Articles on: Getting Started

Getting started with AI for customer service at Crisp

Learn how AI fits into Crisp customer service, and where to start if you want practical results quickly.


Crisp AI works best when it is rolled out as a layered system. Hugo can answer, route, escalate, and use integrations when configured; Search Chatbox AI can reduce simple questions before they become conversations; and agent-facing tools can help your team draft, understand, and handle conversations faster.


This article covers the main AI layers you can start with:



How AI fits into Crisp


The goal is not to turn every AI feature on at once. A stronger setup usually starts with reliable training resources, a tested Hugo configuration, and then the additional AI layers that match your workflow.


The main AI layers to know are:

  • Hugo → the customer-facing AI Agent that can answer questions, route conversations, escalate to humans, and use integrations when configured
  • Search Chatbox AI → the self-service search layer inside the chatbox, useful when visitors should search for answers before contacting your team
  • Workflows → guided automation flows that can collect information, route users, trigger actions, and hand a conversation to Hugo when AI should take over
  • AI Tools in the composer → agent-facing actions such as Predict, Rephrase, Fix grammar, Adjust tone, and Expand
  • Copilot → an internal assistant that helps operators ask questions, understand policies, or draft better replies without sending anything automatically
  • Conversation intelligence → summaries, topics, and transcription features that help your team understand conversations faster


This layered approach helps you keep control. Customers can self-serve when the answer is simple, Hugo can handle repetitive conversations, and human agents can still review, personalize, and escalate whenever needed.



Start with Hugo


If you want the strongest AI foundation first, start from Crisp and open AI Agent from the left menu. This is where you configure Hugo, train it, test it, monitor it, and decide when it should go live.


Configure identity and behavior


Open AI Agent → Agent → Settings to define how Hugo should present itself and what business context it should understand. This includes the agent identity, avatar, business description, sources to use, escalation behavior, auto-resolution settings, and office-hours awareness.


This step matters because Hugo performs better when the business context is explicit. A good business description should explain who you are, what you sell or provide, who your customers are, which channels or stores matter, and which support expectations must be respected.


Useful setup resources:


Train Hugo on reliable resources


Open AI Agent → Train to give Hugo the content it should rely on. Good training matters more than adding many complex prompts. If the content is outdated, vague, duplicated, or contradictory, Hugo will have a harder time producing consistent answers.


The best first training resources are usually:

  • Website content → product pages, documentation, pricing explanations, policy pages, onboarding pages, and other public resources
  • Knowledge base articles → the support answers you already want customers to find and reuse
  • Q&A snippets → short internal answers for recurring edge cases, policy clarifications, or answers that are not worth publishing as full articles
  • Files → TXT, CSV, or PDF resources when the useful content is not already available online


Q&A snippets should be informative. Use them to answer questions, not to define Hugo's personality or behavior. Tone, escalation rules, and behavioral guidance belong in Instructions instead.


Read How to train Hugo AI Agent on your data if you want examples of strong training resources.


Test before activation


Use AI Agent → Evaluate → Playground to test Hugo like a customer would. Ask realistic questions, including easy questions, vague questions, follow-ups, unsupported requests, and cases where Hugo should escalate or trigger a workflow.


Testing helps you spot missing policies, weak snippets, confusing instructions, or routing rules that are too broad. It is also safer than activating Hugo immediately on all incoming conversations.


Once answers feel reliable, go to AI Agent → Agent → Activation. From there, decide whether Hugo should answer new conversations by default, only handle selected channels, or start only when a routing rule detects a specific situation.


Add integrations when Hugo should perform real actions


Hugo is not limited to answering questions from documentation. When you connect native integrations or custom MCP servers, it can retrieve live information or perform actions within the rules you define.


This is what unlocks use cases such as checking an order status, finding a subscription, retrieving account details, or preparing a refund flow without immediate human handoff.


Useful integration resources:



Add AI across the customer journey


Once Hugo is properly configured, you can extend AI to different moments of the customer journey. Some features act before a conversation starts, others help during automation, and others support the human operator inside the Inbox.


Search Chatbox AI


Search Chatbox AI helps visitors look for answers before they open a conversation. It is useful when many users ask simple, repeated questions and your help content is already in good shape.


You can enable it from Crisp by opening Settings → Chatbox Settings → Search Chatbox AI. From there, you can activate the feature, configure frequent searches, add categories, and decide whether visitors must search before contacting your team.


Search Chatbox AI displayed before chat contact


Use this layer carefully. It can reduce repetitive questions, but it should still make it easy for visitors to reach a human when they genuinely need help.


Read How do I setup Crisp Search Chatbox AI on my chatbox widget? for the full setup.


Workflows and Answer with Hugo


Workflows are useful when you want a structured path before, after, or around an AI answer. For example, a workflow can collect an order number, ask the user to choose a topic, update conversation data, route the user to a team, or trigger Hugo only when AI should handle the next part of the conversation.


Use Answer with Hugo when a workflow should hand the conversation to the AI Agent. This keeps the setup aligned with the current Hugo-first experience while still letting you design guided flows.


Common workflow patterns include:

  • Qualification → collect context before routing a sales, support, billing, or partnership request
  • Handoff → gather the required details before escalating to a human agent
  • Self-service → guide users through choices and send them to Hugo when a natural-language answer is better
  • Post-resolution → ask for missing information, collect feedback, or send the next step after Hugo has helped


Read Getting started with Workflows and How to integrate Hugo in your Workflows to build these flows cleanly.


AI Tools in the composer


Inside conversations, operators can use AI Tools to improve or accelerate a reply before sending it. This is useful when the customer should still receive a reviewed human answer, but the first draft can be faster.


The most useful actions to know are:

  • Predict → generate a draft reply the agent can review, edit, and send
  • Rephrase → rewrite the current message more clearly without changing its meaning
  • Fix grammar → clean up spelling and grammar quickly
  • Adjust tone → make a reply friendlier, more formal, or more direct
  • Expand → turn a short answer into a fuller explanation that is easier for customers to understand


This is a good layer to introduce early to operators because it does not automate the customer reply. The agent stays in control of the final message.


Copilot for internal help


Copilot is the internal assistant your team can open from the Inbox sidebar. It is meant for agents, not customers, and it helps them ask questions, understand policies, or draft better replies while keeping the final send under human control.


Because it can rely on the same core resources as Hugo, Copilot is especially useful when an agent needs help finding the right answer quickly without leaving the conversation.


Hugo Copilot in the Inbox


Summaries, topics, and transcription


AI is also useful after a conversation already exists. Conversation summaries help agents catch up faster, topic detection helps teams identify support patterns, and voice-to-text transcription makes audio messages or recorded calls easier to process.


These features are especially helpful when conversations are long, handled by several teammates, or coming from channels where context can be fragmented.



Roll out AI progressively


A simple rollout usually works better than an ambitious one. The best sequence depends on your team, but most workspaces should start with training quality and only then add more advanced layers.


A strong rollout order is:

  • Train Hugo first → website pages, knowledge base articles, Q&A snippets, and files before advanced routing
  • Test in the Playground → catch missing information, unclear policies, and weak instructions early
  • Activate Hugo carefully → start with trusted channels, clear topics, or lower-risk support cases
  • Add Search Chatbox AI → reduce repetitive pre-chat questions once your content is reliable
  • Use workflows for structure → collect information, route requests, and hand conversations to Hugo when useful
  • Scale with agent-facing AI → let operators use AI Tools, Copilot, summaries, topics, and transcription to work faster


Do not use AI to hide a missing process. If customers ask the same unclear question repeatedly, fix the underlying article, policy, or workflow first. Hugo and Copilot will both perform better when the source material is clean.



Next steps


Once your first AI layer is live, the best improvements usually come from better training resources, clearer routing, and better integrations rather than more prompt complexity.


Helpful resources to continue:


Updated on: 04/05/2026

Was this article helpful?

Share your feedback

Cancel

Thank you!