Outbound surveys — the ethical framework

When we designed the outbound side of the AI Call Center, we decided early: we are not building a cold-calling system. We do not call anyone who hasn't previously opted in. That decision had costs (narrower use cases, slower scaling) and benefits (lower complaint rate, better brand perception, GDPR-aligned out of the box). Six months later the numbers vindicate the call.

The four non-negotiable rules

Our outbound module operates under four rules that aren't up for negotiation:

1. Opt-in only. The recipient must have previously consented (at purchase, at registration, or through explicit request) to being called. No purchased lists, no cold lists.
2. Recording disclosure in the first 5 seconds. Before any content begins, the agent says: "this call is being recorded and your participation is voluntary." If asked, it gives details.
3. Hang-up = consent revocation. If the recipient hangs up at any point, that's automatically logged as consent withdrawal. The system never calls them again, and any responses gathered to that point are anonymised (or deleted, by tenant config) within 24 hours.
4. One call, one try. If nobody answers, no automatic redial. A human can decide on a second attempt after 24 hours, but the AI can't do it on its own.

These rules are GDPR-compliant and satisfy the Hungarian NMHH telecoms regulator. Czechia and Poland are also covered. Germany and Austria are stricter and require a separate compatibility review per tenant.

The 6-question NPS survey we standardised on

Most tenants ask for NPS (Net Promoter Score). After six months of iteration we converged on this 6-question structure:

1. "On a scale of 0-10, how likely are you to recommend us?" — The classic NPS question.
2. "What was best about your recent experience with us?" — Open-ended, positive primer.
3. "Was there anything that caused difficulty?" — Open-ended, problem-finding.
4. "If we changed one thing, what should it be?" — Constructive feedback.
5. "Would you be willing to be quoted in a case study?" — Promoter conversion.
6. "Thanks for your time — anything else you'd like to add?" — Closer, opens the door.

With this structure the average call runs 2 minutes 40 seconds. 87% of respondents make it through all 6 questions. The 13% drop-off happens at questions 3 and 4; the open format probably steals time.

The 27% response rate

Most surprising finding: voice surveys land a 27% response rate versus email's typical 8%. We measured this across 4 different industries (e-com, B2B SaaS, healthcare service, education), and it's consistent everywhere.

Why does voice work better? Probably three things:

• People answer when they pick up — there's no "I'll get to it later" inertia like email has.
• Voice feels personal, the process flows.
• The AI waits patiently and lets the answer breathe, no character limit.

What we didn't win

Voice surveys have limits too. Time-of-day dependency is much stronger: 18:00-20:00 yields 31% response, weekends 19%, Monday mornings only 12%. We tune this together with each tenant.

Another limit: numeric or counting questions ("how many times did you use the product in the last month?") get unreliable answers by voice. Those belong in email or a web form.

Complaint rate

In six months and 31,000+ outbound calls we logged a 0.4% complaint rate. That means fewer than 4 people per 1,000 complained about the call (either to the tenant or publicly). The wider industry average is 2-3%; the low number is the result of the four rules.

Complaint breakdown: 41% "called too early" (timing issue, we fix it), 28% "don't remember consenting" (better consent-record presentation helps), 31% other. Not a single complaint was about it being an AI voice; that's an important signal.

What we will never build

Finally, the blacklist that matters for the brand:

• We won't build campaigns based on purchased lists
• We won't hide that an AI is speaking
• We won't record secretly
• We won't call outside the 9-19 window (unless the customer explicitly asked)

These four bans stay. Not because they're technically impossible, but because they're what keeps the product sustainable long-term.