A plain-English look at how large language models turn a customer message into a reply, why they get things wrong, and how to keep answers grounded.
A customer types "Can I change my delivery address?" into your website chat, WhatsApp or Messenger. A few seconds later, a fluent, polite answer appears. It feels like magic, but it is a predictable process, and understanding it is the difference between an agent you can trust and one that occasionally makes things up. This article walks through what happens between the question and the reply, and what you can control.
What a large language model actually is
A large language model (LLM) is a program trained on very large amounts of text to predict what words are most likely to come next in a given context. Models such as Claude, GPT and Gemini all work on this principle. They do not look up answers in a database of facts. They generate text one small piece at a time, guided by everything in front of them.
That has two consequences worth remembering:
- LLMs are excellent at language. They understand paraphrases, typos, mixed languages and vague phrasing far better than keyword matching ever could.
- LLMs are not a source of truth about your business. A general model has never read your refund policy, your shipping times or your price list. Unless you give it that information, it will fill the gap with something plausible.
So the real question in customer support is not "how smart is the model?" but "what is the model allowed to see when it writes the reply?"
From message to reply: the pipeline
When a customer message arrives, a well-built support agent does much more than forward it to a model. Roughly, the steps look like this:
- 01Customer message arrives
- 02Relevant knowledge is retrieved
- 03Instructions and context are assembled
- 04Model writes the reply
1. The message arrives with context
The message comes from a channel such as a website widget, WhatsApp or Messenger. The agent also has the earlier turns of the conversation, so "What about the blue one?" can be understood as a follow-up and not a brand-new question.
2. Relevant knowledge is retrieved
Your help articles, PDFs, FAQs and product pages are too large to hand to a model in full for every question. Instead, they are split into passages. The system finds the few passages most related to the customer's question and passes only those along. This technique is commonly called retrieval-augmented generation, or RAG.
3. The prompt is assembled
The model receives a package that typically contains:
- the agent's persona and rules (tone, scope, what it must never promise)
- the retrieved passages from your knowledge base
- the conversation so far
- the customer's latest message
4. The model writes the answer
Using all of that, the model composes a reply in the customer's language and in your brand's voice. If the retrieved passages contain the answer, a good reply is mostly a clear rewording of them. If they do not, the behaviour depends entirely on the rules you set.
Why LLMs sometimes get answers wrong
Most bad answers fall into a handful of patterns, and each has a practical fix.
| What goes wrong | Why it happens | What helps |
|---|---|---|
| Invented policy or price | No relevant passage was found, so the model improvises | Add the missing content; instruct the agent to say it does not know |
| Outdated answer | The knowledge base still contains an old document | Re-sync sources; remove stale files |
| Wrong passage used | Two articles overlap or contradict each other | Consolidate duplicates |
| Overconfident tone | The persona never says when to hand off | Add explicit escalation rules |
| Off-topic or manipulated replies | A user tries to override the instructions | Enable abuse and prompt-injection filters |
The pattern is consistent: most failures are information problems, not intelligence problems.
Grounding: keeping answers tied to your content
Grounding means the agent answers from your approved material rather than from the model's general memory. You strengthen it in three ways.
Write sources the way customers ask
Customers ask "Can I return something I bought on sale?" not "Returns policy, section 4.2." Q&A pairs and short, plainly worded articles retrieve better than dense legal text. If your policy has exceptions, state them in the same passage as the rule.
Give the agent permission to say "I don't know"
A persona instruction such as "If the answer is not in the provided information, say so and offer to connect the customer with a person" prevents the most damaging failure: a confident wrong answer. Customers forgive an honest handoff far more easily than a false promise.
Keep one source of truth
If the same fact lives in five documents, it will eventually be out of sync in two of them. Fewer, cleaner sources produce more consistent replies.
Choosing and testing the model
Different models have different strengths: some are faster and cheaper, others handle long, nuanced conversations better. For a simple FAQ-style agent, a standard model is often enough. For complex products or multi-step questions, a more advanced model may be worth the extra cost.
Do not choose by reputation alone. Test with real questions:
- Collect 20 to 30 actual customer questions from past chats, including the awkward ones.
- Run them through the agent and mark each answer correct, partly correct or wrong.
- For each miss, decide whether the fix is content, instructions or model choice.
- Repeat after every significant change.
This takes an afternoon and tells you more than any benchmark.
Handoff: the part the model should not do alone
Some questions should never be settled by a model: disputed charges, complaints, anything legal or safety related, and anything where the customer simply asks for a person. A good agent recognises these and passes the conversation to a human, with the history attached so the customer does not have to repeat themselves.
The reply at handoff matters too. Telling a customer exactly when someone will respond is only safe if the system knows the real schedule. Inventing a wait time is worse than saying nothing.
Doing this in AI SNS CHAT
AI SNS CHAT puts each of the steps above into settings you can inspect and change.
Give the agent knowledge. Under Data sources in the sidebar you can add Files, Text snippets, a Website, Q&A pairs, Notion and more. Connectors under Channels bring in help content from tools such as Confluence, GitBook, Zendesk Guide, Google Drive, SharePoint and others. Wiki connectors sync up to 1.5 million characters and 400 pages at a time, and the product tells you if anything was truncated.
Set persona and guardrails. Create an agent with the guided wizard, start from one of the 14 starter personas or the prompt library, and write the rules about tone, scope and when to hand off. Under Agent → Settings → AI you choose the model. Anthropic (Claude), OpenAI, Google (Gemini) and OpenRouter are supported for text, and if a provider is unavailable the agent fails over to the next one. Advanced models and model fallback need the Hobby plan or higher; lower plans use the provider's standard model.
Test before customers do. Use Playground to ask real questions, and the Playground A/B model compare to see two models answer the same message side by side.
Improve from real conversations. Conversations on every channel except email appear under Activity → Chat logs, where you can rate answers, flag, escalate or resolve them. Email lives in the Inbox. In both places you can edit a reply into the ideal answer to train the agent, so a corrected answer becomes future knowledge.
Protect the conversation. Each agent has prompt-injection and abuse filters plus per-device rate limits, configured in the agent's security settings.
Hand off to people. Under Helpdesk → Scheduling you set shifts per person and timezone, and handoffs are assigned to whoever is on shift. Alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. The "what happens next" line the customer sees is added by the server from the real rota; it never invents a wait time. You can also take over any chat yourself, which pauses the bot. Helpdesk needs Standard or higher.
Meter the cost. One reply uses one credit, and the Usage page shows credits per agent. See pricing for plan details and features for the full list of channels, which you can review under connections.
Conclusion
An LLM does not "know" your business; it writes well from whatever you put in front of it. Clean sources, clear rules, honest handoffs and regular testing with real questions are what turn a fluent model into a dependable support agent.
Frequently asked questions
Does the model learn from each customer conversation automatically?
Not by itself. A model's underlying training does not change during a chat. What improves an agent is the material around the model: its knowledge sources, its instructions and corrected answers you add. In AI SNS CHAT you can edit a reply into the ideal answer to train the agent.
Why does the agent sometimes answer from general knowledge instead of my documents?
Usually because no passage matched the question well enough to be retrieved. Add or rewrite content in the way customers phrase the question, and tell the agent in its persona to admit when information is missing and offer a handoff.
Can the agent search the web when my knowledge base has no answer?
In AI SNS CHAT, answering from the web when the knowledge base has none is available only on LINE WORKS, under Behaviour → Web search, and it cites the page used. For other agents, the Tavily action can give real-time web search; actions are limited by plan.
Which AI model should I pick for customer support?
Start with a standard model for straightforward FAQs and test a more advanced one if questions are complex. Compare both on a set of real customer questions in the Playground A/B model compare rather than guessing.
How do I stop people from tricking the agent into ignoring its rules?
Combine clear persona rules with technical filters. AI SNS CHAT offers per-agent prompt-injection and abuse filters and per-device rate limits, and you should keep sensitive information out of the knowledge base entirely.


