How to Achieve 98%+ Accuracy with Customer Service Agents

Lessons from Real Projects

Since January 2025, DeepFlow AI has delivered customer service agents for dozens of leading clients, achieving accuracy above 98% and outperforming their previous human support teams. Every project has taken us through moments when the road seemed blocked, followed by a breakthrough. Along the way, we have also watched several companies' in-house initiatives fail.

Here are five counterintuitive lessons from building customer service agents.

Lesson 1: First Principles Beat Any Open-Source Framework

There is no need to put blind faith in Dify, Coze, or LangChain.

Many of these products were built around academic ideas and demos from the early days of agents.

An academic demo is not enterprise production.

DeepFlow AI's first-principles formula is:

Accuracy = Input quality × Model capability × Transparent pipelines

Input quality = Completeness + Accuracy + Clarity. All three are essential.

Building an agent product means constructing transparent information pipelines around LLM calls, so that information can be inspected and improved throughout the process.

Many companies build their own agents on products such as Dify but hesitate to modify the underlying implementation. They work only within the configuration options offered by the visual canvas. Platforms of this kind have significant limitations in customer service scenarios. Solving many of those problems requires the confidence to redesign them from first principles.

Ask yourself: how many of your agent's configuration choices exist because the low-code platform left you with no other option?

From the outset, DeepFlow AI chose to build its own agents around customer service requirements. That approach has worked well in practice.

Lesson 2: Keep the Architecture Simple and Go Deep on the Technology

If retrieval is inaccurate even across ten documents, what should you do?

• Option A, chosen by 99% of teams: Have an LLM classify the query first, then retrieve from a smaller subset. Two minutes of dragging nodes on a canvas, and it is done.
• Option B: Investigate the underlying retrieval technology, identify the cause, and solve the retrieval problem across those ten documents.

Resist the temptation of Option A. Choose Option B.

The Problems with Option A

1. It adds another LLM call, increasing response time.
2. The text descriptions used for classification are incomplete representations of the categories, introducing new uncertainty.
3. The solution is difficult to generalize.

The Advantages of Option B

1. It solves a retrieval problem at the retrieval layer, addressing the cause without adding architectural complexity.
2. It improves overall retrieval capability while resolving the immediate problem.

For every agent module, product and engineering teams need the ability to investigate and modify the underlying technology.

“Low-code” does not mean “low-tech.” Fewer nodes force you to go deeper on each one.

Do not let your agent canvas grow into hundreds of nodes that become impossible—or too intimidating—to maintain.

Lesson 3: Make Full Use of Elasticsearch

The BM25 search bundled with a vector database is roughly a toy version of search.

DeepFlow's approach is to use Elasticsearch (ES) for BM25 search. We draw on its powerful search capabilities and continuously refine a retrieval solution designed for document chunks.

Agent product managers should also understand Elasticsearch's parameters and search optimization techniques.

Lesson 4: In the AI Era, Be Willing to Do Manual Work

Even the strongest model cannot rescue dirty data.

Information sent to an LLM must be complete, accurate, and easy to understand. Faced with messy documents, spreadsheets, and files, even today's most advanced technology can fall short.

Should you wait for AGI, or invest in manual processing and proofreading now to obtain high-quality data?

Data preparation is a major part of the service DeepFlow AI provides to clients. It is demanding, labor-intensive work. It is also essential to delivering highly accurate customer service agents.

Do not believe promises of agents that require zero human effort and no training.

Lesson 5: Rethink the Organization

The traditional waterfall process—requirements, scheduling, development, and delivery—is a poor fit for today's agent products.

Agent products do not yet have mature methodologies and established best practices. In-house engineering teams cannot reliably complete these tasks within a fixed schedule. Under deadline pressure, they often ship quickly using an existing but unsuitable framework. When problems arise, the response becomes: “That is just how large language models behave.”

DeepFlow AI has built a research-oriented organization. We take the difficult but necessary path, innovate through challenges, and accumulate customer service agent know-how through project delivery. We then provide clients with an end-to-end service, from knowledge to conversation.