Realistic ranges are 40-70% deflection within three months of a well-tuned deployment. Variables: how well-structured your KB is, how often you update the known-issue list, how clean the account-state tool function returns are. Teams that start with a rough KB see lower numbers initially and improve as logs feed back into KB gaps. Teams with mature docs and active incident pages hit the high end fast.

For logged-in users, the bot reads the WordPress user record and any linked account data via a registered tool function. For anonymous users, the bot can collect the email or account ID and call a verified-lookup function (which may require a one-time code) before showing account-specific data. The bot is explicitly forbidden from showing data to unverified sessions. The verification step is the security gate.

Yes for clearly bounded cases. A tool function for 'apply discount code' or 'change plan to monthly' can be called by the bot when policy criteria are met (e.g. refund within first 30 days, no past chargebacks). For anything outside the policy box (partial refund, custom plan, disputed charge) the bot routes to a human with the full case context. Drawing the box conservatively is the right default and easy to widen later.

Help-doc search returns articles. The bot returns an answer. It reads the relevant article into context, applies it to the user's account state, and produces a tailored response. 'Article 27 says X, but your plan is Enterprise so the relevant clause is Y, here is what to do.' The article search forces the user to do that synthesis themselves. Most users do not.

Works with extra guardrails. In regulated sectors, the system prompt is more conservative about what counts as resolvable vs escalatable, the account-state tool function redacts PHI or PII before returning to the model, and the conversation log is encrypted at rest with appropriate retention. The model provider should also be on a HIPAA BAA or equivalent contract for your region. SleekAI's architecture (your WordPress, your API key, your provider contract) makes this straightforward.

Two integrations. Read: the bot pulls the active-incidents feed from Statuspage, BetterStack, or a WordPress incident post type on every conversation, so symptoms matching active incidents get acknowledged immediately. Write: when many users report the same novel symptom in a short window, the bot can flag it as a potential incident candidate to your on-call team. Both prevent the classic 'we did not know it was happening' incident lag.

Weekly review surfaces three patterns: (1) cases the bot resolved successfully (validate the answer was good), (2) cases the bot escalated unnecessarily (often a missing KB article), (3) cases where the bot tried and the user gave up (broken or unclear KB). Each pattern is a backlog item for the docs team. Most teams see KB completeness improve faster with the bot in place because it surfaces gaps users were silently working around before.

Via webhook. When escalation happens, SleekAI posts to your help-desk API (Zendesk, Freshdesk, Help Scout, Intercom, Front, Jira Service Desk) creating a ticket with the customer, the conversation transcript, the suggested category, and any tags. The conversation is attached as a note so the human agent reads the actual back-and-forth, not just a summary. Round-trip stays inside the support team's existing workflow.