Insights from Implementing 100+ AI Agents: Part 2

Over the past year, we have implemented over 100 AI agents in online stores worldwide. Our bots have answered more than 145,000 inquiries and resolved 86,000 tickets, resulting in an average resolution rate of 59%.

In this article, we discuss the technological challenges we overcame to deliver AI agents that truly work.

We found ways to address problems such as:

1. 1. GPT hallucinations
   2. High GPT costs
   3. Latency
   4. Low-quality source content
   5. Lack of control
   6. Difficult bot-agent cooperation

1. Eliminating hallucinations is possible

Hallucinations in AI responses are a common issue. Despite improvements with each GPT iteration, they can still occur, which is unacceptable in customer interactions. We’ve tackled this by integrating dynamic AI reasoning chains into PerfectBot. This means we process user input as a series of lightweight tasks, using GPT and dynamic decision points to refine NLP output.

Additionally, we implemented multi-layer hallucination prevention, ensuring that PerfectBot’s responses are firmly grounded in content sourced from customers’ knowledge bases.

2. Using AI agents must be cost-effective

Using the API offered by OpenAI incurs costs – the higher the model number, the higher the cost for each API request. Not all bot-generated queries need to be handled by the highest model. By introducing adaptive model orchestration into our system, PerfectBot dynamically selects various GPT models for NLP tasks, balancing quality and cost. This has allowed us to offer the most sophisticated AI agent on the market for as little as 99 USD per 1,000 conversations.

Our technology makes PerfectBot one of the most competitively priced solutions today. It is up to 80% cheaper than other similar solutions.

3. Customers want fast responses

Customers on chat are willing to wait a bit longer for human replies but do not want to wait for bot responses. They become impatient after just a few seconds. The higher the GPT model, the longer it takes to generate responses.

Our goal was to minimize this time while maintaining the quality.PerfecBot handles real-world conversations with an average latency of under 2.5 seconds per message.

4. Handling low-quality source content

Our AI agents can base their knowledge on merchants’ help centers, customer tickets, Shopify product details, policies, and blogs.

Some customers have fragmented and sometimes contradictory knowledge that reminds of a situation with puzzle pieces that don’t fit each other. We can quickly transform unstructured content into a validated format, enabling the AI agent to reason and respond concisely.

5. Customers want to have a control

We have just released a new dashboard that includes continuous chat quality control tools. Customers can access metrics, topic popularity, and knowledge gaps. Furthermore, we enable customers to improve the chatbot’s knowledge by using Gorgias Internal Notes.

Internal Notes are messages within helpdesk tickets that are only visible to your team and not to the customer.

6. Bot-agent cooperation is quite complicated

Seamless AI-human cooperation is crucial for effective helpdesk integration. PerfectBot offers dynamic handover scenarios tailored to topics, feedback, hours, and agent availability. This flexible collaboration allows for ticket pick up, closing, or tagging according to business needs.

Last but not least, we have recently observed rising customer expectations due to the new GPT models from OpenAI. This motivates us to continue improving PerfecBot. If you haven’t had a chance to try it, you can start a free 30-day trial with no strings attached.