Manufacturing AI Adoption
From Pain Points to Solutions
A traditional manufacturer's AI selection journey: four real problems, four practical solutions.
No dodging hard questions, just solving problems.
Background: First Mover in the Industry
This is a large paper manufacturing conglomerate running a legacy Yonyou ERP system for their core business processes. The CEO's vision was crystal clear -- not satisfied with AI replacing just 10% of repetitive work, the goal was to automate the vast majority of daily office operations.
Among their industry peers, they are the first mover. Fellow paper manufacturers are closely watching their results -- the industry has a culture of sharing experiences, and once someone blazes a trail, others follow.
Their selection criteria were equally pragmatic: the product must be mature or near-mature, and once chosen, it would be adopted long-term. They had no appetite for the risks that come with early-stage startup products. They weren't looking for a "demo toy" -- they needed a "production-ready tool."
The value of this case study lies not in showcasing a perfect success story, but in honestly documenting every specific problem a manufacturing enterprise encounters during AI selection -- and how COCO responded.
Four Real Problems, Four Practical Solutions
Throughout the evaluation period, the customer raised four core issues. None were theoretical -- each was a real obstacle encountered during actual operations. These problems represent universal challenges that manufacturing enterprises inevitably face during AI adoption.
Inaccurate OCR Document Recognition
The customer tested AI with 10 photos of delivery notes, asking it to extract data into Excel. Despite clear photos, general-purpose AI produced inaccurate results that recurred after correction. COCO's approach: General AI isn't suited for specialized image recognition. Dedicated OCR tools/Skills are needed. The product roadmap already includes document-specific recognition Skills. Short-term solution: integrate professional image recognition toolchains for significantly improved accuracy.
Browser Automation Timeout (Banking Sites)
The company needed to log into banking websites that require token/OTP verification, but AI-driven browser operations were too slow -- tokens expired before they could be used. The fundamental issue: websites are designed for humans, not AI. COCO's approach: Prioritize API/CLI over browser simulation; use remote desktop assistance (Browser Explorer plugin) for initial login, then maintain the session; pursue a dual strategy -- improve tooling while pushing for API-based banking interfaces.
Automatic OTP Forwarding
The company phone receives OTP codes that must be manually forwarded to the Agent each time -- a tedious workflow bottleneck. COCO's approach: Develop a lightweight Android APK installed on the company phone to intercept SMS messages and automatically forward verification codes to the Agent. With clear requirements, the Agent can even build this APK itself.
Data Security & Information Isolation
The customer was concerned about AI handling ERP data securely. COCO's three-layer approach: Layer 1 -- ISO/SOC2 certification with end-to-end encryption; Layer 2 -- Agent information isolation best practices including personal/business separation, internal/external separation, and inter-department isolation; Layer 3 -- data stays in systems, not in the Agent, with query-only permissions and auditable access control. We were candid: hard isolation remains an industry-wide challenge that even Anthropic/OpenAI haven't fully solved -- we offer current best practices.
Bonus: The Skill Memory Problem
The customer also raised a practical pain point: they taught the AI a skill, but it forgot by the next day.
This is a real and widespread issue. The current solution involves visual storage (Git/Lark docs) to categorize and record Skills for precise retrieval. The next version of the Workspace product will productize these best practices -- letting users directly see what their Agent knows, what it has learned, and which Skill to use when, eliminating the need to re-teach every time.
We're not looking for a demo toy. We need a production-ready tool. Our industry peers are all watching our results -- once we blaze a trail, others will follow.
Frequently Asked Questions
Q: What level of image recognition accuracy can be achieved?
It depends on the document type and image quality. General-purpose AI has limitations for specialized document recognition, but accuracy improves significantly with dedicated OCR tools. COCO's product roadmap includes purpose-built recognition Skills for various document types (delivery notes, invoices, bills of lading), with continuous optimization.
Q: Can COCO integrate with existing ERP systems?
Yes. One of AI Agent's core capabilities is acting as a "glue layer" between systems -- it can connect to existing systems via APIs, database connections, or browser automation. The Width case study demonstrates this: a non-technical CEO completed a CRM + Zendesk + DocuSign integration in 3-4 days. Legacy ERP systems like Yonyou can be integrated through similar approaches.
Q: How is data security ensured?
COCO provides three layers of security: ISO/SOC2 certification with end-to-end encryption; Agent information isolation best practices (personal/business separation, internal/external separation, inter-department isolation); and access control with query-only permissions (data stays in systems, not in the Agent, with auditable retention policies). We are transparent that hard isolation remains an industry-wide challenge -- COCO offers current best-practice solutions.
Q: Will the Agent forget previously learned skills?
In the current version, Skill persistence is managed through visual storage (Git/Lark docs). The upcoming Workspace product will productize these best practices -- users will be able to directly see their Agent's Skills, when they're used, and how to manage them, fully solving the "taught today, forgotten tomorrow" problem.