing to fast food and software development.
Specialisation brings greater efficiency and innovation as each specialist completes their part before handing the task to another. However, it can be challenging to achieve the full benefits when departmental silos and scattered data make collaboration difficult. Organisations must also overcome skills shortages or gaps between roles and available workers when filling specialist positions.
"Agentic AI is set to transform what can be achieved through specialisation, as well as changing how humans interact and work with artificial intelligence."
Agentic AI is set to transform what can be achieved through specialisation, as well as changing how humans interact and work with artificial intelligence. Collaborative AI agents working autonomously across business functions will drive new levels of productivity. Analysts estimate that agentic AI in enterprise software applications will grow from less than 1% in 2024 to 33% by 2028, enabling autonomous solutions for 15% of day-to-day work decisions.
Getting the benefits of agentic AI
One of the defining features of agentic AI is proactivity. Agents can autonomously perform tasks with multi-step self-conceived workflows. For example, an agent specialising in retail could rearrange deliveries without human intervention in response to weather-related disruptions or material shortages. The agent can decide whether to proceed, conduct more research or validate data before continuing towards its goal. Agents that work 24/7 reacting to events will adapt and develop new approaches or strategies if circumstances change.
If this sounds like the way you would reason, it’s because it is. Ensuring that an AI agent can reason requires context. This enables it to consider more variables when making decisions. Data provides context and an understanding of your business: if you want AI agents to work with the latest data across the entire business, they need access to it. Agents that work in data silos will create domain-specific progress, but will fail to drive enterprise-wide impact.


Metadata—the who, what and when—helps an agent to determine whether data should be considered, or whether it is outdated or irrelevant and should be ignored. Knowledge graphs tie everything together by exposing metadata, making relationships between information accessible. In this way, an AI agent resolving a payment dispute knows the relationships between vendor and buyer, purchase orders, etc. Knowledge graphs also provide a grounding for data and the AI agent’s reasoning by connecting it to real-world facts. Retail agents need data connected to real-world context to suggest products to customers based on, for example, purchase history and demographics.
Agents also need a way to understand how different pieces of information depend on one another, so that they can consider a task’s wider context beyond a specific component. This is conveyed through a knowledge graph that captures process outlines, like a handbook for AI agents. If you book a holiday, the agents will ask for the destination and dates before the graph points them to flight booking and hotels.
This structured way of inferring new information and making logical connections will help AI agents to reason and derive conclusions. Knowledge graphs also offer the benefit of scale. Write this down only once; you won’t have to rebuild it for each new agent.
The multi-agent journey
The benefits of embedding non-agentic AI across the supply chain have already been staggering for distributors—offering reductions in inventory levels, logistics costs and procurement spending. Agentic AI will build on these benefits. Techniques such as multi-step reasoning and the ability to autonomously run tasks make it perfect for dynamic fields such as supply chains, enabling tasks considered too intricate or with too many interdependent steps to be automated. Non-agentic AI might use historical data to reorder stock, which could lead to overstocking or stockouts when unexpected spikes occur. Agentic AI, by contrast, can continuously analyse a wide array of data inputs to adjust demand forecasts in real time, while collaborating with other agents to optimise logistics or adjust employee shifts.
"Techniques such as multi-step reasoning and the ability to autonomously run tasks make it perfect for dynamic fields such as supply chains, enabling tasks considered too intricate or with too many interdependent steps to be automated."
This ability to collaborate and bridge specialisations offers benefits that are not available to isolated systems. Take the interaction between manufacturing, customer service and the supply chain. For manufacturing organisations to forecast costs accurately, they need to understand the supply chain. Suppliers, material costs and transport all affect manufacturing costs. AI agents can consider all of this in real time, while data from sensors on machines and components help to predict the timescale of repairs and production outages. Passing all this information to customer service agents allows them proactively to contact customers about delivery status, inform them if their delivery is delayed and offer a bonus to keep them happy.

Where are we now
Agentic AI is still developing and, despite its potential, there are challenges. AI agents are only as powerful as the data they are grounded in, so organisations must ensure they have access to high-quality unified data. Isolated agents cannot collaborate or drive cross-company productivity.
"AI agents are only as powerful as the data they are grounded in, so organisations must ensure they have access to high-quality unified data."
Practical goal setting remains just as crucial for agentic AI as for a traditional workforce. You must consider what goals you set for agents and under what conditions they can determine their own goals before human intervention. Agents can make mistakes, especially as they learn. Ensuring there are guardrails in place can mitigate risks.
The promise of agentic AI reads almost like science fiction. AI that can work and act independently to alleviate complex and routine tasks can feel like a far-off prospect. However, if correctly implemented, the benefits are a not-too-distant reality—and science fiction becomes real.

Author: Dr. Philipp Herzig, Chief Technology Officer, SAP
