What Actually Is an AI Agent? (And Why It Matters for Procurement)

If you’ve read anything about AI in procurement lately, you’ve heard the term “AI agent” thrown around. Marketing materials promise “agentic AI” will revolutionise your work. Tech vendors claim their platforms now have “AI agents.”

But what actually is an AI agent? How is it different from the chatbots we’ve been using for years? And why does the distinction matter?

NVIDIA CEO Jensen Huang has the clearest answer: AI agents are “information robots”—digital employees that work alongside you, handling tasks autonomously whilst you focus on strategic decisions. Not someday. Right now.

Let me break down what that actually means, in plain language, grounded in how the companies actually building these systems describe them.

NVIDIA CEO Jensen Huang has the most vivid way of describing AI agents: he calls them “information robots.”

Here’s why that framing is brilliant. Huang explains that there are actually two types of robots, and they both work the same way:

Physical Robots (what we traditionally think of): Humanoid robots, robotic arms in factories, autonomous vehicles. They have bodies and manipulate the physical world.

Information Robots (AI agents): These robots live inside the computer. They process data, understand context, and perform actions within digital environments. Instead of moving boxes in a warehouse, they move information through systems.

Both types use the exact same operational loop: Perceive → Reason → Plan → Act. The difference is the environment they operate in and what they manipulate.

For procurement professionals, the robots that matter most right now are the information robots. They’re already here, already working, already changing how we operate.

But let’s be more precise. Here’s how the major AI companies define these information robots:

OpenAI: “Agents are systems that independently accomplish tasks on your behalf. Agents use an LLM to execute instructions and make decisions.”

Anthropic: “Agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

NVIDIA: “AI agents are advanced AI systems designed to autonomously reason, plan, and execute complex tasks based on high-level goals.”

Perplexity: “AI agents are AI assistants capable of autonomously pursuing user-defined goals by planning and taking multi-step actions on a user’s behalf.”

Notice the common thread? Autonomy. Information robots don’t just respond to prompts. They act independently to achieve goals.

Think of it this way:

Or, as Huang puts it: they’re digital employees working alongside you.

I love this way of thinking about AI Agents.

AI agents operate through a continuous cycle of four key activities. Understanding this cycle is essential to understanding what makes them different from previous generations of AI tools.

AI agents gather and process data from various sources to understand their environment and the task at hand. This involves:

In procurement, perception might mean:

Unlike traditional systems that only see data you explicitly feed them, agents can pull information from multiple sources simultaneously. A procurement agent might be monitoring your contract repository, vendor portal, email inbox, compliance database, and market data—all at once—looking for signals that require action.

At the heart of every AI agent is a Large Language Model (LLM) acting as the orchestrator or reasoning engine. Think of it as the brain.

The LLM’s job is to:

This is where Retrieval-Augmented Generation (RAG) becomes critical. The agent doesn’t just rely on what the LLM was trained on. It actively retrieves relevant information from your proprietary data sources—contract standards, vendor policies, compliance requirements, historical decisions—to inform its reasoning.

In procurement, reasoning might involve:

This isn’t simple if/then logic. The agent is actually reasoning through scenarios, considering options, weighing trade-offs, and making judgement calls—much like a junior analyst would, but informed by your organisation’s complete documented knowledge.

Memory Modules (from NVIDIA’s architecture):

Tools and Integrations (from OpenAI’s framework):

The LLM coordinates all of this, deciding what information to retrieve, which tools to use, and how to synthesise everything into a plan of action.

After perceiving the situation and reasoning through options, the agent takes action. This is where agents differentiate themselves from chatbots most clearly.

By integrating with external tools and software via Application Programming Interfaces (APIs), AI agents can execute tasks based on the plans they’ve formulated.

In procurement, actions might include:

Critical components for action (from OpenAI’s framework):

Tools: Functions and APIs the agent can invoke. In procurement, this might include:

Guardrails: Safety checks that prevent the agent from taking inappropriate actions:

Handoffs: The ability to transfer control to human experts or specialised agents when needed:

The agent doesn’t just tell you what should happen. It makes it happen—within the boundaries you’ve defined.

This is where AI agents truly differentiate themselves from static automation. Through what NVIDIA calls a “data flywheel,” agents continuously improve their performance.

Here’s how it works:

Data Collection: Every action the agent takes generates data—what worked, what didn’t, how long it took, what decisions were made, whether humans intervened.

Feedback Integration: This data, combined with human feedback (corrections, approvals, rejections, ratings), feeds back into the system.

Model Refinement: The system uses this information to:

Adaptive Behaviour: Over time, the agent gets better at:

This learning happens in two ways:

Immediate adaptation: The agent uses feedback from its current session to adjust its approach. If it routes a contract to the wrong approver, it learns from the correction and applies that knowledge immediately.

Systematic improvement: Aggregated data across all users and interactions is used to improve the underlying models, making the agent smarter for everyone over time.

Important: This isn’t the agent making up its own rules. The learning happens within the governance boundaries you’ve set. It’s getting better at applying your policies, not changing them.

Jensen Huang introduces another useful concept: the AI Factory.

Traditional data centres store data. AI Factories manufacture intelligence.

Think about what happens when an information robot processes a contract:

Huang often says “tokens are the food” for information robots. Just as physical robots need energy to operate, information robots need computational resources to generate tokens—the currency of AI work.

For procurement teams, this means:

Procurement-agent platforms need more than a model and a chat interface. They require infrastructure, data pipelines, permissions, monitoring and governance designed around the work the agent is expected to perform.

Let me show you the difference with a real procurement scenario: A vendor’s insurance certificate is about to expire.

See the difference? The agent handles the entire workflow. You only get involved if there’s an exception that needs human judgement.

Not all agents are the same. Here are the main types you’ll encounter:

Reactive Agents These respond to specific triggers but don’t learn or adapt much.

Example: When a contract arrives, extract key terms and check against standards. If there’s a deviation, flag it.

These are the simplest agents. They’re reliable and predictable, but not particularly intelligent.

Goal-Based Agents These have specific objectives and figure out how to achieve them.

Example: “Ensure all active vendors have current insurance certificates.” The agent determines what actions are needed (request certificates, follow up, verify, escalate) to achieve that goal.

Learning Agents These improve their performance over time based on outcomes.

Example: A vendor onboarding agent that learns which vendors typically respond quickly to document requests and which need multiple follow-ups, adjusting its approach accordingly.

Collaborative Agents These work with humans and other agents, understanding when to handle tasks autonomously and when to involve people.

Example: A contract review agent that handles standard clauses independently but routes unusual terms to legal, understanding the boundary between routine and exceptional.

You don’t need to be a data scientist to understand this, but it helps to know what’s happening under the hood.

Modern AI agents are built on Large Language Models (LLMs), like GPT-4 or Claude. But they’re not just chatbots. They’re LLMs wrapped in additional capabilities:

Tools and Integrations The agent can use tools, like:

Think of the LLM as the brain, and tools as the hands. The brain decides what to do. The hands actually do it.

Memory The agent needs to remember:

Without memory, every interaction starts from scratch. With memory, the agent gets smarter about your specific environment.

Planning and Chain-of-Thought When facing a complex task, good agents break it down into steps:

Task: Onboard a new vendor

Agent’s thinking:

The agent can plan multi-step workflows and adapt if something doesn’t go as expected.

Here’s something critical: Good AI agents know their limits.

They’re designed with decision boundaries. Certain actions they handle autonomously. Certain decisions trigger human review.

For example, a contract review agent might:

This is “human-in-the-loop” design. The agent handles the routine 90% efficiently. You focus on the exceptional 10% that requires expertise and judgement.

Understanding this distinction is crucial. Agents aren’t replacements for procurement professionals. They’re tools that handle the work that doesn’t require human expertise, so humans can focus on work that does.

Let me walk you through a realistic example of how an agent works in practice.

Scenario: Your company receives a 50-page software licence agreement for review.

Total time: 5 minutes from receipt to routed for decision Human time saved: 2-3 hours of manual review Human time required: 30 minutes to review agent’s analysis and make final decisions on flagged issues

The agent didn’t replace the human. It did the tedious extraction and initial analysis work, so the human could focus on the judgement calls.

Understanding information robots matters because they’re fundamentally changing what “procurement work” means.

Jensen Huang predicts something profound: you’ll soon be working alongside millions of digital employees. Not metaphorically. Literally.

Your organisation will hire (deploy) these information robots just like it hires people. It will onboard them (train them on your policies and systems). It will manage their performance. And crucially, IT departments will evolve into the “HR departments” for these digital workers.

Think about what this means practically:

Traditional IT role: Provision systems, manage access, maintain infrastructure Future IT role: Deploy agents, train them on organisational knowledge, monitor their performance, manage their “career development” as they learn and improve

For procurement professionals, this creates a critical skill: managing digital employees.

Work is splitting into two categories:

Category 1: Agent-Suitable Work (what your digital employees handle)

Category 2: Human-Essential Work (what you focus on)

If your daily work is mostly Category 1, your role will change significantly. But here’s the opportunity: a procurement professional who can effectively manage 10 information robots—delegating tasks, reviewing outputs, making strategic decisions—is worth far more than someone who does the work of one person really well.

Below this point, I share what it’s actually like to manage seven information robots in production—what they do, how they perform, what I’ve learned about the difference between theory and practice, and what to look for when you’re ready to deploy your own. This isn’t hypothetical. These agents are working right now.

If you’re serious about getting your career and your organisation agentic-first, this is the kind of real-world insight you need. Subscribe to get full access to this article and future deep-dives on managing AI agents in procurement and risk management.

This isn’t science fiction. It’s happening now. Managing information robots means:

Delegation: Knowing which tasks to assign to agents vs. handle yourself. “This contract is standard—agent can review. This one’s strategic—I’ll handle it personally.”

Oversight: Spot-checking agent outputs initially, then moving to exception-based review as they prove reliable.

Training: Providing feedback when agents make mistakes, which improves their performance over time (remember the data flywheel).

Workflow Design: Structuring processes so agents and humans collaborate efficiently. Setting the boundaries for autonomous action vs. human escalation.

Performance Management: Monitoring agent accuracy, speed, and consistency. Identifying when an agent needs retraining or when a human should take over.

The procurement professionals who understand how to work with these digital employees—how to design workflows, interpret outputs, handle exceptions, and focus on strategic work—will be far more valuable than those who resist the change.

This isn’t about AI replacing you. It’s about you managing a team of 10 digital employees instead of doing everything yourself.

Let me make this concrete. I’ve spent more than a year building and managing information robots that work in the background. Here are some of the digital workers I have tested or deployed:

Document Review (for general procurement documents)

DORA Agent (Digital Operational Resilience Act)

ESG Agent (Environmental, Social, Governance)

DPIA Agent (Data Protection Impact Assessments)

These aren’t hypothetical. They’re working right now, processing real documents, making real decisions (within defined boundaries), and improving with every interaction.

On a typical day, I might review 5 contracts that my contract review agent flagged as requiring human judgement, whilst it autonomously processed 20 others. I provide feedback on the 5 (which improves the agent’s future performance), and spend my time on the strategic contract negotiations the agent correctly identified as needing my expertise.

The vendor onboarding agent handles the administrative drudgery—chase documents, verify formats, update systems—whilst I focus on evaluating whether this vendor relationship makes strategic sense for our business.

This is what Jensen Huang means by digital employees. These agents don’t report to me on an org chart, but I manage their performance, improve their training, and decide which tasks to delegate to them vs. handle myself. They’re my team. They just happen to be made of software instead of flesh.

And here’s the thing: building and managing these agents is now part of my job description. Not an “extra” responsibility. Core work. Because the organisations that figure out how to effectively deploy and manage information robots will have a massive competitive advantage over those that don’t.

If you’re evaluating AI agents for procurement, here’s what to ask vendors:

Information robots—AI agents that live inside computers—are here. They can perceive situations, reason through options, and take action autonomously. They’re not chatbots waiting for instructions. They’re not simple automation following rigid rules. They’re digital employees capable of handling complex, multi-step workflows with minimal human supervision.

In procurement, these information robots handle the repetitive work that bogs down your day—document processing, compliance checking, routine follow-ups, data entry—so you can focus on the strategic work that actually requires human expertise.

As Jensen Huang predicts, you’ll soon be managing teams of these digital workers. The question isn’t whether this will happen. It’s happening now. The question is whether you’ll be one of the procurement professionals who knows how to work with them, or whether you’ll be struggling to keep up as your peers leverage these digital employees to do the work of 10 people.

Understanding information robots isn’t optional anymore. They’re already changing procurement work. The question is whether you’ll shape how they’re used in your organisation, or whether someone else will make those decisions for you.

Daniel Barnes is a procurement practitioner, AI builder and founder of World of Procurement. He has nearly a decade of experience across procurement, risk management and supply chain, and has tested AI-agent capabilities in production environments since early 2024. Views expressed are his own.