Skip to main content
Support AI responds to customers based on an approved knowledge base, and it operates in one of three modes, and the path to autonomy is deliberately step-by-step: first, prompts under operator control, then self-response within defined boundaries. The mode is switched in the “AI Support” section, hints and autonomy cannot be enabled without a working role model “Respondent” Autonomous mode is additionally locked with readiness checks. Classification of appeals and mood assessment are included by separate switches and do not depend on the dialogue mode: mood is assessed even when the AI is turned off. analytics.

How AI Is Building a Response

The first message opens a waiting window (8 seconds by default), each subsequent message extends it, short messages and pictures without a signature have a longer window (20 seconds), a total ceiling of 120 seconds, so that the person typing a thought into five messages receives one coherent response, not five scraps. And then you look for validated lessons along the way, and you go through typos and you write “pn” instead of “VPN.” If you don’t know, you go to the operator. If you have the knowledge, the Respondent model forms one of three outcomes: a ready-made answer, a clarifying question, or a recommendation to pass it to a person, each with its own confidence. The client’s screenshots are recognized by a separate role model, Visual Analysis, where the defendant receives a textual description marked as recognition, not fact, up to four images per move, and a simple “hello” responds without calling the model. The client’s text remains data: it’s enclosed in the model request, and the instructions inside it are not to be executed by the model, and the operators’ internal notes don’t fit into the context of the model.

Mode of clues

The hint lives separately from the correspondence and cannot physically leave the client until the operator hits “Send.” The hint card shows the confidence of the model and the sources: the lessons on which the answer is built. Check them before sending. The operator can reject a clue with a reason: wrong fact, outdated, insecure, wrong tone, model didn’t understand the question. Feedback accumulates in analytics, but knowledge itself does not change. If the operator answered in his own words, the clue is marked unused. When the client writes again, while the operator was thinking, the clue becomes outdated: sending a response in an outdated context will not work, a new hint will come from the updated dialogue. Obsolescence is not considered an AI error and is accounted for separately in the metrics.

Self-contained regime

The autonomous mode itself does not allow anything: the AI can respond without an operator only within the autonomous regions that are turned on; the region sets the rules of compliance (languages, keywords, channels) and the confidence threshold (0.8 by default); the answer below the threshold or outside the areas goes to the person. While the AI prepares a response, the ticket is marked with the state “AI processes” so that the queue does not look abandoned. Before sending, the admissibility is checked again: if during this time the operator intercepted the ticket, changed mode or the client wrote again, the auto-answer is canceled.

Readiness checks

The first activation of autonomy requires that you pass readiness checks. Block the inclusion: active approved knowledge, lack of knowledge in quarantine, the Respondent working model, configured human transmission rules and at least one threshold area included. Control scenarios and their runs are recommended, but not blocked. Each check in the readiness card is clickable and leads to the correction site.

Control scenarios

The “AI Testing” section stores control scenarios: questioning and waiting for an AI to behave (respond, clarify, or transfer to a human). There are negative scenarios where the right behavior is failure. The “Run All” button runs scenarios on a live model in the background. The failure of a script marked critical, a signal that autonomy is early to turn on.

Transmission to the human

The transfer is triggered by a whole set of triggers: the client asked for a person, the dialogue ran into a limit of cycles or time, the client’s mood deteriorated, knowledge was not found, the question got into forbidden topics, confidence is too low. The mode of transmission is selected in the settings. Soft: the first request for a person transmits the dialogue immediately. Persistent: AI can offer another attempt, but the repeated request is performed unconditionally, and the cycle and time limits with AI act as a ceiling. Prohibited topics are transmitted to a person in any mode. The request for a person is recognized without calling the model, so it works even when the provider fails. When it’s transferred, the client gets a message about the transfer to the operator, the ticket goes back in line, and the operator is shown a resume: the problem is that the AI has already asked what the client has suggested, what knowledge was used, and the reason for the transfer. The client does not have to repeat everything again. After passing the AI on this ticket, he is silent until the operator clearly presses “Return AI.” The “Intercept from AI” button works without transmission, at any time.

Failure and degradation

The module is built on the principle of fail-closed: if the model is not configured, turned off or exhausted the limit, the ticket simply remains with the operators, without hanging states. The transfer of the mode to “off” cancels the unfinished work of the AI: prompts become obsolete, unassembled moves are thrown.
Pay attention to the “Hint Feedback” in the “Support AI” section: the percentage of rejected prompts and reasons for refusals suggest what the knowledge base is worth finishing.