What an AI customer-service bot actually does
A modern support bot does four jobs: it answers common questions from your help content, deflects tickets that would otherwise reach a human, triages and routes the ones it cannot solve, and assists agents by drafting replies and summarising threads. The best platforms ground answers in your knowledge base (so they cite real policy rather than hallucinate), recognise when they are out of their depth, and escalate to a human with the full conversation attached. Judge any tool on those behaviours, not on demo polish.
The packaged platforms
Intercom Fin is built on frontier LLMs and is known for natural conversation and strong knowledge-base grounding. Its headline feature is resolution-based pricing — you pay per query it successfully resolves — which aligns cost with value but can become expensive at high volume. Best for product-led companies already on Intercom.
Zendesk AI layers generative answers, intent detection, and agent-assist on top of the most established help-desk suite. Its strength is depth of ticketing, reporting, and workflow automation; AI is bundled into higher tiers and add-ons. Best for established support orgs that already run Zendesk.
Freshdesk Freddy offers a capable bot plus agent-assist (suggested replies, summaries) at generally lower price points, making it attractive for small and mid-size teams. The AI is solid rather than class-leading, but the value-for-money is strong.
The custom GPT or Claude route
Building directly on the OpenAI or Anthropic API — typically with a retrieval-augmented-generation (RAG) layer over your docs — gives you total control: choice of model, exact tone, strict data handling, and deep integration with internal systems a packaged tool cannot reach. The tradeoff is that you own everything: the retrieval pipeline, escalation logic, analytics, guardrails, and ongoing maintenance. This path makes sense when packaged platforms genuinely cannot meet a requirement, or at a scale where their per-resolution or per-seat pricing exceeds the cost of building.
How to choose
Run a structured evaluation on your own ticket history rather than trusting vendor claims:
- Resolution rate on your real, representative questions — not curated demos.
- Escalation quality — does it hand off gracefully with context, or loop and frustrate?
- Multilingual coverage if you serve global customers.
- Integration depth with your help desk, CRM, and order/account systems.
- Pricing model mapped against your actual monthly volume (resolution-based vs seat-based vs token-based can differ by an order of magnitude).
Start with the packaged platform that matches your existing stack, measure honestly, and only consider a custom build when you hit a wall the platform cannot scale past.