How Do You Turn Past Zendesk Tickets into an AI Knowledge Base?

To turn past Zendesk tickets into an AI knowledge base in Macha, run a Study over your solved tickets with a four-column schema (question, answer, category, product_area), then push the results into a Knowledge Source your agents search. The next customer with the same question gets answered from how your team actually solved it last time. There are no CSV exports and no scripts, and the tickets never leave your connected Zendesk.

If you want finished help-center-style articles rather than Q&A rows, Macha's AI Knowledge Builder does the same job from the same wizard shape and merges near-duplicates for you. Both routes are compared below.

Why are past tickets worth turning into knowledge?

Every resolved ticket is a small piece of documentation: a real question, the context around it, and an answer that worked. Across a year that adds up to a large, tested FAQ, and it covers the messy edge cases your help center never will, in the words your customers actually use.

The problem has always been scale. Nobody reads 20,000 tickets or hand-writes an article for each one. Once a good answer is sent, it sinks to the bottom of the queue. The knowledge exists; it just can't be retrieved in that form.

Reading a large pile of text and pulling out the useful parts is work AI handles well, provided you can run it across every ticket at once.

How is this different from exporting tickets and scripting it yourself?

You can do this by hand. Export tickets to CSV, write a script that calls an LLM on each row, clean the output, then work out where to store it so something can use it. We've seen technical teams do exactly this in a coding environment, downloading tickets and running AI over them offline.

Macha does it on the platform, over your live Zendesk data, with no export and no code. You connect Zendesk once, describe what you want extracted, and Studies runs the analysis and can drop the results straight into a knowledge source your agents search. The data and the AI sit in the same place.

What is a Study?

A Study (labelled AI Analysis in the sidebar) runs an AI model over a set of records, such as your Zendesk tickets, and fills a structure you define: one row of results per record, one column per field. You can sort and filter the grid, export it to CSV, open a report view, or push it into your agents' knowledge.

The flow is a four-step wizard, Source → Scope → Extract → Review, with a push to knowledge at the end.

Step 1: which source do you pick?

Start a New Study and choose Zendesk Tickets, the source available today (Zendesk has to be connected first, under Connectors). If you have two Zendesk accounts connected, each shows up separately.

This is where the mental model clicks: a Study isn't a chat or a single prompt. It's a batch job that treats every ticket as a row and fills in the columns you define.

Step 2: which tickets does the AI read, and which fields?

This step does two jobs: it decides which tickets get processed (which also sets your cost), and what the AI may read from each one.

Date range. Defaults to the last 30 days. Widen it to cover more history, but a bigger range means more tickets and more cost, and a single run is capped at 20,000 tickets. To go bigger, split the history into several runs by date. For a knowledge base, a 6 to 12 month window of resolved tickets is a good starting size.

Filter (optional). Any Zendesk search query, for example tags:wismo status:solved priority:high. Leave it blank to include every ticket in the range. This is the most useful lever for both quality and cost: scoping to status:solved (tickets that actually have answers) and one topic at a time gives you a clean, cheap first run. The estimate you see later is calculated from this exact query.

What the AI sees per ticket. You tick which fields the AI reads: Subject, Description (first message), Status, Priority, Type, Tags, Created/Updated dates, Requester/Assignee/Group IDs, Custom fields, and the Full comment thread. Macha flags Custom fields and the Full comment thread as expensive, because each adds an API call per ticket and the thread grows the prompt. For most classification jobs Subject and Description are enough. For a knowledge base, Subject + Full comment thread is usually the right pair, because the resolution lives in the replies.

Step 3: how do you define what to extract?

You define a schema: the columns you want filled for every ticket, like a spreadsheet, plus the model and any guidance.

Model. Studies use the same model line-up as agent chat. Extraction is usually simple work, so a fast model such as GPT-5.4 Mini is often enough. Step up to a heavier model only when the columns need judgement.

Instructions (optional). Free-form guidance on how to fill the fields: tone, edge cases, what to ignore, examples. Each column can also carry its own guidance. For a knowledge base, something like: "Write each answer as a clear, self-contained solution a new customer could follow without seeing the original ticket. Ignore internal-only chatter."

Fields to extract. Add one column per piece of information, each with a Column label and an Answer type:

• Yes/No: true/false flags (refund_requested, bug_report).
• Single / Multiple choice: the AI picks from a fixed list you type in (category, product_area).
• Number: counts or scores.
• Short / Long text: free-form answers (question, answer, summary).

For turning tickets into a knowledge base, a good starting schema is:

Those four columns turn a raw ticket into a clean Q&A entry.

Step 4: how do you check the cost before running?

The Review step summarizes the source, account, scope, fields and model, and shows the number of tickets the query matches and the estimated cost before anything runs. A Study is a saved configuration, so you can re-run it later.

Two safety nets matter here:

1. Test run first. A test takes a small sample (typically 10 to 50 records) and runs the same extraction, so you see how well the AI fills your schema before committing. If the answers aren't clean, go back to Step 3, tighten the column guidance, and test again. Iterate on a handful of tickets, not on 10,000.
2. Confirmation gate. Clicking Run study opens a second confirmation that repeats the record count, the estimate and the model. During the run you see live progress, and you can cancel at any time: you only pay for records actually processed. Each run keeps a frozen snapshot of its config, so editing the Study later never rewrites history.

Step 5: how do the results reach your AI agents?

When the run finishes, push it into a Knowledge Source. Every row becomes a searchable document, indexed like any other knowledge. Link that source to an agent and it answers from your real, historical resolutions alongside your help center.

A year of resolved tickets becomes an FAQ your agents search. The next "how do I change my billing date?" gets answered from the dozen times your team already answered it. (See how to connect a knowledge base to an AI agent for the linking step.)

Should you use Studies or the AI Knowledge Builder?

The AI Knowledge Builder lives under Sources and follows the same Source → Scope wizard, but it writes finished articles instead of schema rows. Its scope defaults to the last 90 days and a status>=solved filter, a Skip low-value tickets toggle drops auto-replies, spam and thin threads, and each new article is compared with existing ones so a similar ticket updates or attaches to an article instead of creating a duplicate.

• Use a Study when you want control over the exact columns, or want the same run to feed analysis as well as knowledge.
• Use the Knowledge Builder when the goal is only a clean, deduplicated knowledge base.

What else can a Study pull out of old tickets?

The same wizard extracts anything you can describe. Change the schema and you get analysis of a year of support data:

• Sentiment and CSAT drivers: add sentiment and frustration_reason columns and see what's making customers unhappy.
• True contact reasons: a category column shows why people really write in, which often differs from your existing tags.
• Risk and revenue signals: flag refund_requested, bug_report, cancellation_risk or upsell_opportunity across every ticket.
• Knowledge gaps: a column for "was this answerable from existing docs?" tells you what to write next.

The report view charts every Yes/No, choice and number column, so an admin can read the pattern without a BI project.

Which Macha plans include Studies?

Macha sells one plan priced by monthly ticket volume, starting at $299/mo for up to 750 tickets, and Studies is included on every tier. Studies also run on the $50 free trial, so you can connect Zendesk, build agents and run your first Study before you pay.

What are the best practices for a first run?

• Start narrow. One topic or product area, status:solved, a few months. Get the schema right, then widen.
• Pick the model that fills the schema well. A fast model handles simple extraction; a heavier one earns its place only when the columns need judgement.
• Iterate with the test run. Don't run the full set until a sample looks clean. Inconsistent answers are usually fixed in the column guidance, not by changing the model.
• Choose fields deliberately. The full comment thread gives the richest context for answers; trim fields you don't need.
• Refresh quarterly. Re-run on recent tickets so the knowledge stays current. Each run is a snapshot, so you keep the history.

Frequently asked questions

What is Macha Studies? A feature that runs an AI model over a set of records (like Zendesk tickets) and extracts a structured result for each, which you can analyze, export to CSV, or push into your agents' knowledge base.

Do I need to export tickets or write code? No. Studies run on the platform over your connected Zendesk data, with no CSV exports and no scripts.

How much does a run cost? The Review screen shows the matched record count and the estimated cost before you confirm, and you only pay for records actually processed. Studies are included on every Macha plan, which starts at $299/mo, and on the free trial.

How many tickets can I analyze at once? Up to 20,000 records per run, processed in parallel with live progress. Split larger histories into several runs by date range.

What should the AI read from each ticket? Pick fields in Step 2. For a knowledge base, Subject + the Full comment thread captures the question and the resolution; trim the rest to save cost.

Can my agents actually use the results? Yes. Push a completed run into a Knowledge Source and any linked agent can search it, the same way it searches your help center.

Can I run a Study on the free trial? Yes. The $50 trial covers agents, tickets and Studies; a run draws on that usage.

Put your past tickets to work: connect Zendesk, build your first agent on $50 of free usage with no credit card, and run a Study on your solved tickets once you're on a paid plan. Start free.