Segment + Koji: Turn CDP Audiences Into AI Interviews — and Pipe the Answers Back Into Every Tool

Answer first: Segment is the system of record for who your users are and what they do. Koji's AI-moderated interviews answer the question a CDP structurally cannot: why. The integration runs both directions. Into Koji: a Segment Audience or a forwarded Segment event triggers a Koji AI interview to exactly those people — trial stallers, milestone-hitters, freshly-churned accounts. Back into Segment: the moment an interview is analyzed, Koji fires a Koji Interview Completed track call and updates user traits like last_interview_sentiment and last_interview_themes, which Segment then fans out to your warehouse, email platform, and dashboards automatically. There is no one-click connector to install — Koji exposes a headless API, HMAC-signed webhooks, and a Zapier app, and Segment speaks all three. A working v1 takes about 30 minutes.

If you have ever built a Segment Audience called "activated but silent for 14 days" and wished you could just ask them what happened, this is the integration that turns that audience into a stack of transcripts by Friday.

Why connect Segment and Koji

A CDP is superb at behavioral truth and blind to motivation. Teams usually patch that blindness with low-signal tactics: a quarterly NPS blast, a Slack message to CS asking for "vibes," or five calendar-negotiated interviews a month that sample a rounding error of the cohort. Koji removes the scaling ceiling — its AI moderator runs voice or text interviews around the clock, with no scheduling and no human moderator — but the real unlock is precision targeting. You do not want to interview "users." You want to interview the 212 people in one specific Segment Audience. Segment already computes those audiences continuously; the integration simply hands each new member to Koji.

Because Koji's AI asks adaptive follow-up questions instead of running a fixed script, one triggered interview produces the kind of "why" a survey never reaches — and every answer flows back into the same Segment identity you started from. Platforms like Koji make the depth of a moderated interview available at survey scale.

What flows in each direction

Segment → Koji (trigger interviews from behavior)

• A user enters a Segment Audience (Engage/Personas) or fires a qualifying track event.
• Segment forwards that to Koji through one of three transports (below).
• Koji starts an interview for that respondent, keyed by userId or email, and delivers a personalized link over email, SMS, or an in-product embed.

Koji → Segment (insights become traits and events)

• On each interview reaching analysis_ready, Koji's webhook posts to your forwarder.
• The forwarder calls Segment's HTTP Tracking API: a track event named Koji Interview Completed (properties: top themes, sentiment, quality score, interview mode, transcript URL) plus an identify call setting durable traits.
• Segment fans those traits to every destination — Snowflake/BigQuery, Braze, Customer.io, Mixpanel, your BI dashboards — with zero extra wiring.

Three ways to wire it up

1. No-code with Zapier (about 30 minutes). Use Segment's Zapier destination (or an Audience webhook) as the trigger and Koji's "Start Interview" Zap action as the effect. For the return trip, trigger a Zap on Koji's webhook and use Segment's Zapier action to send track/identify. Best for a first test — see Connect Koji to Zapier.

2. Segment Functions forwarder (about 1 hour). Write a small Destination Function that receives Koji webhook payloads and calls Segment's Tracking API, plus an Audience webhook (or Source Function) that calls Koji's headless API to start interviews. This is the clean production path and keeps everything inside Segment.

3. Reverse ETL / warehouse-first (production-grade). If your audiences live in the warehouse, sync them to a webhook with Hightouch or Census and call Koji's Start Interview endpoint. Koji insights land back in the warehouse as a modeled table you can join to any metric.

Identity resolution: matching interviews to the right user

Two clean patterns:

• Email-keyed: pass the Segment identifier as respondent_email; Koji echoes it on the webhook so you can identify the right user.
• ID-keyed: if your userId is a UUID, pass it as Koji's respondent_id (separate from email). The same key returns on the webhook, so insights attach to the exact profile — even for pseudonymous users.

For sensitive studies, run the interview in anonymous mode: no email is collected, and you simply skip the Segment write-back.

Design the interview so the data is actually joinable

The reason Segment + Koji beats a fancy survey link is Koji's structured questions. Every study can mix six question types — open_ended, scale, single_choice, multiple_choice, ranking, and yes_no — and each produces a typed, aggregatable value alongside the conversational transcript. A scale NPS item or a single_choice "primary reason" becomes a clean property you can write straight back to Segment as a trait and segment your entire base by. See the structured questions guide to design questions that analyze themselves.

High-value Segment audiences to point at Koji

• Activation drop-offs — signed up, never reached the aha event (feed onboarding drop-off research).
• Expansion signals — power users still on a low tier (feed account expansion research).
• Silent churn risk — usage decayed 50% month-over-month.
• Just-churned — subscription cancelled in the last 24 hours (route to cancel-flow exit interviews).
• New power users — crossed a usage milestone; capture testimonials while enthusiasm is high.

Example: closing the trial-conversion loop

1. Segment Audience: trial_started AND NOT subscription_created AND days_since_signup >= 10.
2. Audience entry triggers a 4-question Koji AI interview (one single_choice on the main blocker, two open_ended probes, one scale on likelihood to pay).
3. Interview completes: Koji Interview Completed lands in Segment with theme = "SSO missing", sentiment = "frustrated", quality_score = 0.86.
4. Segment routes it to Snowflake (for the PM's trial-conversion model) and to Customer.io (which enrolls "frustrated / SSO" users in a tailored win-back sequence).

You now segment trial conversion by stated reason, not guesswork — exactly what free trial feedback research is for, running continuously.

Plans and limits

Webhooks and the headless API are included on the Interviews plan (€79/month, 79 credits) and Enterprise. The Insights plan (€29/month) does not include webhooks, but Zapier and CSV import still work, so you can build a manual version on any plan. Interview volume is governed by your credit balance; one completed interview consumes one credit.

The webhook payload, field by field

When an interview reaches analysis_ready, Koji's webhook delivers a compact JSON object your forwarder maps onto Segment. The fields you will use most:

• respondent_email / respondent_id — the identity key you passed in, returned unchanged for joining.
• sentiment — an overall label (for example positive, neutral, frustrated).
• themes — the ranked list of coded themes across the conversation.
• structured_answers — one entry per structured question, each with its typed value.
• quality_score — a 0-to-1 confidence signal you can filter low-effort interviews on.
• transcript_url — a link back to the full conversation in Koji.

Because you decide which of these to forward, you keep tight control over what ever lands in your CDP.

Common pitfalls (and how to avoid them)

• Interviewing everyone at once. A CDP makes it tempting to fire at a 10,000-person audience. Start with a tight, high-intent slice (a few hundred) so you can read every transcript and refine the questions before scaling.
• Skipping identity mapping. If you forward insights without a stable respondent_id, they will not join back to the right profile. Decide your key — email or ID — before you build the return leg.
• Writing raw transcripts to Segment. Keep the CDP clean and privacy-safe by forwarding only structured insights; leave transcripts in Koji.
• No frequency cap. Suppress users who were interviewed in the last N days so a busy audience does not receive repeat invitations.

What to expect: week one vs. quarter one

In week one you will typically wire a single audience-to-interview flow and watch the first transcripts land. By the end of a quarter, most teams run three to five always-on flows — activation, expansion, churn, and testimonial capture — each feeding themed traits back into Segment. At that point qualitative "why" is no longer a quarterly project; it is a live property on every user, refreshed continuously and available to every downstream tool the moment it is created.

Related Resources

• Research Automation with Webhooks
• Connect Koji to Zapier
• Starting Interviews via API
• Structured Questions in AI Interviews
• Mixpanel + Koji
• Amplitude + Koji
• Understanding Themes & Patterns