Training an AI agent in Helpdesky does not mean fine tuning a model or uploading files to a provider. Everything the agent knows about your product comes from your help center, and everything about how it behaves comes from four things you control in the dashboard: its Instructions, the Skills it follows, the Examples of past replies it imitates, and the Suggestions Helpdesky prepares from your team's corrections. This guide covers all four, including the markdown training file that lets you edit the whole skills library in your own editor. If you have not created an agent yet, start with Set Up Your First AI Agent.
Where knowledge comes from
The agent answers from your published articles and enabled data sources. The read only Knowledge section in the agent editor shows what it has to work with, for example 12 published articles and 2 enabled data sources, and whether it uses Semantic search over article embeddings or Keyword search (embeddings not configured).
That split matters for training. When the agent gets a fact wrong, fix the article. When it gets the tone, the process or the policy wrong, fix the instructions or a skill. Trying to teach product facts through instructions produces agents that contradict your help center. The AI Question Insights page is a good place to spot which topics your articles do not cover well enough.
Instructions: the agent's standing rules
The Instructions section is the agent's voice, persona and standing rules. It is sent with every request, so keep it focused. A good set of instructions answers three questions: who is the agent, how does it write, and what should it always or never do.
The Tone presets (Friendly, Professional, Concise, Warm) insert starter text at the top of the box that you can edit freely. Sign-off controls how replies end: No sign-off, First name or Custom text.
A few rules that work well in practice:
- State the outcome you want, not the mechanism. "Ask for the order number before discussing delivery" beats a paragraph about why.
- Put limits in plain words. "Never promise a refund; explain the policy and offer to hand over" is something the model follows reliably.
- Keep it short. Anything that only applies to one kind of question belongs in a skill, not here.
Skills: instructions for one situation
A skill is a named block of instructions the agent follows when its Use when line applies. Skills live in one library per helpdesk, and any agent can pick them up, so a refund policy written once serves your Copilot and your Autopilot alike.
The Skills library on the AI Agents page. Each row shows the Use when line and how many agents use the skill.
Create a skill
Open Inbox > AI Agents, switch to the Skills tab and click New skill. The dialog asks for:
- Name, for example Refund requests.
- Use when: one line that tells the agent when this skill applies, such as "the customer asks for a refund or mentions a chargeback". The agent reads this line to decide whether the skill is relevant, so write it the way customers actually phrase the question.
- Instructions: step by step guidance, policies and phrasing the agent should follow in that situation.
- Enabled: disabled skills stay in the library but no agent uses them. Handy for seasonal policies.
Each row in the library shows the Use when line and a used by 2 agents count. The switch on the right enables or disables the skill, and the pencil opens it for editing. Deleting a skill removes it from every agent that uses it.
Give a skill to an agent
Skills are opt in per agent. In the agent editor, the Skills section lists the whole library with a checkbox per skill; tick the ones this agent should follow. You can also click Create skill right there if the library is empty or missing something.
Skills and Examples in the agent editor. The budget line under the examples shows how much prompt space they use.
When you test in the Playground, the readout under each reply lists the skills the agent used, so you can see at once whether a Use when line triggers for the questions you expect.
The markdown training file
The whole skills library can be exported as one markdown document and imported back. That makes it easy to draft skills in your own editor, keep them in version control, review changes with your team or copy a library between helpdesks.
On the Skills tab, Export markdown downloads ai-agent-skills.md. Import markdown reads a .md file back in.
The format is one ## heading per skill, two optional metadata lines, then the instructions:
# Skills for Acme
## Refund requests
- Use when: the customer asks for a refund or mentions a chargeback
- Enabled: yes
Confirm the order number and purchase date first.
Refunds are available within 30 days of delivery.
If the order is older than 30 days, offer store credit and hand over to a human if the customer insists.
## Shipping and tracking
- Use when: the customer asks where their order is or about delivery times
- Enabled: yes
Ask for the order number, then explain how to find the tracking link in the shipping email.
Things to know before you import:
- Skills are matched by name. A section whose name matches an existing skill updates it; a new name creates a new skill. Import never deletes anything, so removing a section from the file does not remove the skill from Helpdesky.
- Both metadata lines are optional. A missing Enabled line means enabled; a missing Use when line leaves it empty, which you should fill in afterwards because the agent relies on it.
- Names are capped at 80 characters, Use when at 500 and the instructions at 20,000. The file itself can be up to 512 KB.
- After a successful import the toast reads Imported 2 new and updated 3 skills. Sections that could not be read are reported as warnings while the rest is still imported.
Imported skills land in the library but are not assigned to any agent until you tick them in the editor. If you maintain the file as the source of truth, export after editing in the dashboard too, so the two do not drift apart.
Examples: replies to imitate
Instructions tell the agent what to do; examples show it. The Examples section of the agent editor holds past replies from your team that the agent should imitate. They are shown to the model as reference answers on every request, which is the most reliable way to get your team's actual tone rather than a generic support voice.
Click Add example to open the Pick a past reply dialog.
Search your team's replies, or pick from the suggested list.
- The search box finds your team's replies in recent conversations by their text.
- Below it, Suggested: highly rated or frequently reused replies from resolved conversations lists candidates Helpdesky picked for you. Each shows the customer's message, the reply, a Resolved badge and a Reused 3 times count where the same answer has been sent more than once.
- Click Add on a candidate and it appears in the editor as Example 1, Example 2 and so on, with the customer message above the reply. Remove example takes it out again.
Picking examples needs the Messages permission, because it reads real conversation text.
Examples count towards the prompt size on every request, so there is a budget: up to 10 examples and roughly 1,500 tokens in total. The line under the list, for example 2 of 10 examples · about 216 of 1,500 tokens, shows where you stand. If the examples exceed the budget, the last ones are left out of requests and the editor says so. Pick the handful that best show how your team writes, not the longest ones.
Suggestions: learning from your team's corrections
Everything above is you teaching the agent. The Suggestions tab is the agent learning from your team. Two things in the inbox count as corrections:
- Edited before sending. When a teammate changes a Copilot suggestion or a Review draft before sending it, Helpdesky keeps the agent's version and the version that was sent.
- Thumbs down. A thumbs down on a suggestion or an agent reply opens a small What was wrong? box. The reason is optional, but a short one such as "wrong refund window" or "too formal" makes the next step much better. Skip saves the rating without a reason.
Thumbs up are recorded too; they feed the Thumbs up ratio in the agent's analytics rather than the suggestions.
Helpdesky turns recent corrections into proposed changes to an agent's skills or instructions. New suggestions are prepared once a week when an agent has at least 5 new corrections, or whenever you ask. Nothing changes until you approve it.
A suggestion with its before and after text and the corrections behind it.
Generate suggestions now
Pick an agent in the dropdown and click Generate now for Bella. The note underneath says what it costs: generating uses your helpdesk's AI key, one request that reads the agent's recent corrections. The pass looks at the last 90 days, at most 40 of the most recent corrections, and returns up to 5 proposals. If there is nothing new to learn, the toast says No changes suggested this time. Below the button you can see the last run, for example Last run 2 days ago (weekly): 7 corrections, 2 suggestions.
Review a suggestion
Each card on the To review list shows:
- The kind of change: New skill, Edit skill or Edit instructions, with the agent it belongs to.
- A short rationale written by the model.
- Before and After text, so you can see exactly what would change.
- A Based on 3 corrections link that expands the motivating corrections. For each you see the customer's message, what the Agent wrote and what the Team sent, or the thumbs down reason. Only teammates with the Messages permission see the conversation text.
Then choose:
- Approve applies the change immediately. A new skill is created, enabled and added to that agent. An edited skill or edited instructions are saved in place.
- Edit and approve opens the proposed text so you can adjust it before it is applied. Approved suggestions that you changed are marked Approved with edits in the Approved list.
- Reject moves it to the Rejected list and changes nothing.
If someone deletes a skill while a suggestion for it is waiting, the card says so and can no longer be applied. Reviewed suggestions stay in their list so you can see what was accepted over time.
A training routine that works
- Start with a few sentences of instructions and two or three skills for the questions your team answers most.
- Run the agent in Copilot for a week or two. Ask the team to edit suggestions rather than retype them and to use the thumbs.
- Open Suggestions, click Generate now, and approve or edit what makes sense.
- Add two or three Examples from the replies your team is proudest of.
- Check the agent's Analytics for Thumbs up, Suggestions used and Corrections, then repeat.
Once the corrections dry up and the ratings hold steady, you are ready to think about letting the agent send on its own. The guardrails, pacing and hand-off guide covers how to do that safely.