Client Story
Using AI to Automate Order Entry and Optimize Costs
Little room for error with thousands of garments a day
Our client works on custom apparel, producing up to thousands of garments per day for national distributors and organizations. Their orders range from a few hundred garments to hundreds of thousands, often on tight turnarounds. They have a packed daily schedule, and a delay at order entry carries over straight into production, creating a costly waterfall effect.
Differing order submissions
Clients submit purchase orders as PDFs, spreadsheets, custom data feeds, and other formats. Then, the team spends significant time re-entering that data and resolving discrepancies between documents.
A rules-based parser was our first thought and implementation, but it wouldn’t have worked long-term. The formats vary too widely from one client to another for one tool to handle every nuance and required step. Each new client would have required their team to build and maintain yet another integration, which makes scaling costly.
Two systems built around existing internal tools
The first system we built is an AI agent that reads emailed purchase orders and enters them into the company’s production system. It handles client-specific formats and has been trained on customer patterns to improve accuracy over time. Once the AI agent processes an email, the sender receives a reply with the order’s ID in the system along with some details about the intake workflow, confirming it was processed successfully or explaining why it couldn’t be processed. This process also supports some back-and-forth needed to get the order right in case the AI agent needed additional information while it mapped with data in the database.
The second is an internal AI assistant connected to a limited view of the company’s live order and production data. Staff check on orders or generate reports via the company’s enterprise subscription of AI chat tools that they already had, like Claude and Slack’s AI assistant feature. The assistant helps them quickly get answers and reports from their databases, including finding data discrepancies before it’s too late.
Other features and functionality include:
Custom AI Solutions Delivered
AI Assistant
Recently, with a deeper understanding of AI systems and infrastructure, we upgraded our model to significantly boost its intelligence, reaching a top-tier frontier level, while maintaining the same costs. Through various cost optimization strategies, we managed to keep expenses unchanged and more than doubled the model’s intelligence. We recently finished another initiative that is expected to reduce inference costs by 50% while keeping nearly the same level of large language model intelligence.
Order Entry Agent
As more staff used these systems, the organization and our AI engineers evaluated a potential switch to a less expensive AI model but ultimately decided against it. Claude Sonnet produced significantly better results, and the potential reduction in order-entry accuracy wasn’t worth the cost savings.
The AI assistant starts finding cost optimization opportunities
With access to shipping cost data, the AI assistant quickly gave the team clear insight into shipping costs, so they can review and manage them with confidence. This effort resulted in identifying and implementing significant cost optimization opportunities in shipping.
What changed
- Order entry that once required manual work now runs automatically, around the clock with human oversight. New third-party order sources go to the existing AI agent, eliminating the need for a separate integration for each and significantly speeding up the new client onboarding process.
- Staff pull answers from live data faster and with low latency through Claude or our AI Slack assistant integration instead of manually searching for an answer.
- When the organization’s ERP (the core system that runs orders and production) went down for several hours, staff could still reach their data through the AI assistant.
Staff can access the AI assistant on desktop or mobile, primarily through the Claude and Slack apps, on their phones and desktops. Within weeks, the assistant’s message volume increased by 400%, proving its usefulness and usability.
Under the hood
All production AI calls start in a self-hosted workflow backend and go through OpenRouter, which reaches many model providers through one API and one bill. During a provider outage, we moved traffic to another provider automatically without a configuration change, so order processing and AI assistants kept running.
The AI assistant queries order data through an MCP server (Model Context Protocol, a standard way for AI assistants to call outside tools). A per-tool allowlist controls what each person can ask, and access requires the company’s single sign-on account on the company’s domain.
We enforce Zero Data Retention, so we block any provider that logs or retains request data. In addition, we automatically redact sensitive information and typical multiple injection techniques. Inference is limited to a vetted subset of roughly 10 providers.
Share the process your team keeps re-keying
We see this same workflow in manufacturing, distribution, logistics, and fulfillment companies, especially ones where staff re-enter orders that arrive in each client’s own format or search a system by hand to answer routine questions. The systems we built for this apparel manufacturer handle mission-critical production processes every day.
Reach out to our team to share where manual entry or hard-to-reach data is slowing your people down. We’ll learn how your orders and data move today, then build an AI solution tailored to that.