Using call transcripts from tickets to build the knowledge base for Customer Agent

[Hakkinen33]

Hi everyone,

our client records support calls on tickets, with HubSpot generating transcripts and AI call summaries. The goal is to turn recurring issues and their resolutions into knowledge base content, so that Customer Agent can autonomously answer the most common and simple requests.

Has anyone done something similar?

1. Can call transcripts or summaries be used directly as a source for Customer Agent, or do they need to become knowledge base articles first?
2. How do you extract and classify recurring issues from calls (reports, exports, AI tools, ticket properties)?
3. Any tips on handling personal data in recordings when building the knowledge base?

Thanks!

[Berangere]

Hey @Hakkinen33 and Happy Wednesday!

These are great questions, thanks for asking the HubSpot Community!

To start with, I’d suggest checking these resources that can help:

• Understand the customer agent
• Review call recordings and transcripts
• Generate customer agent knowledge from customer interactions

And I’d love to put you in touch with our Top Experts: Hi @danmoyle, @CHawkins and @EugeneAtKiwi do you have suggestions to help @Hakkinen33, please?

Have a lovely day and thanks so much!

Bérangère

[Jignesh-at-Klarion]

I’d be careful about treating recurring call content as knowledge that should automatically flow into the Customer Agent.

Adding knowledge to a Customer Agent is really an expansion of what you are asking the agent to handle autonomously. That deserves a more deliberate process than simply approving a generated answer.

When we work with clients, we generally start top-down: continuously analyze why customers are contacting support, then decide:

What should the Agent handle?
What should be escalated to a human?
What customer friction should be fixed upstream instead?

That last category is especially important. A recurring support question does not necessarily mean “create an answer for the AI.” Sometimes doing that just masks a frustrating product, process, or customer-journey problem that should be fixed.

Once you decide that a use case belongs with the Customer Agent, then you can build the appropriate knowledge, policies, customer context and actions around it. I’d also test it end-to-end against historical tickets and edge cases before putting it in front of customers.

HubSpot’s Generate Knowledge beta looks useful to me as a way to surface recurring questions and draft potential knowledge. But I’d treat those as inputs into this process, rather than as knowledge that an admin should simply approve and publish.

For anything that materially expands what the agent can resolve, I’d also want the relevant business owners involved. A seemingly simple answer can depend on product rules, customer tier, commercial policy, exceptions, escalation logic, or downstream processes.

Jignesh Shah

[MahdiEqbal]

I wouldn’t feed the raw transcripts straight into the Customer Agent.

I’d use them more as source material: extract the issue, what actually solved it, and maybe product or version context, then group similar cases together

Once the same issue shows up enough times, turn that into a reviewed knowledge article the agent can use

That review step is pretty important IMO. Calls can contain customer-specific details, temporary workarounds, or even troubleshooting that turned out to be wrong later

For PII I’d also remove it before the AI/classification step if possible, rather than trying to clean it only when publishing the final article.

So roughly: calls then extract or classify then group recurring issues then human review then approved knowledge

[SBhat3]

Hi @Hakkinen33 ,

You’ve already got the right idea here. The main thing I’d change is the workflow: I wouldn’t use the raw call transcripts as the Customer Agent’s knowledge source.

HubSpot’s current Customer Agent setup supports knowledge bases, website pages, URLs, files, and short answers as content sources. I don’t see call transcripts or call summaries listed as a direct Customer Agent source.

There is actually a pretty useful newer option for this use case, though. HubSpot now has a Customer Agent knowledge-from-customer-interactions beta that can review resolved tickets, group similar questions, and generate FAQ-style Q&A pairs for you to review. There’s also a Knowledge Base Agent beta that can analyze closed tickets and generate draft knowledge base articles from recurring knowledge gaps.

So I’d approach it like this:

Call → transcript/summary → ticket → identify recurring issue → review the resolution → turn the validated answer into an FAQ/KB article → use that as Customer Agent knowledge.

The important part is the human review between the transcript and the final knowledge source. A transcript can contain customer-specific details, guesses from the rep, outdated workarounds, or information that only applied to that particular account. I wouldn’t feed those straight into an autonomous agent.

For identifying the recurring issues, I’d start with the ticket data rather than trying to manually read hundreds of transcripts. Look for repeated ticket subjects/categories, resolution patterns, and the actual questions being asked. If the new customer-interaction knowledge feature is available in the portal, it can do some of that clustering for you and produce Q&A drafts that you can review.

And definitely be careful with PII. HubSpot specifically warns that content used by Customer Agent can appear in responses, including content from private sources, so personal, confidential, sensitive, or customer-specific information shouldn’t be put into the agent’s knowledge sources.

For example, instead of turning:

“John from ABC Ltd couldn’t reset his password because his old email address was still attached to the account…”

into a knowledge article, I’d extract the general solution:

How do I reset my password if I no longer have access to my original email address?
Follow these steps…

That gives the Customer Agent something reusable without exposing anything about John or ABC Ltd.

So yes, this is definitely a workable approach, and HubSpot has actually added some features recently that make the workflow much easier. I’d just treat the call transcript as the raw material, not the final source of truth. The KB/FAQ should contain the cleaned-up, validated resolution that you actually want the Customer Agent to give customers.

[danmoyle]

Hi @Hakkinen33! Short answer: yes, this is a use case HubSpot has built toward, and there are native paths for it. Some of the features we’re talking about here might still be in some level of beta.

The key distinction I want to call out is that call transcripts are useful for finding patterns, but they aren’t a standard Customer Agent source on their own. Customer Agent accepts synced KB articles, web and landing pages, uploaded files, public URLs, and short answers you write yourself. So the work is turning what happens on the calls into reviewed, customer-safe content before the agent relies on it.

Native options worth knowing about in my opinion and in my experience:

Customer Agent knowledge generation (beta). It analyzes tickets from the last 90 days, or a segment you choose, clusters similar questions, and drafts FAQ-style Q&A pairs. You accept, edit, or decline each one.

Knowledge Base Agent (beta). It scans closed tickets, finds gaps, and proposes article drafts. It needs Service Hub Professional or Enterprise, an existing KB, at least five closed tickets, customer conversation data enabled, and credits.

Conversation Intelligence. It gives you AI call summaries, including a Support Rep summary built around the issue, root cause, actions taken, and outcome. Transcript search is there too, and tracked terms are available on Enterprise.

Data Agent and smart properties. These can read call transcripts and write structured values into ticket properties.

Classifying recurring issues: I wouldn’t let a free-text summary become your taxonomy. Build a small set of dropdown properties on the ticket, like issue family, product area, customer intent, root cause, resolution type, and self-service eligibility. Let the transcript help fill them in, and keep a human reviewing. Then report on what repeats most, what’s suitable for deflection, and which categories have no linked article.

For personal data, this is the part I’d slow down on. HubSpot says that even private content added to Customer Agent can influence its public responses. Treat your calls as research, not as published knowledge. Strip names, account details, invoice numbers, and anything internal, and write the answer as a neutral pattern and add a “do not reuse” flag for calls involving payment data, security issues, legal matters, or another customer’s information. Also review your admin settings for customer conversation data, and confirm your recording consent and retention rules separately.

Here’s how I’d start. First I’d pick one high-volume, low-risk topic, like login help or standard how-tos. Then I’d build a segment of resolved, self-service-eligible tickets from the last 90 days. Review every draft before accepting it. And finally, I’d test the agent in Help Desk reply recommendations before it faces customers, and measure for 30 days before expanding.

Hope that helps!