Crisp AI: Complete Guide to Hugo, Features & Limitations (2026)

If you're reading this, someone has probably asked you to "add AI to our chat", and Crisp keeps coming up as the cheap option. That's fair, because it is genuinely cheap. But I'd want to know whether cheap is the same as right for your team, and what you're signing up for when you switch Hugo on.

A few situations tend to bring people here:

1. You already run Crisp's live chat and you're wondering whether Hugo is worth turning on.
2. You're shopping for low-cost AI chat and weighing Crisp against Intercom's Fin or Tidio's Lyro.
3. You're already on Zendesk or Intercom and want to know if you can bolt Crisp's AI onto what you have. (Short answer, up front: you can't, and that turns out to matter a lot.)

I'll walk through what Crisp's AI actually is, how Hugo works, what you can train it on, where it falls short, what it costs, and who should pick it, versus layering AI onto the helpdesk you've already got.

What is Crisp AI?

Crisp is a bootstrapped French SaaS company, based in Nantes, that started life as a live-chat widget and rebuilt itself into what it now calls "the AI Customer Support platform". Crisp AI is a whole stack of AI features layered onto Crisp's core chat product rather than a single tool, and the flagship piece is Hugo, Crisp's autonomous AI agent.

Hugo is bundled into Crisp's paid plans. There's no standalone Hugo subscription you can buy on its own; to use the AI, you use Crisp.

Around Hugo sit a handful of other AI features, some newer than others:

• Hugo: the autonomous agent that answers customer questions and runs no-code workflows.
• MagicReply: the older AI-draft feature, originally built on OpenAI models, now folded into an "Answer with Hugo" workflow block.
• AI Overlay: an AI-powered self-service search widget that Crisp says can "deflect 30-50% of simple requests".
• Copilot: an agent-facing assistant that suggests answers to your human operators and, in Crisp's words, "never writes to the customer".
• AI Writing tools: composer actions like Predict, Rephrase, Fix grammar, Adjust tone, Expand, Summarize and Translate.

In plain English: Hugo is the brain that talks to customers, Copilot is the brain that helps your agents, and the Magic and Overlay features are the older self-service layer around them.

G2: Crisp scores 4.5/5 from 194 reviews on G2. "Until recently, we had to outsource our AI chatbot service, but Crisp made massive improvements in that, and now we don't have to pay for our outsource anymore." via a verified G2 reviewer.

How easy is it to set up Crisp's AI (Hugo)?

Crisp markets Hugo as fast to deploy, with "try it yourself in under 5 minutes" on Hugo's homepage. The documented setup is a four-step loop: train the AI on your content, build a workflow, test it in the Playground, then deploy and measure.

The onboarding itself is no-code for the standard path. You give Hugo a name and avatar, write a description of your business, set escalation rules and office hours, point it at your knowledge sources, and test it in the Playground: a sandbox where you can chat with Hugo, switch AI models on the fly, and open a "Guardrails drawer" that shows which sources Hugo used for each answer.

Reality is a little messier than "under 5 minutes". Word-on-the-street (and the G2 reviews) is that the AI is "quite hard for beginners to grasp", and getting Hugo genuinely good takes ongoing tuning rather than a single afternoon.

There's also a billing wrinkle worth knowing before you start: Hugo credits are consumed during the free trial, so testing and training Hugo bills against your usage. The MCP Server for custom tool integrations is Plus-tier only and does need engineering work.

What channels does Crisp's AI work in?

Hugo deploys across web chat, WhatsApp, Messenger, Instagram and email. The wrinkle is that channels are aggressively tier-gated, so which ones you get depends on how much you pay:

• Free: chat widget and contact form only.
• Mini ($45/mo): adds email, Telegram, Facebook Messenger and X (Twitter) DMs.
• Essentials ($95/mo): unlocks the full omnichannel inbox, including WhatsApp Business, Instagram, SMS, Viber, Line and Discord.
• Plus ($295/mo): keeps all of the above, plus white-labeling and 100+ integrations.

The notable gap is voice. Crisp has no native phone or PSTN support; any phone capability comes through third-party integrations like Aircall or Ringover, and the Twilio plugin is SMS-only. There's also no native support for WeChat, KakaoTalk or TikTok, which matters if your audience leans toward those channels.

What are the limitations of Crisp's AI?

Crisp's AI limitations show most in daily use, where Crisp's marketing reads more polished than the experience some long-term customers describe. The most pointed critique on G2 comes from a six-year customer who rated Crisp 1.5 out of 5 in April 2026: "Crisp's adoption of AI… has unfortunately been slow… It feels like the focus on AI is 90% marketing, 10% implementation" via a G2 review.

Here are the limitations that come up most, and the ones most likely to weigh on your decision:

• Crisp is all-or-nothing. This is the big one. Because Hugo is bundled into Crisp's platform, you can't add Crisp's AI on top of an existing helpdesk (the way you'd switch on Fin inside Intercom, or drop a third-party agent into Zendesk). If you already run Zendesk or Intercom, there's no way to keep your stack and just plug in Hugo: using the AI means moving your whole support operation onto Crisp, which is a migration rather than a quick bolt-on.
• Crisp won't train its AI on your past tickets. Crisp is explicit that it does "not use your Inbox Messages to generate automated AI answers to your users for privacy reasons". Your resolved conversations only ever inform agent-facing Copilot suggestions, never the answers Hugo gives customers. That's a meaningful limit, because past tickets are some of the richest training material a support AI can have.
• The best AI features are gated behind the $295 Plus tier. Multilingual AI, the AI Writing Assistant, the MCP Server, AI site search, the spam filter, auto-tagging and conversation summaries are all Plus-only. On the cheaper plans, "Crisp AI" is a noticeably thinner product than the marketing suggests.
• There's no AI in the mobile app. The same long-term G2 reviewer flags it directly: "there's no AI implemented in the app at all (arguably it's the place you need it the most)."
• Analytics are thin. A Capterra reviewer notes you're "unable to extract raw data when doing an export via analytics, only topline numbers come through." If you rely on custom reporting or deep trend analysis, that's a known weak spot.

This is also where our own experience comes in. We used Crisp ourselves in the early days, and it made sense as a free live chat on our site. But as we grew, the feature set started to feel limiting, and we moved to Intercom.

The snag with that move is the one I keep coming back to: when you outgrow Crisp and switch helpdesks, you can't take your AI agent with you. All the training and tuning you put into Hugo stays behind.

What knowledge can I train Crisp's AI on?

Hugo and MagicReply pull from six types of data source:

• Answer Snippets: short Q&A pairs you write yourself, up to 1,000 characters per question.
• Knowledge Base articles: synced automatically with Hugo on create, update or delete.
• Web Pages: crawled by Crisp's bot from one or more domains, though only one domain re-imports at a time.
• Files: TXT, CSV and PDF only, and only one document imports at a time.
• Inbox Messages: your past conversations, but only for agent-facing Copilot suggestions, never for customer answers (see the limitation above).
• MCP integrations: live data and actions from connected systems like Shopify and Stripe natively, plus custom servers on Plus.

Crisp also auto-imports knowledge bases from a long list of other tools when you migrate, including Zendesk, Intercom, Freshdesk, Gorgias and Help Scout. A per-block confidence slider lets you limit how willing Hugo is to answer when it's unsure, which is a reasonable hallucination guard.

What features does Crisp's AI have?

Crisp's AI surface area is broad but stratified by tier:

• The AI chatbot and no-code workflow builder start on Essentials.
• The Hugo agent is available from Mini upward, but limited by AI credits.
• The internal Copilot and AI knowledge training start on Essentials.
• The richer features (multilingual AI, the AI Writing Assistant, the MCP Server, AI site search, the spam filter, auto-tagging and conversation summaries) are Plus-only.

A genuinely useful detail in Crisp's model: workflows themselves don't consume Hugo credits, only AI-generated answers do. So deterministic, rules-based chatbot flows are effectively free, and you only pay when the AI brain actually writes a reply. The Playground lets you test with different models, and Plus adds access to 100+ integrations including HubSpot, Salesforce, Zapier and Stripe.

How do I improve Crisp's AI responses?

Crisp's documented improvement playbook is a layered rollout: train Hugo on your existing materials, test in the Playground, activate on trusted channels first, then add the self-service Overlay, then structured workflows, and finally the agent-facing AI tools. Crisp recommends weekly reviews of unsuccessful answers and only expanding the AI's scope once you've hit around 75% accuracy on existing question types.

The weak spot is tooling for this. There's no published evaluation framework, no documented A/B testing for prompt variants, and no quantitative feedback dashboard in the public docs.

The Playground is the closest thing to a QA workbench, and optimization is largely manual. As one G2 reviewer put it, the AI "requires you to build custom workflows, and this isn't easy."

Play video

What resolution rate can I expect from Crisp's AI?

Crisp's headline claim is "resolve up to 60% of your support tickets" on Hugo's homepage, with the homepage framing it as "automate 50% of your inquiries" and the self-service Overlay claiming it can "deflect 30-50% of simple requests".

Crisp's own published customer numbers cluster in the 40-60% band, and I read these as vendor-selected wins rather than a fair average. AFS Foiling reports "around 60% of requests are fully automated"; Emma App says Hugo "resolves approximately 40% of queries autonomously"; Spider-VO reports "around 40% of incoming requests are now fully automated". A Capterra reviewer independently reports "a 40% reduction in queries… due to their AI handling mundane queries".

Two caveats matter when you read those numbers. First, there's no independent third-party benchmark of Crisp's AI; every figure is either Crisp's own marketing or a customer self-report.

Second, the line between deflection and true resolution is blurry. If a conversation that escalates to a human counts in the "resolved" figure, the headline overstates how much Hugo handles on its own, and Crisp doesn't publicly document what counts as a Hugo-resolved conversation. (If that distinction is new to you, we wrote up containment vs deflection vs resolution separately.)

For context, the field median across 38 AI support vendors and 195 rated deployments sits at about 70%, from our own resolution-rate benchmark research. That number comes with a health warning, though: the metric label moves it more than capability does.

A vendor reporting "resolution" tends to look about 12 points higher than one reporting "automation" or "deflection", because each counts a different event. Crisp's "automate 50% / deflect 30-50%" framing isn't measured the same way as a resolution number, so treat any cross-vendor comparison as indicative rather than precise.

What AI model does Crisp use?

Crisp is flexible on models and lets you pick the one behind Hugo. It offers a choice across five model families:

• OpenAI GPT-5 and GPT-5.1 as the default general-purpose options.
• Anthropic Claude 4.5 Sonnet and Haiku for more complex reasoning.
• Mistral Large and Medium as EU-hosted alternatives.
• Google Gemini in technical preview.
• A proprietary internal model that Crisp develops and hosts itself on Scaleway infrastructure in Paris, recommended for organizations that want maximum GDPR isolation, since no data leaves Crisp's own servers.

Two internal models run alongside whichever LLM you pick: a guardrails model that checks responses against your knowledge, and a spam-fighting model. Crisp is also clear that Hugo isn't trained on your inbox conversations and doesn't reuse other customers' messages as a training source.

What languages does Crisp's AI work in?

The numbers vary by component. Crisp's chatbox UI is translated into 60+ languages, the Hugo agent reportedly covers 85+, and LiveTranslate (its real-time message translation) handles 50+. Cross-language retrieval works, so a French-trained knowledge base can answer questions asked in English.

There's a real French-and-EU strength here. Crisp is a French company, it offers Mistral models, and its pricing page proudly reads "Made in Europe". The watch-out is that packaged Multilingual AI is a Plus-only feature, so multilingual support as a built-in capability sits at the top tier.

How secure is Crisp's AI?

Crisp's security story is strong on GDPR and light on enterprise certifications. Both sides deserve a precise look.

On the strong side, Crisp is fully GDPR-compliant as a French company under EU law, with a designated DPO and fast response windows for data requests and breach notifications. All customer data sits in the EU: messaging servers in the Netherlands, plugin servers in Germany, and the proprietary AI model in Paris.

If EU data residency is your priority, Crisp delivers it. The flip side is there's no US data-residency option, which can complicate some US enterprise procurement.

On the lighter side, Crisp is candid in its own documentation that it is not SOC 2 audited (in its words, "Crisp hasn't been over any SOC2 audit"), and there's no ISO 27001 certification and no publicly confirmed HIPAA BAA.

Who is using Crisp's AI?

Crisp publishes a long list of named customers and Hugo case studies, weighted heavily toward EU and francophone SMBs and Shopify-based e-commerce. Named examples include Decathlon, Emma App (around 40% autonomous resolution), Reedsy (which switched from Intercom and Zendesk), Hoxton Mix, AFS Foiling (around 60% automation), B&B Hotels and Etnia Barcelona.

One thing to read carefully is Crisp's customer count, which varies wildly by source. Its marketing cites "10,000 companies" using Hugo, while its About page and third-party profiles cite "600,000+", a gap that almost certainly reflects the difference between paying customers and everyone who's ever installed the free widget.

Crisp itself is bootstrapped, run by a roughly 20-person remote team out of Nantes, and was founded in 2015 by Baptiste Jamin (CEO) and Valérian Saliou (CTO). No Fortune 500 customers show up in independent sources, and Crisp's center of gravity is firmly SMB.

How much does Crisp's AI cost?

Crisp uses flat per-workspace pricing rather than per-seat, which is a real positive for small teams:

• Free: $0 forever, 2 seats, no AI credits.
• Mini: $45/mo, 4 seats, $5 of AI credits (about 90 Hugo conversations).
• Essentials: $95/mo, 10 seats, $25 of credits (about 450 conversations).
• Plus: $295/mo, 20 seats, $75 of credits (about 1,350 conversations).

Seats are capped per plan, with extra agents at $10 each per month. Hugo itself is billed per conversation, from the first AI-handled message until it's resolved, routed or escalated, at roughly $0.05 to $0.10 each.

The included credits are a dollar value rather than a fixed conversation count, and beyond them you can enable Pay-As-You-Go; without it, Hugo simply stops answering when credits run out. Credits reset on the 4th of each month and don't roll over.

That per-conversation cost is genuinely low, roughly 20× cheaper than Intercom's Fin at around $0.99 per resolution. The watch-out is what the cheapest plans really get you: with only $5 of credits, Mini covers fewer than a hundred AI conversations a month, and the genuinely useful AI features don't arrive until Essentials and Plus.

Does Crisp's AI have a free trial?

Crisp offers a 14-day trial with all features and no credit card required, alongside a permanent Free plan that stays available indefinitely (2 seats, 100 customer profiles, the chat widget, but no AI credits and no email channel).

The gotcha to plan around: Hugo credits are consumed during the trial, and Pay-As-You-Go isn't available during it. So your testing and training burn through the trial's credits, and when they run out, Hugo stops, which can cut your evaluation short right when you're getting useful signal.

Is Crisp's AI worth it?

Whether Crisp's AI is worth it depends on who you are, and the picture is more mixed than the marketing.

The scores on the doors tell a consistent story for the right buyer: Crisp scores 4.5/5 on G2 (194 reviews), around 4.5 on Capterra, 4.1 on the Shopify App Store and 4.8 on Product Hunt. The enthusiasm is real among the SMBs and e-commerce teams Crisp is built for.

If you're weighing Crisp against running AI inside the helpdesk you already have, here's the quick version:

The bottom line is straightforward. If your volume is moderate, your knowledge base is tidy, your channel mix is web plus WhatsApp plus email plus e-commerce, and you operate in the EU, Crisp's flat pricing and EU residency are real advantages.

If any of those fail (enterprise compliance, complex reporting, an existing helpdesk you want to keep), the limitations stack up quickly. And if the lock-in question is what's nagging you, I'd take it seriously: the AI you train on Crisp is AI you can't take with you.

For the wider field, see our roundup of the best AI customer service agents for 2026.

What are the pros and cons of Crisp's AI?

Pros

• Genuinely cheap AI: at roughly $0.05-$0.10 per Hugo conversation on flat per-workspace plans, Crisp is one of the lowest-cost ways to put an AI agent on your chat.
• Rare model choice with a strong EU story: five model families, including a Paris-hosted proprietary model, give Crisp an unusually flexible, GDPR-friendly setup.
• Fast, no-code setup with broad SMB channels: the four-step Playground flow and omnichannel reach make it quick to get live for a small team.

Cons

• All-or-nothing platform: you can't run Hugo on top of an existing helpdesk, so using the AI means moving everything onto Crisp.
• No training on past tickets: Crisp won't use your resolved conversations to answer customers, which caps how good the AI can get.
• No SOC 2 or ISO 27001: Crisp is GDPR-strong but lacks the enterprise certifications procurement teams expect.
• Best AI is Plus-gated, and absent on mobile: the most useful AI features sit behind the $295 tier, and the mobile app has no AI at all.

FAQs