AI in customer service is the use of AI to answer, complete a permissioned task, or reach a human — across the channels you already publish. A website chatbot is one interface inside that box, not the box.
Guzli wrote this page. Desk research of public explainers and product pages, read 19 September 2026. Zendesk’s AI in customer service overview and AI agent vs chatbot pages are the category language used below. Salesforce closed its acquisition of Fin (formerly Intercom) on 10 September 2026. Guzli sells a published-plan version of the box for small teams; this is not a ranked vendor list.
If you are buying a widget, use the buyer’s guide. If you need the phone in the same agent, use best AI chatbot that also answers the phone. Category boundaries: AI customer support software, chatbot vs AI agent.
What is in the category
Zendesk’s explainer says AI-powered tools “automate repetitive tasks, route requests intelligently, and equip agents with real-time context.” Their agent-vs-chatbot page draws the useful line: chatbots converse; agents can act across connected systems.
In practice the stack is:
- An inbox — Salesforce, Zendesk, Gorgias, or Guzli’s help desk
- Retrieval — answers from your help articles and, when connected, solved tickets
- Guardrails — what the system must refuse
- Handoff — a person can take over with a summary
- Permissioned actions — order lookup, subscription change, booking — only with rules
- The channels you publish — chat, email, SMS, WhatsApp, Slack, inbound voice
Named commerce tools, not “every integration”: Shopify, Stripe, BigCommerce, WooCommerce, plus Calendly or HubSpot if you use them.
What it is not
- A chat widget with no ticket path — that is a lead magnet
- A phone-only answering service — see best AI answering service
- A promise that AI replaces agents. Humans still own exceptions, sensitive cases, and anything the policy does not cover.
When it works — and when it does not
Use it when volume is repetitive, the policy is written down, and escalation is obvious.
Do not use AI-only support when the knowledge base is stale, nobody owns weekly review, or you want the bot to issue refunds it is not allowed to issue.
Guzli’s public offer (19 September 2026, /pricing/): Hobby $49/mo chat-only; voice from Standard $149 + 16¢/min. Handoff desks: Salesforce, Zendesk, or Guzli’s help desk. Not a Shopify App Store app.
Fin’s public offer (fin.ai/pricing): $0.99 per outcome; Voice is custom / sales-partnered. Salesforce closed its acquisition of Fin on 10 September 2026. Use those pages for live numbers — not screenshots.
A five-step rollout
1. One channel, one page
Pricing or a billing FAQ. Leave the rest of the site on the current inbox.
2. Connect a tight source set
Help center + the two policy pages that generate tickets. Measure accuracy on the top intents before you crawl the blog.
3. Write the escalation rule
Customer asks for a human, or the bot is unsure, or the action is out of bounds. Confirm the receiving agent sees the summary.
4. Add one action
Shopify order lookup, Stripe change under policy, or Calendly / Cal.com booking. Not five actions in week one.
5. Review weekly
Unanswered questions, escalation reasons, wrong actions, leads that were actually qualified. Then expand.
Business-case math: AI chatbot ROI and the ROI calculator.
Challenges
- Stale knowledge — resync and one source of truth
- Weak handoff — customers fight the bot; customer satisfaction (CSAT) falls on the transfer
- Unsafe actions — start with lookups; keep refunds permissioned
- No measurement — deflection without a re-contact window is a vanity number
FAQs
Is AI in customer service the same as a chatbot?
No. A chatbot is one interface. The category also includes routing, agent assist (drafts for a human), voice, and permissioned actions that change a system of record — the database that owns the order, subscription, or ticket. Humans still own exceptions.
What should I test first?
One channel and one page (pricing or a billing FAQ), the top intents on that page, then one permissioned action, then whether the receiving teammate sees the handoff summary. Do not crawl the blog or add five actions in week one.
How do I reduce invented answers?
Retrieve from help articles and solved tickets you control, refuse when the source is missing, and escalate when the customer asks for a human, the bot is unsure, or the action is out of bounds. That is design, not a model name.
When should we not use AI-only support?
When the knowledge base is stale, nobody owns weekly review, or you want the bot to issue refunds it is not allowed to issue. Start with lookups; keep refunds permissioned.