AI Knowledge Builder

You already wrote the answer.
Once, to one customer, in 2024.

The Knowledge Builder reads tickets your team has already solved and writes them up as articles your agents can search — so the second person to ask gets the answer in seconds.

Kostenlose Testphase starten Read the docs

Solved ticket Closed · Mar 2024

#18402 · Order stuck in transit for 5 days

Resolved by Priya · 4 replies · 2h 40m

Customer

Tracking hasn’t moved since Tuesday. It just says “in transit”. This was meant to arrive Thursday.

Priya · internal note

Checked the carrier portal — last scan is the origin depot, no handover scan. Looks like the depot missed it. Reissuing rather than waiting on a trace.

Priya · to customer

The parcel was missed at the depot, so I’ve sent a replacement out today on express and refunded the delivery fee. Tracking follows within the hour.

Solved well. Then archived, where nobody could find it again.

Generated article Editable

How do I check where my order is when tracking has not moved?

What is happening

The parcel has an origin scan but no handover scan, which usually means the depot missed it rather than the parcel being lost in transit.

How to resolve it

Reissue on express rather than opening a carrier trace — a trace takes 5 to 7 days and usually ends in a reissue anyway.

What the customer is owed

Refund the delivery fee when the depot is at fault. The item itself is replaced, not refunded, unless the customer asks.

chunked & embedded editable merged from 4 tickets

One ticket in, one article out. Point it at a quarter of solved tickets and you have a knowledge base written entirely out of work your team has already done.

Your best documentation is sitting in your archive.

A ticket about an order stuck in transit gets solved once, properly, by someone who knew to check the depot scan. Then it closes, and that knowledge leaves with the conversation. The next customer with the same problem joins the same queue and gets the same investigation from scratch.

Nobody has time to write that up as an article. So it never gets written, and the same two hours get spent again next month. The Knowledge Builder does the writing part, at the scale of your whole archive.

Order stuck in transit for 5 days

Mar 24

Discount code not applying at checkout

Apr 24

Duplicate charge after retry

Jun 24

Warranty claim outside 30 days

Aug 24

4,812 solved · 0 searchable

The obvious objection

Won’t this produce four thousand near-identical articles?

No, because articles are compared before they are saved. Every generated article is embedded and checked against what already exists on that source. If it is semantically close to one that is already there, it merges into it instead of creating another entry.

So forty tickets about the same depot problem become one article that gets better each time, not forty near-duplicates that make the base worse to search than the archive it replaced.

#18402 Order stuck in transit for 5 days seed

#19117 Tracking not updating since Monday 0.91

#19640 Parcel hasn’t moved, still at depot 0.89

#20233 No delivery scan for a week 0.87

One article

Order shows “in transit” with no tracking movement

Merged from 4 tickets · similarity above 0.85

What it actually produces.

A source like any other, sitting in your Sources list with an AI Knowledge Builder tag — and every article titled as the question a customer would ask, because that is how it gets found.

August Solved Tickets Knowledge Base

Built from Zendesk tickets

AI Knowledge Builder 42 articles

How do I check where my order is? Editable

Why doesn’t my webhook fire after a workflow runs? Editable

How do I handle a refund request outside the window? Editable

How much does it cost to ship a large item? Editable

What does the 30-day warranty cover? Editable

Why are workflows delayed after a bulk import? Editable

Two ways to write each article.

Most tickets need a straight write-up. Some need someone to go and look at the order first — so you can choose per run, and pay accordingly.

Prompt mode

Default

One model call per record. Fast, cheap and predictable — the right choice for the bulk of an archive, where the ticket already contains the answer.

One call per ticket, so cost scales linearly and predictably
Best when the resolution is explained in the thread itself
Fastest way to turn a large backlog into a first draft base

Agent mode

Richer, slower

One of your own agents investigates each record before writing — searching knowledge, pulling the order, checking the account — then writes the article from what it found.

Uses the agent’s read tools to fill in context the ticket left out
Noticeably better on tickets where the fix took real investigation
Write tools are stripped, so an article run can never change a ticket

What a batch does to each record

Five steps, in order, per ticket. The first two exist to stop you paying for records that were never going to produce anything.

Already done?

Tickets that produced an article on a previous batch are skipped, so re-running is safe.

Worth it?

A relevance filter drops auto-replies, spam and thin threads before generation. On by default, and editable.

Write

Prompt mode calls the model directly; agent mode investigates first, then structures the result.

Embed

Title and body are vectorised so the article is retrievable by meaning, not keywords.

Merge or add

Close to an existing article, it merges. Genuinely new, it becomes its own.

Four features that sound alike. They are not.

The Knowledge Builder writes articles out of records you have already resolved, so the answer is retrievable next time.

Studies — AI Analysis

Run one analysis across thousands of records and get a structured grid back.

Agent Simulations

Run your agent on real tickets with every write held back.

Agent Evaluation

Grade the conversations that already happened, against the agent’s own instructions.

Questions about the Knowledge Builder.

Short answers. If something is missing, ask us — we will tell you straight.

The Knowledge Base is where knowledge goes in — you upload documents, connect sources, crawl a site. The Knowledge Builder produces that knowledge out of records you already have. Its output is an ordinary knowledge source, so the two are complements, not alternatives.

Articles are written from the record in front of it, and in agent mode from what its read tools actually returned. Every article stays editable, and we would expect you to review a sample of the first batch before trusting the rest.

Not by default — the scope filter starts at solved tickets, because an unresolved thread has no answer worth publishing. You can change the query if your workflow differs.

They are read, because that is usually where the real diagnosis lives, and the article is written from the resolution rather than quoted from the thread. Choose which fields are included per run.

Yes. Every article is editable in a rich text surface, and the whole set is browsable from the sources list with an AI Knowledge Builder tag so you can tell generated articles from uploaded ones.

Bereit, dein Team mit KI zu turbo-boosten?

Starte in Minuten. Verbinde deine Tools, konfiguriere deine Agenten und lass die KI den Rest erledigen.

Demo ansehen Kostenlose Testphase starten

500 Freikredite · ohne Zeitlimit, ohne Kreditkarte

You already wrote the answer. Once, to one customer, in 2024.