Grounded in your knowledge base and resolved tickets, it answers routine questions with citations, drafts replies for agents, and escalates with context when it is unsure.
If a large share of your tickets are "how do I" questions with a documented answer, a copilot pays for itself in agent time. If most tickets need account access or judgement, start with agent assist and keep humans in the loop.
The copilot sits on top of the same retrieval layer as our enterprise RAG builds. What makes it a support product is the confidence gate, the hand-off with context, and the feedback loop that turns every edited draft into an improvement. We run exactly this pattern inside Eventify, our own event platform, where an AI support bot handles organiser and attendee questions.
The confidence gate is not one number. It combines the retrieval score of the best passage, the agreement between the top passages, a topic classifier for sensitive categories, and the customer's account state where the integration allows it. The thresholds are set from the pilot data and retuned as the evaluation set grows. The hand-off always includes the draft and its sources, so the agent gains time even when the system was not sure.
The copilot lives inside the helpdesk. For each new ticket it shows a drafted reply, the articles it used and a confidence indicator. The agent edits and sends. Every edit is logged.
In chat, the help centre or the ticket form, the copilot answers directly when the gate passes. It always offers a route to a person, and it never pretends to be one.
Both modes share the conversation design and guardrails we use in AI chatbots, including handling of abusive input, repeated questions and attempts to extract other customers' data. The prompt injection defence page covers the controls in detail.
Help-centre articles, internal macros, product docs and anonymised resolved tickets, each tagged with its source type so official content wins ties. Stale and duplicate articles are flagged during the audit.
Replies in your tone, with the articles used shown alongside. Variables such as plan names and limits are pulled from the account record, not guessed from the article.
Threshold logic, sensitive-topic rules and hand-off payloads that drop into the helpdesk's existing assignment and SLA workflows.
Deflection rate, draft acceptance and edit distance, escalations by reason, and a ranked list of questions with no good source. This is the list your content team works from each week.
Real tickets with approved answers, scored for correctness, tone and policy compliance on every prompt or model change. The method follows our guide to evaluating LLM outputs.
One to two weeks. We cluster recent tickets by intent, map each cluster to the articles that should answer it, and find the gaps. Output: the first-release scope and a content fix list.
Helpdesk connection, knowledge ingestion with source tagging, anonymisation of ticket history. Retrieval tuned on the audit's question set.
A pilot team of agents uses drafts for two to four weeks. We track acceptance and edits daily and fix the top failure modes each week.
Topics whose scores and edit rates pass the agreed bar are opened to customers. Sensitive topics stay human by rule. Thresholds are reviewed monthly.
Weekly gap list to the content team, monthly evaluation review, and a retainer if you want us to keep tuning. Monitoring follows our LLMOps practice.
Deflect, assist and escalate: the full solution including voice and triage.
ReadHow we run a support copilot inside our own event platform.
ReadGet help-centre content ready before the copilot reads it.
ExploreConversation design, channels and guardrails for customer-facing bots.
ReadAgent assist, almost always. It uses the same retrieval and the same knowledge base, but a human reviews every answer, so the risk is low and the feedback is rich. Once edit rates fall and the question set scores well, the same system can be opened to customers for the topics it handles best.
It does not guess. Below the agreed threshold, or on sensitive topics such as billing disputes and cancellations, the conversation goes to a human with the draft answer and the sources attached. The agent sees why the system was unsure, which also tells your knowledge team what is missing.
Zendesk, Salesforce Service Cloud, Freshdesk, Intercom, HubSpot and ServiceNow are the common ones. The copilot reads knowledge articles, macros and resolved tickets, and writes drafts back into the agent's workspace. See customer support AI for the wider solution.
Yes, with care. Resolved tickets are the richest source of real answers, but they contain personal data and outdated advice. We anonymise, deduplicate and date-filter them, and keep them separate from official articles so the copilot can prefer the official source.
Yes. Our event platform Eventify includes an AI support and knowledge bot for organisers and attendees, built on the same retrieval approach described here. The case study describes what we learned running it.
Agent assist is usually live in 6–10 weeks, including knowledge audit, helpdesk integration and a pilot team. Customer-facing deflection follows once the evaluation set and the pilot edit rates support it, often another 4–6 weeks. Talk to us about your queue.
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.