How We Built a Voice-First Blog Agent at Kontactic

A few weeks ago we built a small internal tool: a voice agent that calls the founder twice a day, lets him talk for a few minutes about whatever happened that day in the business, and then turns the recording into a draft blog post. It is not a product. It is a piece of internal tooling. But the reason we built it is worth writing down, because it touches on a problem most operator-led companies have — the best material for your blog is locked inside your operators, and they will never sit down and write it.

This is a working note about that problem and how we are trying to solve it.

The blank-page problem for operators

The people doing the actual work — clearing customs, fighting with a Coupang category manager, filing a KC certification revision, explaining VAT to a confused founder in California — almost never write blog posts. Not because they cannot write. Because the moment they sit down in front of a blank text editor, the operational context evaporates and what comes out is generic.

In our experience, the interesting material is almost always a specific deviation from the textbook answer. A customs officer who interpreted an HS code unusually. A Coupang policy that changed in a way the official documentation does not yet reflect. A Korean entity registration that hit a snag because of a non-resident director's address format. These are exactly the things that make our guide to setting up a Korean entity as a non-resident foreigner useful — and they are also exactly the things that disappear from memory inside 48 hours.

Why a phone call, and not a Notion doc

We tried the usual things first. A shared Notion page for "blog ideas." A Slack channel called #content-raw. A Friday review meeting where someone is supposed to write down what they learned that week.

None of it worked. Not because the team is undisciplined, but because writing — even bullet-point writing — is a different cognitive mode from operating. Switching modes has a cost, and on a busy day the cost is not paid.

A phone call has none of that friction. You answer it the way you would answer any call. You talk for two or three minutes. You hang up and go back to whatever you were doing. The agent handles transcription, structuring, and a draft pass.

Two calls a day is deliberate. One in the afternoon, when most customs and platform issues have surfaced. One in the evening, when there is enough distance from the day to notice patterns. We tried once a day and it missed too much. We tried three and it became a chore.

What the agent is actually doing

The agent is not magic, and it is worth being precise about what it does and does not do.

It records a short conversation. It transcribes it. It identifies whether the source material contains a real operational topic — a regulation, a specific client situation, a process detail — or whether it is throat-clearing. If there is a topic, it drafts a blog post in our voice, pulls the relevant facts from our internal knowledge base, adds the right internal links, and submits the draft to a review workspace. A human still reads every draft before it is published.

What it does not do:

• It does not invent statistics or regulations. If the founder did not say a number, no number appears in the draft.
• It does not publish anything autonomously. Every post is reviewed.
• It does not replace subject-matter posts that require a specialist's hand — for example, a deep technical post like our piece on when a foreign EMC report is enough for KC certification was written the slow way, by the engineer who actually does that work.

The editorial constraints we baked in

The interesting design work was not in the transcription or the LLM call. It was in the constraints.

A voice agent left to its own devices will produce hype-shaped content. It will say "seamless." It will say "unlock." It will say "in today's competitive global marketplace." Our brand voice rules explicitly forbid all of those phrases, and the agent's drafting prompt enforces them at generation time.

It also enforces structural rules that matter for how readers — and answer engines — consume the post. The opening paragraph has to be a real answer, not a wind-up. Headings have to map to subtopics a searcher actually cares about. Internal links have to be contextual and pulled from a real index of published posts, not invented. When the agent does not have enough material to write 1,800 words honestly, it writes a shorter Journal-style note instead of padding.

That last point matters. A founder who is mid-sprint will not write 1,800 words. He will, however, talk for three minutes about the week's most annoying customs issue. The agent's job is to make that three minutes turn into something useful — and to refuse to turn it into something dishonest.

Why this is a Kontactic Journal post and not a marketing post

We thought about writing this as a "look at our cool AI tool" piece. We decided not to.

The honest framing is that this is a piece of internal tooling that helps us run a content function with a small team. It is not a product. It is not for sale. It is not a differentiator we want clients to evaluate us on. What clients should evaluate us on is whether the operational guidance in the rest of this blog — pieces like our decision framework for Rocket Growth versus cross-border selling, or the founder's note on sequencing operational readiness before ad spend — actually helps them make decisions.

The voice agent is upstream of all of that. It is part of why the rest of the blog reads the way it does.

What we are still figuring out

Two open questions remain.

The first is how to handle multilingual source material. The founder's first instinct on a call is often to switch to Korean for nuanced operational terms — 위탁판매대행, 통관, KC 인증 — and then back to English for the framing. The current agent handles the switch reasonably, but the structuring of code-switched transcripts is still rough.

The second is how to keep the agent honest about scope. A three-minute voice memo is sometimes genuinely thin — the founder was tired, or the day was uneventful, or he was demonstrating the system to his wife rather than reporting a real operational lesson. The agent has to recognize those calls and refuse to pretend they are blog posts. The current heuristic is conservative: when in doubt, write a short Journal note about the meta topic rather than fabricating substance. That is what this post is.

Common questions

Is the voice agent a product Kontactic sells? No. It is internal tooling. Our service tiers — Spark, Flame, and Blaze — are described in our guide to selling on Coupang as a foreign brand. The blog agent is not part of any of them.

Does an AI write your blog? A drafter assists. A human reviews every post before publication. Technical posts are written by the operator who actually does the work.

Why publish this at all? Because the question "how does Kontactic produce so much specific operational content with a small team" comes up often, and the honest answer is: by lowering the friction between what our operators know and what ends up on the page.