Can the bot work in several languages?

Yes, modern models handle many languages well, including Ukrainian. Test each language separately on real messages; retrieval and policy content must exist in (or be matched across) those languages.

Which helpdesks can you integrate with?

Any with an API — commonly Zendesk, Freshdesk, Intercom, HubSpot, Salesforce Service Cloud and custom systems.

Do we have to tell customers it’s a bot?

In the EU, the AI Act’s transparency rules generally require informing people they’re interacting with AI unless it’s obvious. It’s also simply good for trust. See EU AI Act for product teams.

Guide · Customer support AI

An AI support chatbot that doesn’t embarrass you.

Everyone has seen the screenshots: a support bot inventing a refund policy, agreeing to sell a car for one dollar, or insulting its own company. Those bots weren’t built badly by accident — they were built without scope, grounding, limits and a way out to a human. Here’s the blueprint for one that helps customers and protects your brand.

By Nythrex EngineeringUpdated 24 September 2026 3 min read

01
Customer message
Web, app, messenger
02
Understand
Intent · language · account
03
Answer or act
Approved content · scoped tools
04
Check
Policy · confidence · tone
05
Reply or hand off
Human gets full context

How support bots embarrass companies

01Invented policies

The model fills a gap with a plausible refund or warranty rule that doesn’t exist — and customers screenshot it.

02Jailbroken persona

Users talk the bot into rude, off-brand or absurd statements, then post them.

03Trapped customers

No way to reach a person, or a hand-off that forces the customer to repeat everything.

04Confident about stale content

An old help article outranks the current one; the bot quotes last year’s prices.

05Actions without limits

A bot that can issue refunds without caps or verification is a fraud channel.

The blueprint

1
Scope it on paper first
List the top intents by volume. Mark each: bot answers, bot acts (with which tool and limits), or hand off. Everything unlisted is a hand-off.
2
Ground every answer
Use RAG over approved help content and policies. Answers cite sources; no source, no answer.
3
Actions through narrow tools
“Check order status”, “cancel subscription”, “issue refund up to X after verification” — each with validation and audit, never generic access.
4
Guardrails on input and output
Detect prompt injection and abuse, keep the persona on-brand, block topics you never want discussed, and check replies against policy before sending.
5
Hand-off that respects the customer
One click to a human, with the conversation summary, detected intent and account context attached. Hand off automatically on low confidence, frustration or sensitive topics.
6
Measure and review weekly
Unanswered questions, hand-off reasons and thumbs-down replies drive content fixes and new test cases.

Metrics that tell the truth

Metric	Why	Trap to avoid
Resolution without hand-off	Real work taken off the team	Counting conversations where customers simply gave up
Customer satisfaction on bot conversations	Quality as customers see it	Only surveying resolved conversations
Hand-off quality	Agents shouldn’t have to re-ask	Ignoring the agent side of the experience
Answer accuracy on the test set	Catches regressions before customers do	A test set that never grows
Cost per resolved conversation	Keeps the business case honest	Ignoring model usage growth

Frequently asked questions

Example projectSupport copilot for a B2B SaaSIllustrative RAG project: a Zendesk support copilot drafting cited replies from help articles and tickets — discovery, PoC, architecture, lessons.ServiceRAG developmentRAG development: AI assistants that answer from your documents and data with citations, permissions and measured accuracy — built in your cloud.GuideLLM evals for non-ML teamsHow to evaluate an LLM feature without a data-science team: build a test set, choose metrics, use LLM-as-judge carefully, run evals in CI and read the results.GuideEU AI Act for product teamsEU AI Act for teams building on LLMs: risk categories, provider vs deployer, chatbot transparency, high-risk areas and dates after the 2026 AI Omnibus.IndustryAI for e-commerceAI for online stores: product content, search, support, returns, reviews and back-office automation — integrated with your platform and carriers.

Want a second opinion on your project?

Tell us what you’re building and where you’re stuck. We’ll reply within one business day with the most practical next step — even if that step isn’t us.

Start a project