“Up to 98%” is the share of incoming customer conversations the AI resolves fully, with no human needed. It is a ceiling we observe when a complete knowledge base meets a highly repetitive question mix — not an average, and not a promise. Most deployments settle between 60% and 90%+, driven mostly by how much of your inbox is repeat questions.
A conversation counts as AI-resolved when the customer’s question is answered end-to-end by Clare AI — no human agent joins, and the customer doesn’t ask for one. The admin console that ships with every deployment reports this directly: AI-resolved conversations ÷ all conversations, over any date range. It’s the same definition Intercom and Zendesk use to bill you per resolution — we just don’t charge for it.
The high end appears under three conditions at once: the knowledge base genuinely covers the questions customers ask, the deployment is connected to live data (orders, deliveries, account status) so the AI answers specifics instead of deflecting, and the question mix is dominated by repetitive, factual asks — “where is my order?”, delivery times, menus, deposit and withdrawal questions, opening hours, returns.
Honesty first: we run a small number of production deployments today, and we publish what we observe in them. This is our own measured experience — not an industry benchmark, and we don’t present it as one.
| Pushes it up | Pushes it down |
|---|---|
| A complete, well-organised knowledge base (your top questions, written out) | Sparse knowledge base — the AI can’t answer what it was never told |
| Live integrations: real order, delivery and account lookups | No integrations — “check your order status” answers stay generic |
| Repetitive question mix (status, pricing, policies, hours, how-to) | Dispute-heavy or judgment-heavy mix (refund exceptions, complaints, negotiations) |
| Clear escalation rules — the AI hands hard cases to humans early | Forcing the AI to attempt everything (hurts trust and the metric) |
| A few weeks of tuning: adding answers the AI missed | Judging it on day one, before the knowledge base has been fed |
Pull your last 100 support conversations. Count how many are variations of your 20 most common questions. That percentage is a realistic first-year ceiling for AI resolution — because those are exactly the questions a knowledge base can fully cover.
E-commerce (order status, returns, stock): 75–90%
iGaming (deposits, withdrawals, bonuses, KYC status): 70–90%
F&B (orders, delivery, menu, hours): 80–95%
Complex B2B (contracts, disputes, custom quotes): 40–60%
The cost calculator assumes roughly 70% AI resolution when comparing against per-resolution vendors — deliberately below our best observations, so the savings math stays conservative.
It is not an average across all customers. It is not guaranteed for your mix. It is not instant — the metric climbs over the first weeks as your knowledge base fills the gaps the AI reports. And it is not a reason to remove humans: the escalation path to a real person is a feature, not a failure.
The chat bubble on our homepage is not a scripted demo — it’s the actual product, answering from our own knowledge file. Ask it something hard.