AI help center that solves it before the ticket
Articles in several languages, a search that understands what the customer means, an assistant that only answers from your documentation, and order lookup. Whatever isn't solved there reaches AgentDesk already classified, and every search with no results tells you which article is missing.
If customers can't find the answer, they open a ticket
When the help center lives inside the helpdesk, it often mixes customer-facing articles with the team's internal templates. That's why we don't copy it as is: internal material moves into AgentDesk as canned replies, and public content is reviewed and completed with copy from your website and official policies before it goes into the new help center.
A search engine that only matches words falls short, because customers don't write the way articles do. So when a keyword search finds little, it's topped up with articles that are close in meaning, and searches that still return nothing are logged so the missing article gets written.
What we want from a help center is that questions with a published answer never reach the queue, and that the ones that do arrive with the order number, photos and category already filled in.
What's inside the help center
Four parts that share the same knowledge base and talk to AgentDesk.
Knowledge base with provenance
Every article records where it came from (the previous helpdesk, the website, official policy or written in-house) and the date a person last confirmed it's still accurate.
- Organised with the same categories the team uses
- Accuracy review every six months by default
- Drafts aren't published until someone reviews them
Keyword and semantic search
It searches full text first, ignoring accents. If that finds little, it adds semantic results from embeddings stored in PostgreSQL with pgvector. If the AI service is down, keyword search keeps working.
- Related articles from the same category
- Searches with no results are logged so the missing article gets written
An assistant that answers from your articles
It retrieves the articles closest to the question and answers only with what they say. If the information isn't there, it says so and offers to open a request. If the customer asks for a person, it opens one straight away.
- It doesn't write links on its own, so it can't invent URLs
- It asks when a question could mean two things
- Message limit per conversation and a daily spending cap
Order lookup and requests
The customer enters the order number and the email used to buy. If they match, they see the status, the carrier and the tracking link, and if the order is still being prepared they can cancel it themselves. For anything else there's a guided request form.
- After three failed attempts nothing about the order is shown
- Up to 12 photos or PDFs per request
- If the customer already has an open request, they're offered to add to it instead
- reCAPTCHA on the form
How we build the knowledge base
The portal itself takes us about two weeks. The content is what takes time, and there are no shortcuts there.
Inventory
We export what you already have in your helpdesk and separate public content from internal material. Often a good part of the articles are team templates rather than text written for customers.
Gaps, backed by data
We match the categories of real tickets against the articles that exist. Views and votes carried over from the old system help decide what to rewrite first.
Writing and translation
We fill in with copy from your website and your official policies. Missing translations are prepared by the AI and stay as drafts until a person reviews them.
Connection to AgentDesk
The questions for each request type are defined in AgentDesk, so the form asks for exactly what the team needs. Order lookups also go through AgentDesk: the help center never calls the WMS directly.
Maintenance
Searches with no results, thumbs-down votes and questions the assistant couldn't answer are grouped by similarity. That list tells you which article to write or fix, but a person makes the call.
From the help center to the team's inbox
When the customer can't find the answer, the request reaches AgentDesk with what's needed to handle it.
Search or ask
Articles, assistant or order lookup
Guided request
Reason, order details and photos
Lands in AgentDesk
With category and complete data
Gets resolved
By the AI in active categories, by a person otherwise
What the assistant does and what it's not allowed to do
It does
- Answer with what the retrieved articles say
- Ask when a question could mean two things
- Suggest a request if nothing is solved by the third message
- Open the request as soon as the customer asks for a person
It doesn't
- Make up policies, deadlines or prices that aren't in an article
- Write links
- Create tickets on its own: the customer confirms the request
- Show order details if the number and email don't match
Before launch we tested it against attempts to bypass its instructions and to look up other people's orders.
What we check before going live
Order lookups without intermittent errors
Order and request lookups depend on the server's network. We test them from the production server before opening, so a network issue doesn't leave requests failing on day one.
AI reachable from production
If the AI service has network restrictions and doesn't allow the new server, neither the assistant nor semantic search works. We check this before signing off the launch.
The same request, several times
Customers often open the same request more than once. If the email already has one open, the form says so and offers to add to it.
Keeping the old help center's rankings
The old help center URLs redirect with a 301 to the new ones. Each language has its own sitemap, hreflang tags and a prerendered version for search engines.
Questions about the AI help center
What's the difference between an AI help center and a chatbot?
A generic chatbot answers with whatever the model knows. This assistant searches your articles first and only answers with what it finds; if the answer isn't there, it offers to open a request. The help center is also a website with indexable articles, search and order lookup, so it works for people who don't want to chat.
Can the assistant make up an answer?
It has explicit instructions not to: if the information isn't in the articles it retrieves, it has to say so and offer a request, and it doesn't write links. No model is infallible, which is why sensitive policies (returns, deadlines, refunds) should live in clear, reviewed articles, and customers' negative votes get reviewed.
Can we migrate the help center from our current helpdesk?
Articles can be exported, usually through the API. Our advice is not to copy them as they are: separate public content from internal templates, rewrite the worst-rated articles, and 301-redirect the old URLs so you don't lose rankings.
Does it work in several languages? Who translates the articles?
Yes, it's multilingual and customers see the help center in their own language. Articles you already have translated are imported with their translations. Missing languages are translated by the AI and kept as drafts until a person reviews them before publishing.
Is it safe for customers to check an order without logging in?
To see an order you need both the order number and the email used for the purchase, and both must match the WMS. After three failed attempts nothing is shown and no request about that order is opened, so nobody can go through order numbers by trial and error.
How long does it take to set up a help center like this?
The portal takes us about two weeks. The content depends on what you already have: how many articles need reviewing or writing, in how many languages, and whether it runs alongside the AgentDesk rollout. That's the part that sets the calendar.
What usually goes with a help center
See AI help center in action
Tell us which tool you use today and how many languages you support, and we'll walk through it with your cases in mind.