Pack Assist
Lower AI costs and 24/7 sales coverage for a packaging supplier with Pack Assist
A human-in-the-loop AI chatbot for the custom packaging industry
Custom packaging quotes need exact specs, but CSRs gathered them by hand and after-hours visitors got no reply. We built a chat widget backed by a 12-stage AI pipeline that answers from a knowledge base, captures quote specs in real time, and lets CSRs approve or edit every reply before it goes out.
Industry
Packaging & Manufacturing
Client
B2B Packaging Supplier
Engagement
End-to-end AI sales qualification platform
Outcome
Reduced AI costs with accurate, 24/7 coverage
Tech Stack
React, Tailwind CSS, Socket.IO, Node.js, Express, OpenAI, LangChain, Pinecone, MongoDB
How do you stop an AI chatbot from making up prices?
We added strict guardrails to the pipeline. A no-pricing rule blocks the bot from quoting a price under any wording, and a 350 GSM rule keeps material answers inside company policy.
This matters in packaging because a wrong quote costs real money. The bot can explain box styles and materials, but the moment a chat turns into a price question, it hands off to a person. Guardrails are code, not prompts, so they hold even when the model is confident and wrong.
Can a human review AI chat replies before customers see them?
Yes. In Approval Mode, every AI draft goes to a CSR queue where a person can edit or approve it before it is sent.
Admins can flip to Direct Mode once they trust the bot on routine questions. And a one-click takeover lets a CSR jump into any live chat. The AI pauses, the chat history stays intact, and the structured spec data the bot already captured stays with the conversation.
How does an AI chatbot cut LLM costs?
It skips the LLM whenever it can. A fast regex check answers greetings and small talk with 0 API cost, and a complexity classifier routes simple questions past the rewriter and vector search, saving about 1.2 seconds per query.
Only questions that need real knowledge reach the full pipeline with its dual RAG lookup. Every token is counted with tiktoken and shown in USD on the dashboard, so the team sees exactly what each conversation costs to run.
How does Pack Assist capture packaging quote specs automatically?
Zod-based extraction reads each chat and writes box style, size, quantity, material, and contact details into the CRM in real time.
Before Pack Assist, CSRs chased these specs across scattered email and chat threads. Now the spec sheet fills itself while the visitor types. When the chat mentions a visual style, the system can also pull up to 5 matching gallery images into the CSR panel to speed up the reply.
Does the AI get smarter from CSR edits?
Yes. When a CSR edits a reply, the approved edit is saved and vectorized into Pinecone.
Future answers use dual RAG across the main knowledge base and those past human edits. So every correction a CSR makes teaches the bot, and the same mistake gets less likely over time. The knowledge base is managed from the admin dashboard, no engineer needed.
What happens if the AI provider goes down?
The pipeline falls back on its own. LangChain fallbacks route requests from OpenAI to OpenRouter the moment rate limits trigger, so chats keep flowing.
The backend is built for load too. Per-user mutex locks stop race conditions, MongoDB runs pooled connections, and rate limiting, CORS whitelisting, and Helmet keep the API stable and locked down.
How do you add a chatbot like this to a website?
With a single script tag. The chat widget is a React app served from Netlify, so the client's site just loads one script and the widget appears.
A domain whitelist in the admin dashboard controls which sites can load the widget, and ban controls block abusive visitors. No changes to the client's backend are needed.
How long does it take to build a human-in-the-loop AI chatbot?
We shipped Pack Assist between May and July 2026, about 3 months.
That covered the embeddable widget, the Node.js and Socket.IO backend, the 12-stage AI pipeline, and the CSR and admin dashboard with its approval queue, trigger builder, and analytics. The system now handles over 10,000 active records in production, and we verified a 100% document match across 10 collections (10,340 records) during migration.
Business Impact
Pack Assist reduced wasted AI spend and gave the supplier a more reliable 24/7 sales channel.
0/7
Sales coverage
Always-on qualification and handoff
0x
Cost control
Hybrid flow reduced unnecessary LLM calls
0%
Agent visibility
Live dashboard for handoffs and lead context
0 flow
Unified workflow
Sales, RAG, and support in one system
Frequently asked questions
Common questions about Pack Assist and human-in-the-loop AI chat.