GuideUpdated 5 min read
AI customer support: how it works, what it costs, and how to roll it out
Campfire guide to AI customer support: how an AI agent answers from your knowledge base, how vendors price it, what to automate first, and rollout steps.
AI customer support means software that reads a customer question and replies on your behalf, using a language model grounded in your own help content, and passes the conversation to a person when it should. This guide explains the moving parts, how the main vendors charge for it, what is sensible to automate first, and how Campfire approaches it.
How an AI support agent works
Most systems follow the same loop. The question arrives through a channel such as a website widget, email or WhatsApp. The system retrieves the passages of your knowledge base that look relevant, gives them to a language model with instructions about tone and limits, and drafts an answer. A confidence check decides whether to send the answer or hand the conversation to a human.
- Grounding. Answers are built from your articles rather than the model's general knowledge, which is what keeps replies specific to your product and lets the agent cite the article it used.
- Handoff. A configurable confidence threshold decides when the agent stops and a person takes over. Without a handoff path an AI agent is a liability, not a saving.
- Channels. The same agent can answer on a web widget, email and messaging channels, so the quality of one knowledge base carries across all of them.
- Agent assist. The same retrieval can draft replies for your human team to review and send, which is a lower-risk first step than letting the agent answer alone.
What to automate first
Start with questions that already have a single, stable, documented answer, and keep everything that involves money, accounts, legal commitments or judgment with people. A good first list comes from your own inbox: export or skim a few weeks of conversations and group them by topic.
- Good candidates: how-to questions with an existing article, order or account status lookups that are already self-serve, opening hours, supported channels, plan and pricing explanations.
- Keep with humans: refunds and disputes, anything emotionally charged, security or privacy requests, bugs that need investigation, and anything the knowledge base does not clearly answer.
- If an article is missing for a common question, write the article first. The agent is only as good as the content it retrieves.
How vendors price AI support
AI is billed differently by almost every vendor, and the unit (an outcome, a resolution, a conversation or a credit) is defined differently too. The figures below were read from each vendor's public pricing page in October 2026; check the page before you buy.
| Product | How AI is priced | Source |
|---|---|---|
| Intercom (Fin) | $0.99 per outcome; minimum monthly commitment for helpdesk use | intercom.com/pricing (as of 2026-10) |
| Zendesk (AI agents) | Some automated resolutions included per agent; then $1.50 committed or $2.00 pay-as-you-go each | zendesk.com/pricing (as of 2026-10) |
| Help Scout (Envoy) | $0.75 per resolution after a 3-month free trial of unlimited resolutions | helpscout.com/pricing (as of 2026-10) |
| Front (Autopilot) | Add-on starting at $0.05 per conversation | front.com/pricing (as of 2026-10) |
| Plain (Ari) | 100 credits per resolution; plans include 1,000 or 2,500 credits a month, top-ups from $10 per 1,000 | plain.com/pricing (as of 2026-10) |
| Crisp | Plans include a monthly AI credit allowance ($5, $25 or $75 on paid plans) | crisp.chat/en/pricing (as of 2026-10) |
| Campfire | No per-resolution fee. Managed AI uses a monthly token allowance (Starter 50K, Growth 500K); Growth can use your own provider keys | campfire.is/pricing |
Per-resolution pricing has a useful property: you only pay when the AI succeeds. It also means the bill rises as automation improves. Flat or allowance-based pricing makes the bill predictable but comes with caps. Work out your own monthly conversation volume and the share you expect the agent to handle before comparing. The Intercom pricing calculator models one such case.
Questions to ask any vendor
- What exactly counts as a billable resolution or outcome, and what happens when a customer asks for a human afterwards?
- Can I set a spending cap or a hard limit, and what does the agent do when it is reached?
- Which models does it use, can I bring my own provider keys, and is my conversation data used to train models?
- How does it decide when to hand over, and can I tune that threshold per channel?
- What does the agent do when the model provider is down? A canned reply is not a good answer.
- Can I see the sources it used for each answer, and measure resolution rate myself?
- Which features sit behind higher plans or add-ons, and which channels cost extra?
A rollout plan that limits risk
- Write or tidy the knowledge base. Cover your twenty most common questions with one clear article each.
- Run it in assist mode. Let the AI draft replies that your team approves, and note where drafts are wrong or the article is missing.
- Enable autonomous answers for one channel and a narrow topic list. Set a conservative confidence threshold so unsure answers go to a person.
- Review conversations weekly. Read a sample of AI-handled threads, fix the underlying articles, and expand the topic list only when the sample is clean.
- Track resolution rate and customer satisfaction from your own data. Treat any vendor-quoted resolution rate as marketing until you have measured yours.
How Campfire approaches AI support
Campfire is a support inbox with an embeddable widget. Its AI agent answers from your knowledge base using semantic retrieval, shows the article behind each draft, and hands the conversation to a human when confidence falls below a threshold you configure. Agents also get inline draft suggestions beside the thread. Source access is available on request.
- Bring your own model: point Settings, AI Providers at OpenAI, Claude or Gemini via OpenRouter, Groq, Together, or a local Ollama instance. You pay the provider directly and Campfire adds no per-resolution markup. See Bring your own AI.
- Managed AI is also available at the same plan price, with a monthly token allowance per plan.
- If every AI provider is unavailable, Campfire does not send a canned reply. On messaging channels it records the failure and hands the conversation to a human.
- Analytics show resolution rate split between AI, agent and escalated conversations, plus response times and CSAT.
- Campfire does not train models on customer data, per its data processing addendum.
What Campfire does not do yet: no voice or phone support, no native mobile app, no SAML or SCIM, no product tours, and SOC 2 Type II is in progress rather than complete. Plans have usage caps (Growth: 2,500 conversations and 500K managed AI tokens per month).
To try it, start the 14-day trial (no card required) or look at the alternatives comparisons to see how it stacks up against the tools above.