How AI helps support teams find information, analyze inquiries, and prepare responses without taking over the conversation with the customer.
* Although this article refers to Intercom throughout, the process works with any CX platform that offers internal notes and API access.
Not every conversation with a customer should be automated
A customer asks why their refund hasn't arrived. Before replying, a support agent opens the payment system, checks the transaction status, reviews the customer's history, and looks for any outstanding issues.
Five minutes of investigation. Thirty seconds to write the reply.
What if AI could handle those five minutes of work, while the agent remained responsible for the conversation?
There are customers and situations where the response itself matters more than the information it contains. These include large B2B accounts, high-value customers, and complex issues involving money and access. Often, companies don’t want to take people out of support at all. They want conversations to stay human, with real people answering, but they want to take the routine load off those people.
In these cases, tone, attention to detail, and the sense that there’s a real person on the other end who’s accountable for the outcome are crucial. If you hand these conversations over to a machine, you risk losing the very reason the customer stays with you.
That said, a person doesn’t necessarily have to do all the work manually. Before each response, a support agent typically follows the same process: opens the back office, checks the payment, reviews the verification status, clarifies the customer’s terms, and only then writes the response. The response itself takes a minute, but preparing for it takes significantly longer.
It is this preparation that can be handed over to AI. The task of responding to the customer remains with the human.
How the AI Support Assistant Works
A four-step process:
Request. The customer sends a message via Telegram, email, or live chat. All channels feed into Intercom.
Investigation. Echomind reads the message, checks the knowledge base, and calls the functions you’ve authorized.
Draft. The result appears in an internal note attached to the conversation: what was determined, what response is suggested, and what risks exist.
Human Decision. A support agent opens the conversation, where the checks have already been completed and the draft is ready. The support agent edits the text, approves it, and sends it themselves.
The customer sees a standard message from their support agent. The internal note is visible only to the team. Until your team chooses a different scenario, no AI response reaches the customer.
What Can Be Automated in Practice
Below are three requests from different areas. Each one shows what the customer wrote, which functions the AI called, what the system returned, and what draft ended up in the internal note.
Example 1. Refund Status
Example 2. Account Access
In this example, the AI does not write “account unlocked” because the system has not confirmed this. It only states what is actually known: verification has been completed, and the unlock is pending approval.
Example 3. Individual request
In all three cases, the support agent receives the facts that have already been gathered and a draft text. They verify that the text matches the situation and add what they personally know about the client.
How Human-in-the-Loop Works
Echomind prepares the draft, but it doesn’t decide what happens next. Every draft ends in one of three outcomes.
The agent sends the draft as is. In the refund example, the facts are clear and the draft matches them. The agent checks the text and sends it.
The agent adjusts the draft. In the plan renewal example, the customer may be a long-term account. The agent keeps the facts but changes the tone, adds a personal note or offers a call. The data doesn’t show the relationship, the agent does.
The agent rejects the recommendation and keeps investigating. In the account access example, the customer mentions that the lock appeared right after a suspicious login. The lookup didn’t cover that, so the agent discards the draft and continues the investigation, for example together with the security team.
In all three cases the agent stays responsible for the conversation. This setup works for three reasons.
First: in this workflow, AI does not make sensitive decisions. Refunds, account unlocks, exceptions to the rules, and special conditions stay with a human.
Second: AI does not promise the customer anything that has not been confirmed by the system. The draft is based on API responses, not on assumptions.
Third: humans see what the data doesn’t show. The customer may be frustrated after a third request this week. Or they may be a key partner. This context changes the tone of the response, and sometimes the response itself.
Data and Control
Security is usually the first concern for companies, so here’s a brief overview of how it works.
Infrastructure. Echomind is hosted on our servers.
Models. The AI models run on our servers too, so customer conversations never go to third-party cloud LLM providers.
Storage. Chat history is stored temporarily for 24 hours. This speeds up processing and keeps context between messages. We do not retain it any longer than that.
Training. Customer data is not used to train the models.
System Access. The AI reaches your systems only through the API functions you approve, and each function returns a limited set of data. It has no direct access to your databases.
Beyond Drafts: Evaluating Support Quality
Preparing better responses is only one part of improving customer support. The next challenge is making sure those responses consistently meet your team’s quality standards.
This is where Judge LLM comes in. It evaluates completed conversations based on criteria you define, and it doesn’t stop at AI-assisted ones. Conversations handled by the team alone are scored by the same criteria. So Echomind helps prepare the replies and helps control their quality.
How it works:
Every completed conversation is evaluated, not just a sample;
Each criterion is scored from 1 to 10, and you define what each score means;
Conversations handled by agents alone and AI-assisted conversations are scored by the same criteria;
Weaknesses point to specific areas for improvement: the knowledge base, guidelines, or support agent training.
Example of the “payment escalation” criterion:
10: The procedure was followed, the transaction was verified, data was collected, and the escalation was performed correctly.
5: Some steps were skipped or the escalation was incomplete.
1: The escalation was performed immediately, without the required checks.
Judge LLM does not replace manual QA. The final decision rests with your quality assurance department, and the system indicates which conversations require priority attention.
What Changes for Your Support Team
Time per reply changes too. As an illustration: if preparing a reply takes an agent 12 to 15 minutes today, the same request could take 4 to 6 minutes when the data is already gathered and a draft is ready.
How to Get Started with a Pilot
There’s no need to launch a full-fledged bot for all customers right away. It’s better to follow these four steps with a small team:
Connect Telegram, email, live chat and your knowledge base to Intercom.
Start by reading data. Begin with read-only functions: payments, terms, and account status.
Review with support agents. Look at the draft quality, draft preparation time, and complex cases.
Add Judge LLM. Evaluate each conversation based on your quality criteria.
Want to try it out with your team? Launch a pilot at echomind.space.
Run a pilot on your chat
We pick one conversation segment together and fix success criteria up front. No rip-and-replace.