Case study
How Order.co Used Agentic AI in Procurement to Automate Vendor Ordering
Redefining B2B Procurement with AWS and Anthropic.
Order.co processes nearly $1 billion in annual purchasing volume across thousands of vendor platforms. phData automated that ordering end-to-end in six weeks using agentic AI in procurement. Built on Anthropic’s Claude via Amazon Bedrock, the system logs in, selects items, validates cart accuracy visually, handles payment, and submits confirmed orders without human intervention. It achieved a 100% success rate across vendors.
success rate: end-to-end order automation across vendors
At a glance
| Industry | Procurement technology / B2B SaaS |
| Scale | 160 employees; nearly $1B in annual purchasing volume |
| Challenge | Manual browser-based vendor ordering that could not scale without adding headcount for every new client or vendor |
| Technology | Anthropic Claude (via Amazon Bedrock), Amazon Nova, Amazon Bedrock AgentCore, LangGraph, AWS Fargate (ECS), Amazon API Gateway, AWS Lambda, Amazon SQS, Amazon DynamoDB, Amazon S3, AWS WAF v2, AWS Secrets Manager, Playwright, Bright Data proxy |
| Timeline | Six-week proof of concept |
| Result | 100% success rate across vendor platforms; orders processed in minutes vs. hours |
| Phdata Service | AI & Machine Learning, AWS, Anthropic |
The problem
Why traditional procurement automation fails at scale
Traditional RPA struggles to scale in procurement because vendor websites change constantly, breaking scripts built on fixed click sequences. Order.co had invested in RPA and found it too brittle for a vendor network that updates continuously. CAPTCHAs blocked bots. Payment flows varied. The engineering effort required to maintain scripts across dozens of vendors was eroding the efficiency gains. Order.co needed automation that could interpret a page, reason through changes, and act correctly.
Agentic AI in procurement refers to AI systems that perceive live vendor websites, reason about the required steps, and execute purchasing workflows autonomously, adapting in real time to layout changes, security prompts, and edge cases.
Robotic Process Automation (RPA) is software that automates tasks by replaying a fixed sequence of recorded clicks and keystrokes, with no ability to interpret or adapt to changes on the page.
LangGraph is a framework for building stateful, multi-agent AI pipelines where each node has defined inputs, outputs, and branching logic for success and error conditions.
Amazon Bedrock AgentCore is an AWS runtime for deploying and running AI agents at production scale, providing the execution environment for LangGraph pipelines inside containerized Fargate tasks.
How phData built the agentic procurement system
Four decisions shaped the delivery.
01
phData built a multi-agent LangGraph pipeline powered by Anthropic's Claude, replacing RPA
Each order runs through a six-stage LangGraph graph: clear-cart, shop, cart QA, ship-and-pay, checkout QA, and complete-purchase. Anthropic’s Claude Sonnet 4, accessed via Amazon Bedrock, serves as the reasoning model at each stage. Two dedicated QA agent nodes compare intended vs. actual cart contents and checkout totals, using vision plus structured data validation, before any order submits. A separate error-ID agent classifies failures (out-of-stock, minimum quantity, payment declined) with structured codes, giving operations teams actionable context rather than a generic failure status.
02
The architecture separates ingestion, orchestration, and execution into independent, isolated layers
Order jobs enter through API Gateway, pass AWS WAF v2 (OWASP rules, bot control, per-tenant rate limiting), and are persisted to DynamoDB before compute spins up. An SQS queue decouples job acceptance from execution. Each job triggers a dedicated ECS Fargate task running the Bedrock AgentCore runtime and a Playwright headless browser. One Fargate task per order provides strong isolation: a failure in one vendor session cannot affect another, and every task’s logs, state, and screenshots are independently traceable in CloudWatch and S3.
03
All browser traffic routes through a Bright Data residential proxy to maintain reliable vendor access at scale
Vendor sites use bot detection, IP rate limiting, and behavioral signals to block automated traffic. Routing through Bright Data’s residential and ISP IP pools makes requests appear as organic user traffic. This was a non-negotiable production requirement: without the proxy tier, vendor bot controls would block agents regardless of how well the AI reasoned about the page. Credentials, proxy auth, and API keys are managed in AWS Secrets Manager with least-privilege IAM policies.
04
State and audit artifacts are centralized in DynamoDB and S3 for replays, analysis, and compliance
Job metadata and current phase persist in a DynamoDB Jobs Table from the moment an order is received. Detailed per-step agent state, including cart contents and QA outcomes, lives in a separate DynamoDB Orders State table. Screenshots from every browser interaction are stored in S3 and referenced by URI. This structure enables idempotent job replays when failures occur, post-hoc analysis of straight-through rate, and a complete audit trail for every order placed.
The result
100% order automation across vendors
The six-week proof of concept placed real orders across all vendors at a 100% success rate. Orders that previously required manual staff hours now complete in minutes.
Order.co’s end customers get faster order confirmation and consistent execution regardless of which vendor site the order routes through. The QA gates before cart submission and before checkout mean only verified, accurate orders reach vendors. When something does go wrong, structured error classification tells operations teams exactly what failed and why, cutting time spent diagnosing exceptions.
The engagement also opened a capability Order.co did not have before: continuous vendor catalog expansion. Because browser agents can navigate any vendor site, Order.co can now search for new suppliers, compare pricing, and add products without manual effort. A business that previously knew three or four vendors for a given product can now find all ten, capture better pricing, and expand its margins at scale.
success rate: end-to-end order automation across vendors
Ready to see what a similar engagement could look like for your organization?
What the client said
I was really impressed that phData brought in highly talented, high-quality people into our very first calls. Those same engineers are helping us build the ability to place orders entirely automatically on any vendor's website, regardless of what that website looks like. I would highly recommend phData to anyone in the industry.
What agentic AI in procurement means for your operation
Agentic AI in procurement is no longer experimental. Any organization that places orders through vendor portals, manages logins across multiple suppliers, or relies on staff to navigate disparate checkout flows faces the same ceiling Order.co hit.
The Order.co engagement demonstrates what production-grade agentic AI in procurement actually requires: a multi-agent pipeline with explicit QA gates, residential proxy infrastructure for reliable vendor access, isolated ephemeral compute per job, centralized state and audit artifacts, and structured error handling that routes failures to humans with actionable context. phData has designed and deployed this stack. For organizations running manual browser-based workflows at scale, the path to automation is concrete and the timeline is short.
Frequently asked questions about agentic AI in procurement
What is agentic AI in procurement?
Agentic AI in procurement refers to AI systems that autonomously execute purchasing workflows, from vendor login and item selection through cart validation, checkout, and order submission, without requiring human intervention at each step. Unlike traditional procurement software that requires structured inputs and manual approvals, agentic AI perceives live webpages and makes decisions in real time. phData deployed this for Order.co using Anthropic’s Claude Sonnet 4 via Amazon Bedrock, achieving a 100% order success rate across 7 vendor platforms in six weeks.
What are the main use cases for agentic AI in procurement?
Production deployments include automated vendor ordering (logging in, selecting items, and submitting orders across multiple supplier sites), vendor catalog expansion (discovering new suppliers and pricing), and structured exception handling (classifying order failures by type and routing them to humans with evidence). phData’s Order.co engagement covers all three: the system places real orders, continuously surfaces new vendors, and emits structured error codes when a QA stage detects a mismatch.
What technology stack powers phData's agentic procurement system?
phData’s system uses Anthropic’s Claude via Amazon Bedrock for agent reasoning across a six-stage LangGraph pipeline. Amazon Bedrock AgentCore provides the runtime inside isolated ECS Fargate tasks. The infrastructure includes API Gateway, AWS WAF v2, Lambda, SQS, DynamoDB, S3, and Secrets Manager. Playwright handles browser automation; Bright Data’s residential proxy ensures vendor sites treat agent traffic as organic user sessions.
How does phData ensure only accurate orders are submitted by the AI agent?
The LangGraph pipeline includes two dedicated QA stages. The qa-cart-agent compares intended order contents to actual cart contents using visual signals and structured data before proceeding to checkout. The qa-checkout-agent verifies totals, taxes, shipping costs, and discounts before the final submit. When either stage detects a mismatch, a dedicated error-ID agent classifies the failure type (out-of-stock, minimum quantity not met, payment declined, AVS failure) and routes the job to human review with full screenshot artifacts attached.
Related case studies
Technology
AI & Data Strategy for a National Procurement Platform
Uncover how governed metrics, master data, and contract intelligence powered a 24‑month AI & Data plan for a national procurement platform.
Technology
Biotech Leader Transforms and Scales Analytics Platform on AWS
Explore how a global biotech leader worked with phData to modernize analytics on AWS—empowering 2,000+ users with faster, more reliable insights across the org.
Technology
How a K-12 Edtech Company Built a Personalized Learning Platform for Millions of Students
Discover how one of the largest K-12 edtech companies built a production-grade personalized learning platform with Claude on AWS Bedrock and Snowflake, saving $1M+ annually.
Get started
Want to put agentic AI to work in your procurement operation?
Talk to our experts about what browser agents can automate in your environment.
