Elastic – a company developing tools for data search and analysis – has published a description of how its tech support is structured. What is interesting is not that they use artificial intelligence (this is almost standard by now), but precisely how it is built into the process.
AI Helps, Humans Verify
The essence of the approach is simple: when a request comes in from a client, a system based on artificial intelligence and RAG technology (retrieval-augmented generation – a method where the model doesn't just generate text but first searches for relevant information in a knowledge base) drafts a response. But this draft does not go straight to the client.
Instead, a support engineer receives it. They read, check, supplement, or rewrite it – and only then hit send. That is, artificial intelligence handles the routine part: searching documentation, formulating a basic answer, and selecting links. Meanwhile, the human is responsible for accuracy, context, and appropriateness.
Elastic calls this «expert-verified solutions». The phrasing sounds like a marketing ploy, but there is very concrete logic behind it: AI speeds up work but does not replace expertise.
Why Companies Are Reconsidering Fully Automated Tech Support
Why This Is Even a Topic for Discussion
Many companies are now trying to automate tech support completely – so that AI answers directly, without human participation. Sometimes this works: for simple questions like «how to reset a password» or «where to download the installer», language models cope quite well.
But when it comes to complex technical products – databases, search engines, distributed systems – an error in the answer can be costly. A client might apply the wrong solution, break something in the production environment, or spend hours debugging in the wrong direction.
Elastic decided not to risk it and kept the human in the decision-making chain. AI is used as an assistant that saves the engineer's time but doesn't make decisions for them.
How RAG Technology Powers AI-Assisted Support
How It Works Technically
RAG is an approach where the model first accesses an external knowledge base (documentation, tickets, articles), extracts relevant information, and then forms an answer based on it. This helps avoid «hallucinations» – situations where the model confidently outputs something plausible but factually incorrect.
In Elastic's case, the knowledge base includes their own documentation, history of resolved tickets, known issues, and so on. When a new request arrives, the system searches for similar cases and assembles a draft answer from them.
The engineer sees not only the draft itself but also the sources the model relied on. This allows them to quickly assess how relevant the answer is and adjust it if necessary.
Benefits of AI-Assisted Support with Human Oversight
What This Yields in Practice
Elastic claims that this approach speeds up tech support work. The engineer doesn't need to search documentation manually every time, recall how a similar problem was solved a month ago, or write an answer from scratch. The draft is already ready – it just remains to check and, if necessary, refine it.
At the same time, the quality of answers doesn't drop because the final word is still with the human. Clients receive accurate, verified solutions, not automatically generated text that might turn out to be incomplete or erroneous.
Why This Is Interesting
This is a fairly balanced example of using artificial intelligence in a production environment. Elastic isn't trying to replace humans but isn't ignoring automation opportunities either. Instead, the company built AI into where it truly helps – routine operations – and kept humans where expertise is needed.
Such an approach is called «human-in-the-loop». It requires more resources than full automation but reduces risks and maintains control over quality.
In the context of tech support, this is especially important. Clients contact support not to get a quick but inaccurate answer. They need a solution that works. And if this requires the engineer to spend a couple of minutes checking the draft from the AI – that is a reasonable compromise.