Categories

AI Agent

The AI Support Agent automatically replies to clients based on the company's Knowledge base and the defined communication rules.

The AI Agent:
  • answers clients' common questions;
  • uses information from the provided Knowledge base;
  • follows the defined communication style (tone of voice);
  • understands when a case needs to be handed over to a live agent.
The AI Agent doesn't fully replace operators; instead, it takes on the first contact and routine cases, reducing the load on the support team and staff costs.

The AI Agent runs on powerful language models: cloud-based DeepSeek-V3.2 (Non-thinking Mode), DeepSeek-V3.2 (Thinking Mode), GPT-5, GPT-5 mini, GPT-4.1 mini, as well as Open Source models OpenAI o4-mini and DeepSeek 70B/671B. All of them can navigate the conversation context and large Knowledge bases.

The Usedesk AI Agent uses its own universal prompt, developed by us specifically for support teams.
You can use other language models, but we recommend the ones that showed the best quality — OpenAI o4-mini and DeepSeek 70B/671B.

Only the message text is sent to the model; the client's contacts are not sent. If a client has included personal data in the message themselves, it is cleared before being sent to the model.


Contents


How the AI Agent works

The AI Agent's work is built on three elements:

1. Knowledge base

The Knowledge base is the main source of information for the AI Agent. For it to work correctly, it's important that the Knowledge base is structured, up to date, and written in simple, clear language.
The AI Agent works differently with Knowledge bases of different sizes:
  • if the Knowledge base has up to 150,000 characters, the entire Knowledge base is loaded in full — the AI Agent analyzes all of it with every client request and forms a response based on the relevant fragments;
  • if the Knowledge base has more than 150,000 characters, an orchestrator is used (a paid feature): the topic of the client's request is determined → relevant articles are selected from the Knowledge base → the AI Agent responds based only on these materials. This makes it possible to work with large Knowledge bases without losing response quality.

2. Prompt (communication rules)

The prompt defines the AI Agent's role, the style and tone of responses, acceptable wording and stop words, and behavior when information is missing.

3. Scenarios and escalations

Scenarios describe the logic of the AI Agent's behavior in non-standard situations: when to transfer the dialog to an operator, how to respond to repeated questions, and how to act when a client is dissatisfied.


How to implement the AI Agent: a checklist

If you want to see how the AI Agent works with your materials, we offer to test it for free on your Knowledge base for 3 days. The Knowledge base doesn't have to be created in Usedesk — any format will do.

Important! During the free testing of the demo AI Agent, we check only one parameter — the correctness of responses based on your Knowledge base. This way you can make sure the AI Agent answers frequent questions correctly, relying on the information provided.

At this stage, we don't configure tone, style, or complex behavior logic. But as part of the free testing, we fix small inaccuracies in facts or wording once.


Stage 1. Preparing the Knowledge base

What kind of Knowledge base is best suited for training the AI?

Example — https://docs.usedesk.com/

A single source of knowledge. It's better to store materials centrally, for example, in a Knowledge base on our platform or in a separate wiki/Confluence/Notion.

Format: Q&A or articles. Q&A (question-answer) is ideal for frequent and short questions: pricing plans, delivery, warranty. Articles/guides — for complex questions, for example, step-by-step instructions or policies.

Knowledge base structure: categories by topic, for example: “Pricing plans and payment”, “Technical questions”, “Warranty/returns”, “Integrations”.

Article structure:

  • a clear title — for example, the question worded the way a client would put it;
  • a brief answer/explanation at the beginning, 2–3 sentences;
  • an extended answer — a detailed instruction;
  • tags/keywords at the end for AI search;
  • simple texts without bureaucratic jargon;
  • consistent wording, one article = one scenario, not “everything about everything”;
  • contextual examples and cases: how clients actually phrase questions and what they might use the article for.

How to prepare the Knowledge base?

1. Identify the most frequent client questions:

  • choose 3–5 main topics, for example, “Payment”, “Delivery”, “Login”, “Returns”;
  • you don't have to gather the entire Knowledge base at once — we proceed in stages, and a portion of the topics is enough for the demo setup.

2. Write articles for these topics.

3. Keep the size limit in mind:

  • at the first stage, the Knowledge base size is no more than 150,000 characters, ≈ 100–120 FAQs or 60–80 Word pages;
  • if the base is larger, we split it into parts and work in stages.

4. Assign someone responsible for the Knowledge base within the company — they will keep the articles up to date and coordinate with us.


Stage 2. Demo environment

1. Send us the prepared part of the Knowledge base in any convenient format: a table, documents, Wiki/Confluence, or a link to the Knowledge base in Usedesk.

2. We'll create a demo environment: we'll connect your Knowledge base to a test AI Agent and give you access for review. The demo environment is provided in Telegram. Requesting contact details, the welcome message, the “Show an example” button, and the rating are attributes of the demo environment only; in the real AI Agent we remove or edit them.

3. You check the AI Agent's work against the Knowledge base:

  • Ask real questions, the way your clients do. The minimum number of questions for testing is 20.
  • Ask questions strictly about the part of the Knowledge base you provided. If there's no answer in the Knowledge base, the AI Agent will make things up.
  • When the output is correct, mark the response with a heart — we'll see all the reactions.
  • When the output is incorrect, mark the response with a thumbs down and write a detailed comment about what the correct answer would be in this case and why. If the answer actually was in the Knowledge base, provide a link to the section and a quote. We'll collect all the comments and make changes to the AI Agent's work in the next round.
  • If the output is incorrect for a question that is in the Knowledge base, you need to run a frequency test — check the answers with at least ten requests. This is necessary because in 3% of cases the model may hallucinate and give a wrong answer — and this can happen on the very first answer. Keep in mind that one wrong answer is not a guarantee that the AI Agent is broken.
  • To save time, you can record comments about recurring wrong answers as voice messages — we'll transcribe the recordings, process them, and add them to the Knowledge base or the prompt.
  • During the test, note right away in which situations and under what conditions an operator should be called, so that this information can later be added to the list of rules for calling an operator.

4. We refine the AI Agent. As part of the free demo, we fix small inaccuracies once for free — we adjust wording and correct errors.


Stage 3. Test launch of the AI Agent after payment

The paid launch (100,000 rubles) includes:

1. Configuring the tone, style, and structure of responses.
You'll need to send a description of your Tone of voice. We configure the wording, communication style, and structure of responses to match your service standards. At this stage, the AI Agent starts to sound like your employee — with the right tone, level of empathy, and response style.
2. Configuring behavior in non-standard situations + escalations.
You'll need to send a description of the scenarios: what the AI Agent does in situations where the answer logic goes beyond the Knowledge base + scenarios for transferring to an operator.
Next, we fine-tune the AI Agent's behavior logic beyond the Knowledge base:
  • when to call an operator;
  • how to respond to a dissatisfied client;
  • how to respond to non-standard requests;
  • how to behave when there's no information;
  • which scenarios to handle on its own and which to hand over to agents.
3. Connecting to Usedesk and checking the logic:
  • checking all current Usedesk triggers;
  • adjusting conflicting triggers;
  • creating new triggers for operators and the AI Agent to work together;
  • configuring routing, statuses, and escalations;
  • checking the AI Agent's work on real traffic;
  • fixing technical details.

At the start, if you wish, we can launch the AI Agent in the test mode “Advanced suggestions” — the AI Agent will work only within your team: offering ready-made responses to agents without sending them to clients.


Stage 4. Launching the AI Agent

1. We turn on the “first contact” mode: the AI Agent responds to the client directly, and an agent steps in only in complex cases.

2. We fine-tune automation in Usedesk:

  • we eliminate conflicts with auto-replies, SLA, and auto-closing;
  • we automatically mark tickets involving the AI Agent with separate tags.

3. You monitor the results: you measure the % of questions resolved by the AI Agent, the % of escalations, and customer satisfaction (CSAT).

4. During the first two weeks, we actively make edits to the AI Agent.


Stage 5. Stabilization

1. Expanding the Knowledge base — we gradually add new topics from you (each iteration over 150,000 characters is paid).

2. Onboarding your team:

  • agents understand how to adjust the AI Agent's work and give feedback;
  • you get a process for continuously improving the Knowledge base.

3. Regular improvements:

  • you analyze errors weekly and edit articles,
  • once a month, optimize triggers in Usedesk.


Possible conflicts between Usedesk automations and the AI Agent

1. Auto-replies to incoming messages

  • What Usedesk does: sends the client a message “Thank you, we've received your request…”.
  • Problem: if the AI Agent has already replied, the client may receive two messages in a row (the AI Agent + the auto-reply).
  • Solution: disable auto-replies for the channels/queues where the AI Agent responds, or set up the condition “do not send if the bot_answered tag is present”.


2. Auto-routing to departments/agents

  • What Usedesk does: immediately assigns the ticket to a specific agent or group.
  • Problem: the AI Agent should respond in the ticket first, but the message goes to an agent. It turns into a “tug of war”.
  • Solution:
    • give priority to the trigger “first message → assign to the AI Agent”,
    • or use a separate bot_to_agent tag and build routing on it.

3. SLA timers and escalations

  • What Usedesk does: if an agent hasn't replied within X minutes → forwards the ticket to someone else.
  • Problem: the AI Agent has replied, but the system still thinks “the agent hasn't replied” → the escalation may happen prematurely.
  • Solution: specify in the triggers: “if there was a response from the AI Agent, do not count it as an SLA violation”.


4. Auto-closing tickets

  • What Usedesk does: if the client hasn't replied within 48 hours, the ticket is closed.
  • Problem: the AI Agent may close the dialog prematurely, or, conversely, the ticket may close even though the AI Agent promised that an agent would get in touch.
  • Solution: keep a separate process for tickets involving the AI Agent: only an agent closes them (at first), or specify a clear rule for when a ticket can be closed if it contained a dialog with the AI Agent without a transfer to an agent.


5. Greeting/closing message macros

  • What Usedesk does: automatically adds “Hello, NAME” or “Have a nice day”.
  • Problem: the client may get a duplicate — the AI Agent said hello + the system said hello again.
  • Solution: either remove the macros for the AI Agent's responses, or set up the AI Agent to respond in the right style from the start.


6. Keyword triggers

  • What Usedesk does: if the client writes “cancel”, “return” → the system assigns it to a specific agent.
  • Problem: the AI Agent may also try to answer this question → the ticket ends up in the wrong place.
  • Solution: give priority to keywords: the AI Agent tries to respond first, but if the “complex topic” label is present → straight to an agent.


7. Newsletters or mass notifications

  • What Usedesk does: sends templated messages when statuses are updated, during promotions, etc.
  • Problem: the AI Agent may send its response, and the system may “cover” it on top with another template.
  • Solution: set up exceptions for the AI Agent's tags.


The main risk is duplication or a conflict in response priority. To solve this:

  1. Create a separate tag for the AI Agent (bot_answered).
  2. Set up a Usedesk trigger that checks this tag and prevents the conflict.
  3. Decide who's in charge: if the AI Agent has replied, then auto-replies and greetings don't kick in. Other automations shouldn't trigger on top of the AI Agent.