The architecture behind CoSupport AI is covered by US patent US11823031B1, "Systems and methods for a multi-model message generation architecture", granted on November 21, 2023. CoSupport AI Assistant is built on this architecture.
In plain terms, the patent describes two routes for a support message:
| What the patent covers | What it means for support teams |
|---|---|
| A message that starts a new thread, or includes an error log, is matched by similarity to a stored problem, and the reply is built from that problem's known solution | Repeat questions and error reports get a fast, consistent answer without calling a large generative model |
| Later messages in a thread go to a response model together with the earlier messages | Replies take the earlier messages into account, so customers don't have to repeat themselves |
| The response model is trained on threads of support messages and replies | Answers are shaped by real support conversations |
| Text such as an email address can be replaced with a placeholder token during pre-processing | Messages are normalized before the model reads them |
Key Technologies Under the Hood of AI Assistants for Customer Support
AI assistants and copilots have become common. ChatGPT and Microsoft Copilot are two examples of tools that professionals and curious users rely on for everyday tasks: writing text, creating images, summarizing information, brainstorming campaigns, analyzing data, and writing code.
An end user is rarely interested in the technology under the hood. Which tool would you prefer for text summarization, ChatGPT or Bard? It depends on the task and on personal preference. You'll hardly dig into which model is better, GPT-4 or Gemini Pro.
Business is different. Say you choose an AI copilot for your customer support. The type of model behind it matters. Let's use one fairly complex ticket to see the difference between two types of models: one built on a Recurrent Neural Network (RNN) and one on the Transformer architecture.
The query:
"Last week, I purchased a blue jacket from your online store. However, when the package arrived yesterday, the jacket was green. I've checked my order confirmation email, and it clearly states that I ordered a blue jacket. Could you assist me in exchanging this for the correct item, and also, could you confirm if the return shipping costs will be covered due to this error?"
A hypothetical answer from the RNN model:
"Hello, thank you for reaching out. To exchange an item, please visit our returns and exchanges page. You can find the link on our website. Thank you for shopping with us."
The response from a Transformer model:
"Hello, I apologize for the inconvenience caused by the color mismatch in your order. Yes, we will assist you with exchanging the green jacket for the blue one you originally ordered. Additionally, since this was our error, we will cover the return shipping costs. Please refer to the instructions in your email to initiate the exchange process."
Even without a background in AI, you can see the difference in how well each answer understands the request.
The RNN model's response is generic. It doesn't acknowledge the color mismatch or the question about return shipping. That's a typical limitation of RNNs, which struggle to keep context over longer text.
The Transformer model's response covers every point in the query: the color mismatch, the exchange, and the return shipping costs. It keeps the context of the whole message and gives a precise answer.
Both answers are polite. But at scale, the better answers decide whether customers stay. That's why it's worth knowing which models and architectures work better, and why.
AI Tools Capabilities in Customer Support. From Early Days to Cutting-Edge Innovations
The Evolution of Technologies
The long-term goal of AI research, and generative AI in particular, is an AI agent that handles a wide range of tasks as well as a person: understanding natural language, reasoning, solving problems, writing, and responding with empathy. (Artificial general intelligence, or AGI, goes further: a system that is as capable as a person and can learn new skills on its own.)
Customer support sits where AI overlaps with Natural Language Processing (NLP): the ability of AI agents to understand and generate human language, work out what the user wants, and respond in context.
Here's how approaches to NLP evolved, using the same blue jacket question:
"Last week, I purchased a blue jacket from your online store. However, when the package arrived yesterday, the jacket was green. I've checked my order confirmation email, and it clearly states that I ordered a blue jacket. Could you assist me in exchanging this for the correct item, and also, could you confirm if the return shipping costs will be covered due to this error?"
| Time | Milestone in NLP | Example of the response | What changed |
|---|---|---|---|
| 1950s–1960s | Rule-based systems | "To return an item, fill out a return form." | Basic, rigid response with no personalization or context |
| 1970s–1990s | Statistical models | "You can return jackets within 30 days." | Recognizes the item category, but no personal context |
| 1980s–2000s | Recurrent Neural Networks (RNNs) | "I see you bought a jacket. You can return it within 30 days using the included label." | Recognizes the purchase and gives more specific instructions |
| Late 1990s–2010s | LSTMs | "I noticed you purchased a blue jacket last week. For a hassle-free return, please use the return section on our website." | Understands both the item and the timing of the purchase |
| 2014–present | Gated Recurrent Units (GRUs) | "Hi, thank you for choosing our service. Based on your previous interactions, we have some personalized recommendations for you." | Similar to LSTMs with fewer parameters, so faster and simpler; remembers earlier context |
| 2017–present | Transformer models (e.g., BERT, GPT) | "Hello! Regarding your recent purchase of a blue jacket, you can easily return it by following the steps outlined on our returns page." | Personal and specific, acknowledging the item and giving clear steps |
Illustrative replies written for this article.
Looking through the table, approaches evolved toward three capabilities:
- grasping the context;
- handling more complex requests;
- personalization.
Today's systems are good at all three, but they aren't perfect, and AI tool makers still face real challenges. Those challenges are where one AI agent differs from another.
Challenges
Today's AI tools based on large language models (LLMs) give fast, relevant, context-aware answers, but there are still problems to solve:
- Contextual understanding. LLMs can struggle to keep track of long conversation threads, because of memory and computational limits.
- Speed and efficiency. Fast responses matter in support, but complex models can be slow to generate replies.
- Personalization. Answers should reflect the customer's history, preferences, and situation, which means connecting the AI to your CRM and helpdesk.
- Accuracy and relevance. Answers have to be correct as well as on topic.
- Computational capacity. Advanced LLMs need a lot of computing power for training and real-time answers, which makes scaling expensive.
- Ethics and privacy. Customer data has to be used responsibly and kept private.
- Human handover. Conversations need to move smoothly from the AI to a person when needed.
When you compare AI support tools, check each one against these seven areas.
The Ways to Succeed in Customer Service: CoSupport AI Patented Approach
With the challenges clear, here's how they can be addressed. Every AI company finds its own path within its technical capabilities, budget, legal requirements, and ethical guidelines. Let's look at how CoSupport AI builds its products, for two reasons:
- the company has built AI products since 2020;
- CoSupport AI Assistant is built on the patented multi-model message generation architecture.
CoSupport AI has two products. CoSupport AI Agent resolves repetitive customer requests on its own. CoSupport AI Assistant works inside your helpdesk and gives human agents a ready draft for each ticket, along with summaries, suggested resolutions, and translation. The sections below show how the patented approach works in the Assistant, where a human agent reviews and sends every reply.
Contextual Understanding
Context awareness keeps a conversation coherent. When a customer discussing a blue jacket return later writes "the item", a context-aware AI knows what they mean. That prevents repeated questions and lets the AI build on what was already said.
How it works in CoSupport AI Assistant
The Assistant learns from your knowledge base and historical tickets, and reads the earlier messages in the thread before drafting a reply, so customers don't have to explain their situation again.
Speed
A fast response solves the customer's problem sooner and shows you take their time seriously. Handling tickets faster also means a team can manage more of them with shorter waits.
How it works in CoSupport AI Assistant
This is where the multi-model message generation architecture comes in. It sends each message down one of two routes:
- A new conversation, or a message with an error log. The system looks for a stored problem similar to this one and builds the answer from its known solution.
- A later message in the same thread. The message goes to a response model together with the earlier messages, so the reply takes them into account.
The patent describes matching each message to the technique suited to it as a way to give more accurate replies than one technique for everything. On the speed side, the Assistant has a draft ready before your agent opens the ticket.
Personalization
Personalization makes customers feel understood. Answers that reflect a customer's situation solve the problem more effectively and build loyalty.
How it works in CoSupport AI Assistant
The Assistant drafts each reply using your knowledge base and business systems, and it reads the earlier messages in the thread, so each reply fits both the current question and what came before.
Computational Capacities
Computing costs grow with volume, so efficient computation decides whether AI support stays affordable at scale. It also affects speed: less computing per reply means faster replies.
How it works in CoSupport AI Assistant
The patented multi-model architecture routes each message to the technique suited to it: known solutions for new threads and error reports, and a trained response model for follow-up messages, instead of sending every message through the same process.
Accuracy and Relevance
Accurate answers solve the problem the customer actually has. Relevant answers fit their specific situation. Together they build trust and cut down on follow-up messages.
How it works in CoSupport AI Assistant
A few things keep the Assistant's drafts accurate and relevant:
- Data integration. It drafts replies using your knowledge base and business systems.
- Contextual awareness. It reads the earlier messages in the thread, so drafts build on what's already been said.
- Learning from your own content. It learns from your knowledge base and historical tickets, so drafts follow your policies and the way your team answers.
In the patent, the response model is trained on threads of support messages and replies.
Ethical and Privacy Concerns
Privacy and ethics decide whether customers trust you with their data. Laws like GDPR make compliance mandatory, and responsible data handling protects customers from identity theft and data breaches.
How it works in CoSupport AI Assistant
CoSupport AI is ISO 27001 certified and GDPR and CCPA compliant. The security page has the details.
Seamless Human Handover
Some issues are too complex or sensitive for AI. A good handover moves them to a person with the full context, so the customer doesn't repeat themselves, and combines AI speed with human judgment.
How it works in CoSupport AI Assistant
With the Assistant, a human agent reviews and sends every reply. The agent brings empathy and experience to the final answer, starting from a draft that's ready before they open the ticket.
Conclusion
NLP has come a long way, from rule-based replies to models that understand context and personalize answers. Today's AI assistants handle tasks from data analysis to personalized customer service, and the field is still developing fast.
Competition between AI providers works in favor of the people who use these tools: new ideas reach products quickly, and existing products keep improving.
CoSupport AI has been building AI for customer support since 2020. Patent US11823031B1, granted in November 2023, covers the multi-model message generation architecture that CoSupport AI Assistant is built on.
The architecture sends each message down one of two routes. A new conversation or an error report is matched to a known solution, and later replies are written with the earlier messages in view, by a response model trained on threads of support messages.
The patent covers the routing approach behind how CoSupport AI answers support messages.
