I want to learn how to create AI Agent in Procurement Procedure

Welcome! I'm excited to help you design and build an AI agent for your procurement workflow.

Before we dive into the architecture and start mapping out the workflow, I'd love to learn a little more about your background so we can tailor this session perfectly to you.

Could you tell me if you are a procurement professional, a developer, or someone looking at this from a strategic business perspective? Also, do you already have a specific procurement task in mind that you would like to automate first (like drafting RFQs or checking contracts)?

I'm intern

It's fantastic to have you here! As an intern, you are in a perfect position to learn this because designing AI agents is quickly becoming one of the most valuable, forward-looking skills in the entire supply chain space.

Before we build anything, we need to layout our plan. I've designed a structured 4-step path to guide you from zero to a fully scoped, tested procurement agent.

Your path

Understanding Core Agent Architecture Components

Designing the Procurement Workflow

Implementation, Testing, and Governance

Let's kick things off with Step 1: Defining the Scope and Purpose of Your Procurement Agent. The biggest mistake people make is trying to build an agent that does "everything" right away. In reality, works best when we start with a highly specific, high-impact task. By narrowing our focus, we can guarantee accuracy and build something that actually saves real-world hours.

To find that perfect starting point, we want to look for tasks that are frequent, repetitive, and rely heavily on document reading or writing. Here are three classic procurement tasks that are absolute prime candidates for an AI agent:

RFQ (Request for Quotation) Drafting: Gathering requirements and automatically writing emails to suppliers asking for pricing.
Vendor Comparison: Parsing messy, incoming PDF quotes from different suppliers and compiling them into a clean, side-by-side comparison sheet.
Purchase Order (PO) Processing: Reviewing incoming purchase requests against company policy and automatically drafting the PO in the ERP system.

Which of these three tasks sounds the most interesting or relevant to what you see at your internship? Or do you have another repetitive task in mind you'd like to tackle?

Drafting RFQs sounds perfect

Drafting RFQs is a brilliant choice! It's highly structured, involves clear templates, and directly affects how quickly your team can get pricing back from suppliers. Now that we have our target task, we need to design its concrete objective, target metrics, and guardrails to make sure it runs successfully.

Setting Clear Objectives and KPIs

Before writing any code or prompts, you must define exactly what "success" looks like for your drafting agent. Without clear and guardrails, it's impossible to measure whether your agent is actually improving the procurement process or just creating more mess.

Objective	Key Performance Indicator (KPI)	How We Measure It
Speed up sourcing	Cycle Time	Time elapsed from initial purchase request to finalized RFQ draft
Maintain standards	Policy Compliance	% of drafts that pass internal procurement policy checks on the first run
Reduce manual effort	Automation Rate	% of RFQs generated entirely by the agent without needing rewrite

By establishing these metrics early, you give your agent a clear standard to meet. For instance, you might set a target of reducing the time to draft an RFQ from 45 minutes down to just 2 minutes, while maintaining a 99% compliance rate with your company's approved vendor guidelines.

How Can I create an AI agent that automate emails to sending vendors?

To automate email dispatch, we have to look closely at the boundary of your agent's autonomy. Deciding exactly where the AI's power ends and where a human must step in is called setting the human-in-the-loop (HITL) requirement.

The Spectrum of Autonomy

When you design an agent to interact with external partners (like sending emails to vendors), you have three main ways to structure its autonomy. Choosing the right one is a balance of speed and safety:

Autonomy Level	How It Works	Best For
Co-Pilot (High HITL)	Agent drafts the email; pauses and waits for you to review and click "Send" in a dashboard.	Complex, high-value, or custom RFQs with new vendors
Supervisor (Medium HITL)	Agent drafts and queues the email; sends automatically unless a human rejects it within a 15-minute window.	Routine re-orders from trusted partners
Autonomous (No HITL)	Agent drafts and sends the email directly to the vendor with zero human intervention.	Very low-value, standard commodities (tail spend)

For an intern starting out, the Co-Pilot model is almost always the best place to begin. It allows you to gain trust in the system, catch any weird formatting or tone issues, and ensure the is outputting the correct supplier names and pricing structures before anything goes live.

Which level of autonomy feels right for the RFQ drafting agent you want to build?

Okay, let's start from 0.

How can I create this agent for free?

Building an AI agent completely from scratch for free is a fantastic way to learn because it strips away the mystery of expensive, enterprise platforms. When you strip away the flashy user interfaces, you are left with three fundamental components that make up any agentic system.

The Basic DIY Stack

To build this without spending a cent, we can bypass complex framework libraries and write a simple script using standard, free tools. Your DIY agent relies on three core blocks to perform its work:

Component	What It Represents	Free Tool Solution
Input	The raw RFQ details and requirements	A free Google Sheet or Google Form
Brain	The LLM that draft the actual email	Free-tier API key (e.g., Gemini or Groq)
Output	Sending the email draft to the vendor	A Python script using a free Gmail SMTP connection

By connecting these three pieces, you create a complete loop: your Python script reads new RFQ requests from your spreadsheet, packages that data into a prompt for the free LLM API to process, and then uses a free email protocol to send the resulting draft to your inbox for review before it goes to the vendor.

This "from zero" approach gives you complete control over how the prompt is structured and how your data is handled.

Shall we look at how these three components pass information back and forth, or would you like to explore which free LLM APIs are best to start with?