AI + HumanFeedback and NPS surveys

Feedback calls where AI asks and your team fixes it the same day

The AI places the survey calls, asks one score and one open why, and tags what customers say. Your team takes the unhappy customers, so feedback becomes a fix, not a slide.

  • One score, one open why, no 20 questions
  • Can say it is an AI, offers a person
  • Low score reaches your team same day
Feedback and NPS surveys: Thursday 6:15 pm · two days after a repair visitAI agent on the line
AI agentOne score, one open why, no 20 questions
OutcomeDetractor reached a person on the same call. Reason tagged, no repeat.

After the visit or delivery

Two days later, one score and one why. The AI asks everyone sampled.

After a support case

CSAT or effort, asked the same way every time, reason tagged.

The unhappy answer

A 4 about a late part reaches a service lead the same day.

The happy customer

A thank-you, and a review link by text after the survey ends.

01 · Why AI + Human

Why customer feedback will run on AI and people together

Most customers never answer a survey. Pew's telephone polls reached 6% in 2018, and online surveys vary widely. A person-only team cannot call every customer after every visit, and an AI-only survey stops at the score. Your detractors need a person who can repair the problem, and 87% say a way to reach a human is essential.

AI agentsLow cost per call, instant scale, consistent
  • Calls every sampled customer with the same neutral question
  • Reaches customers who ignore email, at the right hour
  • Tags the reason in the customer's own words, on every call
On its ownCannot judge a high-value relationship or make an exception, and can miss the mood behind a flat 7.
Your peopleJudgment, authority, care, new solutions
  • Call back a detractor and repair what went wrong
  • Hear anger or worry and decide what the account deserves
  • Turn a pattern of reasons into a fix
On their ownCannot call every customer after every event, and a rushed caller can lead the answer.
Every customer asked, every unhappy answer answeredOne desk, one record of every call
  • The 4 about the late part reaches a service lead the same day, with the transcript
  • Leaders see tagged reasons and sample sizes, not a bare score
  • Agents spend time on repairs, not on dialing a survey list
Side by side
People only (traditional)
AI only
AI + Human
Reach
A few hundred calls a week at most
Whole sample, any hour
Whole sample, with people on the hard ones
Question consistency
Wording drifts by caller
Same neutral script every call
Same script, plus a person's follow-up
Reading the customer
Strong at tone and doubt
Hears words, may miss mood
Mood flagged, a person takes over
Fixing a problem
Can fix, if told in time
Records it, cannot decide
Detractor reaches an owner same day
Cost per survey call
Highest: paid time on every call
Lowest: billed per second
Low on surveys, people spent on repairs
How to think about the split
Ask with AI, decide with people

Let the AI ask the question and record the answer. Keep refunds, apologies and account decisions with a person who can say yes.

Close the loop before widening it

Add a survey moment only when detractors reach a person within your target. A score with no follow-up teaches customers that answering is pointless.

Keep the survey a survey

No offers or upsells on the call. A genuine survey is treated differently from a sales call, and a pitch erases that.

The short answer

When should a company survey customers by phone instead of email, and what happens to unhappy answers?

Use a phone survey when email is ignored, when customers are older or high-value, or when you need the reason behind a score. An AI agent places the call, asks one score and one open why, and tags the answer. A low score or an angry answer goes to a person the same day, with the transcript. Keep the survey free of selling.

Reviewed November 2026 · 15 sources cited below
  • AI asks: one score (CSAT, NPS or effort) plus one open why
  • People take: detractors, angry answers, high-value accounts
  • Survey calls stay free of selling, or they become marketing calls
  • Judge it by loop closure time, not response rate alone
02 · The stakes

Most customers never answer a survey, and the ones who do are not a random sample.

Response rates are low on every channel, so a score from a few respondents can mislead. A call reaches people email misses, but only if it is short, honest about who is calling, and followed by action.

6%

typical response rate in Pew Research Center telephone polls in 2018, down from about 9% earlier (cold random-dial polls, not customer lists)1

44.1%

average response to online surveys across 1,071 rates in published research, but 20.3% where samples exceeded 2,500 (2022 meta-analysis)2

62%

of customers in a survey of 75,000+ reported contacting a company repeatedly to resolve an issue (Customer Contact Council, HBR 2010)3

87%

say a way to reach a human is essential when GenAI is used (Gartner, 3,566 customers, Feb to Mar 2026)4

03 · The whole journey

Where AI and your team gather feedback across a customer's whole life with you

Feedback is not one survey. It follows the customer from the first order to the day they might leave. Here is where AI agents and people on one desk ask, listen and close the loop at each stage.

Purchase and onboardingSTAGE 1
First useSTAGE 2
After a support caseSTAGE 3
Relationship checkSTAGE 4
Advocates and add-onsSTAGE 5
At-risk customersSTAGE 6
Lapsed customersSTAGE 7
The moment
The customer ordered or signed up and wonders if it went right.
The first week is when confusion becomes quiet churn.
The ticket is closed, but the customer may still be unhappy.
Months pass and the customer's view of you drifts unseen.
A happy customer is willing to recommend but nobody asks well.
A detractor or a quiet account is closer to leaving than it looks.
A customer left and nobody learned why.
AI agent
Calls a day or two after delivery or setup: one score, one why, tags missed steps.
Asks how first use went, offers a text link to help, tags what confused them.
Asks CSAT or effort and the reason, and flags repeat contact or a transfer.
Runs a sampled NPS call, respects contact caps, tags reasons by segment.
Thanks promoters, then texts a review link after the survey ends, where allowed.
Flags low scores and cancel words, offers a person now or a same-day call-back.
Places a short, consented exit call asking the main reason, tagged by theme.
Your team
Calls back the customer whose setup went wrong, with the details already in hand.
Takes the customer who is stuck and fixes it, then logs the reason.
Reopens the case or calls back the customer who says it was not fixed.
Calls high-value accounts personally and brings the themes to leaders.
Handles any referral or upgrade conversation the customer opens.
Listens, repairs and decides what to offer, with the transcript in view.
Takes the customer who wants to come back or still has a complaint.
What to track
Early problems caught, onboarding score with sample size
First-use problems fixed before they become cancellations
CSAT, effort score and 7-day repeat contact
NPS with sample size, themes with owners
Reviews and referrals asked in the right moment
Detractors contacted in time, loop-closure time
Lapse reasons captured, win-back conversations

Read down a column for one stage, or across a row to see what the AI and your team each do from first contact to win-back.

04 · AI or person

Which feedback calls the AI owns, which it starts, and which go straight to your team.

The AI is good at asking the same neutral question, at the right time, to every customer in the sample. A person is better at an upset customer, a high-value account and any answer that needs a decision.

AI handles it end to endAI starts, a person finishes with the contextStraight to a person
Call type
Volume
Who takes it
Why
Post-delivery or post-visit check
Two days after the install, the delivery or the service visit: "How satisfied were you, 1 to 5? What made it that?"
high
AI handles it
One short, neutral script and a fixed window after the event. The AI calls everyone sampled, not only the ones who answer email.
Support case CSAT and effort score
After a closed ticket: "How much effort did it take to get this solved, 1 to 5?"
high
AI handles it
Asked the same way on every case. The reason is tagged, such as transfer, wait or repeat contact.
Relationship NPS pulse
Quarterly 0 to 10 recommend question to a random sample of active customers
medium
AI handles it
Sampling, contact caps and a fixed script matter more than speed. The AI applies them on every call.
Customers who ignore email surveys
Older or less digital customers, and people who never open the link
medium
AI handles it
A call reaches people the link misses. Offer a text link or a call-back time if they are busy.
Detractor with a fixable problem
A 4 because the part was late and nobody called
medium
AI, then a person
AI gets the score and reason, offers a person now or a same-day call-back. The person fixes it and logs the outcome.
Promoter: thanks, review or referral
A 10 from a customer who offers to tell a friend
low
AI, then a person
Only after the survey ends, and only where your rules allow. AI thanks them and texts a link; a person handles any referral.
Angry, distressed or legal-threat answer
"I'm filing a complaint." A safety issue. A billing error that cost money.
low
Straight to a person
High stakes and high emotion. The AI recognizes it, stops the survey and transfers with the transcript.
High-value account feedback
A top account's renewal review or a named-contact relationship check
low
Straight to a person
The account owner should ask, listen and decide. Use the AI to book the time and pull the history.
Customer says they want to cancel
"Honestly, I'm thinking of leaving." Surfaces mid-survey.
low
Straight to a person
A retention offer is a judgment about the account. The AI stops the survey and hands over the reason in the customer's words.
TakeawayThe AI asks and records. Your team decides and repairs. Add a call type to the AI only when your detractor hand-offs reach a person within your own time target.
05 · One call, start to finish

One feedback call, from the first sentence to a fixed problem.

The AI gets the score and the reason. A person gets the unhappy customer, with everything the customer said. The loop closes in the same day, not the next quarter.

Step 1

Picks the sample and the moment

The campaign selects customers by rule, skips anyone surveyed recently or in an open dispute, and schedules the call a day or two after the event, inside calling hours.

Step 2

Opens with who, why and AI

The AI names the business, says it is an AI and that the call is recorded, and asks for two minutes. It does not sell. It offers a person at any point.

Step 3

Asks one score and one open why

It asks a neutral 0 to 10 or 1 to 5 question, then "what was the main reason?". The customer answers in their own words. It asks one follow-up at most.

Step 4

Tags the reason and checks the score

It reads back the score, tags themes such as delay, price or staff, and flags a low score, anger or a cancel signal as it hears it.

Step 5

Hands the unhappy customer to a person

A detractor is offered a person now or a same-day call-back. The agent sees the score, the reason, the transcript, the mood and the order or case.

Step 6

Closes the loop and reports

The agent fixes it or sets the next step. AI Assist proposes the notes. The outcome is logged with the score, so response, loop time and themes are reported with sample sizes.

06 · Start small

Start with one feedback moment and one small sample. Prove the loop before you widen.

Pick one event, such as a closed support case or a completed service visit. Write the three questions, name who takes detractors, and set the time target before the first call. Measure the loop, not only the score.

Start with this

Post-visit or post-case feedback call to a random sample

The event is clear, the customer remembers it, and a low score has an obvious owner. A wrong survey call costs a minute of one customer's time, not the relationship.

How calls reach the AI

One outbound list: customers whose case or visit closed two days ago, sampled at random. The dialer calls inside quiet hours. A person is one ask away, and unhappy answers route to a named queue.

Week 1

Baseline and design

  • Record today's survey response rate, score, and how long a detractor waits for a call
  • Write one score question, one open why and one stop-or-person line, with neutral wording
  • Set sampling rules, a contact cap per customer and the consent and hours rules
  • Name the person or queue for detractors and set a same-day target
Weeks 2 to 3

Small sample, every call read

  • Call about 100 customers a week and read or listen to every transcript
  • Check wording for leading phrases and customers who felt pushed
  • Test that each low score reaches a person with transcript and reason
  • Count opt-outs and "are you selling something" reactions
Weeks 4 to 6

Full sample, tagged themes

  • Run the full sample with the response and completion rates reported with sample sizes
  • Review tagged themes with the owners who can fix them, weekly
  • Compare phone responders with your email responders by age and segment
  • Track time to first contact for every detractor
Weeks 7 to 10

Add a second moment

  • Add a second event, such as delivery or a quarterly NPS pulse
  • Set pass marks for the new moment before turning it on
  • Test the optional promoter thank-you text, after the survey ends and where allowed
  • Decide in writing: widen, hold or roll back

Expand when all of these are true

  • At least 90% of detractors contacted by a person within your time target (planning: same business day)
  • Complaints or opt-outs from survey calls below your own threshold, reviewed weekly
  • Phone and email samples compared, and no segment missing from the results
  • Every reported score shown with its sample size and response rate
07 · What to measure

Eight numbers that show whether feedback calling is working.

A score alone says little. Pair it with how many answered, who answered, and how fast unhappy customers heard from a person. Take baselines from your own survey tool first. Targets are planning ranges, not promises.

Metric
Typical today
Planning range with AI and people
Response rate1
Completed surveys divided by customers sampled and called. Report by segment and by channel, not one blended figure.
6% cold phone polls (2018); 44.1% online, published studies
Set from your own email baseline, then compare
Completion rate
Surveys finished divided by surveys started. A drop points to a long script or a wording problem.
No independent benchmark found
85% or higher on a 2-question script
Score with sample size5
CSAT, NPS or effort average, always shown with the number of responses and the share of the sample that answered.
Your last quarter, by segment
Report only segments with 100 or more responses
Detractors contacted in time
Share of detractors (0 to 6 on NPS) that a person reached within your target, such as one business day.
No independent benchmark found
90% or higher within one business day
Loop-closure time6
Hours from the low score to the first person-to-customer contact, then to a recorded fix. Report the median and the slowest 10%.
Your current time, often weeks
Median under 1 business day to first contact
Repeat-contact rate3
Customers who contact you again about the same issue within 7 days, among survey respondents. Match on customer ID and reason.
62% reported contacting repeatedly (survey, 2010)
Below your pre-survey baseline
Theme coverage
Share of open reasons tagged to a theme with an owner. Sample 25 transcripts a week to check the tags.
No independent benchmark found
80% tagged, each theme with a named owner
Opt-outs and complaints
Customers who say stop, or complain about the call, divided by customers called. Review each one.
No independent benchmark found
Under 1% of calls, reviewed weekly

Planning ranges are targets to test in a pilot, not promises. Measure your own baseline for two weeks before you change anything.

08 · The return

What feedback calling is worth to a company surveying 3,000 customers a month.

Use your own numbers in the calculator. The example is one company calling 3,000 customers a month after a purchase or a case. The AI finishes most calls and reaches customers who would never answer an email.

A worked example. Move the sliders to match yours.

3 minaverage call length
$5,300loaded monthly cost per person
30 a daycalls one person handles
$40 · 5%value of one unhappy customer reached and recovered before they leave · share of recovered calls that become one
Estimated monthly gain
$8,284a month
Today, per call
$7.07
With VoiceGenie
$1.27
1,800more calls answered a month
90extra unhappy customer reached and recovered before they leaves a month
$3,600value of those outcomes
600calls your team still takes, with context

AI at $0.14 a minute with volume credit applied, plus Insights seats for the people who take handed-over calls. Phone line charges are your carrier's and are not included. Illustration, not a quote. Open the full calculator

Where these numbers come from

  • Cost per person is a US estimate for a caller on loaded pay, about the level used on the customer support page. Use your own survey team's cost.
  • Calls per person per day of 30 assumes short 3-minute calls with dialing and notes. Replace it with your own dialer data.
  • Missed share of 90% is a planning figure for customers who never answer today's email survey. For context, Pew's 2018 phone polls reached 6% 1 and online surveys averaged 44.1% 2.
  • AI finishes 80% assumes a two-question script. Detractors and angry answers go to people and are not counted as finished by the AI.
  • Outcome value and conversion are placeholders for one saved customer. Set them from your own churn and revenue data.

What the dollar figure leaves out

  • Your team spends time repairing problems, not dialing a list to ask how it went
  • Leaders see the reasons behind a score, in customers' own words, with sample sizes
  • Customers hear from a person quickly after a bad experience, before they write a review
09 · Rules to get right

Rules to get right before an AI places a survey call.

A genuine survey is treated differently from a sales call, but an AI voice, a cell phone, call recording and personal data still bring rules. This is a checklist for your counsel, not legal advice. Rules marked US apply only there.

US

AI voice calls count as artificial-voice calls7

The FCC ruled in February 2024 that AI-generated voices are artificial under the TCPA, so a call needs the called party's prior express consent. Written consent applies if it includes advertising or telemarketing. Identify your business at the start.

US

Consent to call cell phones, and proof of it8

Artificial-voice calls to cell numbers need prior express consent. Ask counsel whether the way you collected the number covers survey calls. Keep the wording, date, number and channel for each customer.

US

A survey cannot be a sales pitch9

The FTC says a call made solely to conduct a survey is exempt from the Do Not Call provisions. A caller who takes a "survey" and also offers to sell must follow them. Keep offers and upsells out of the call.

US

Calling hours and stop requests10

The FTC rule limits telemarketing to 8 am to 9 pm at the customer's location. Pure surveys sit outside it, but use that window as a floor, and a tighter one where your counsel advises. Treat any "stop calling" as final and suppress the number.

US states

Recording consent in all-party states11

California and other all-party consent states need consent before recording starts, not after. Say it in the first sentence. Ask your vendor whether it can use your call audio for its own purposes.

UK/EU

UK and EU: research vs marketing, and data basics12

The ICO says genuine market research is not direct marketing, but a survey with promotion, or run to build a marketing list, is. Pick a lawful basis, tell customers how you use their data, and honor objections.

Four questions to ask any voice AI vendor, including us

  1. Can the AI say it is an AI and that the call is recorded in the first sentence, before any content?
  2. How do you suppress a number after a stop request, across every campaign and both AI and human agents?
  3. Is any of our call audio or transcript used to train models? Is that in the contract?
  4. Show me the hand-off: what does the agent see when a detractor is transferred?
10 · Where it plugs in

Where feedback calls read the customer and write the result.

A survey call needs the customer, the event and the owner, and it writes the score back. VoiceGenie connects to HubSpot, Salesforce, Zoho and Google, and to anything else through the open API, webhooks or Zapier.

CX and survey platforms

  • QUQualtrics
  • MEMedallia
  • SUSurveyMonkey
  • DEDelighted
  • ASAskNicely
  • TYTypeform

CRM and help desk

  • HUHubSpot
  • SASalesforce
  • ZOZoho
  • ZEZendesk
  • FRFreshdesk
  • SEServiceNow

Order, visit and delivery records

  • SHShopify
  • NENetSuite
  • SEServiceTitan
  • GCGoogle Calendar
  • SHShipStation

Phone system and alerts

  • TWTwilio
  • TETelnyx
  • VOVonage
  • Any SIP trunk or PBX
  • SLSlack
  • ZAZapier
Keep your numbers and carrierTwilio, Plivo, Telnyx, Vonage or any SIP trunk or PBX. Put the AI behind one menu option, on overflow or after hours.
CRM and calendarsHubSpot, Salesforce, Zoho and Google connect directly.
Everything elseThe systems above connect through the open API, post-call webhooks or Zapier.
11 · One platform

One platform, set up once, so every call improves AI and people

You do not need one vendor for the dialer, another for the AI agent, a third for the agent desk and a fourth for AI tools that help your people. VoiceGenie covers all four on one record. You connect your systems, write your script and set your rules once. Each corrected answer or QA finding then improves the AI and your agents together.

The usual way

Four tools, four set-ups

Dialerown contract · own integration
own training · own reports
Voice AI agentown contract · own integration
own training · own reports
Agent deskown contract · own integration
own training · own reports
Agent assist and QAown contract · own integration
own training · own reports
With VoiceGenie

One platform, set up once

AI agents
AI-native dialer
Human Agent Desk
AI Assist
Shared by the AI and your people
  • Your phone numbers and carrier
  • CRM and calendar connections
  • Knowledge and call rules
  • Recordings, QA and analytics
One customer and case record

Connect the CRM and help desk once. The AI knows the event it is asking about. The agent sees the same record after hand-off.

One script and one set of rules

Write the questions, quiet hours, contact caps and stop handling once. They apply to every campaign, AI or human.

One recording and QA view

Every survey call and every call-back is recorded, transcribed and QA-scored in one view. A finding fixes a script or coaches an agent.

12 · What goes wrong

Five ways feedback programs go wrong.

Surveying everyone and acting on no one
Decide who owns each theme and each detractor before the first call. A survey with no follow-up teaches customers that answering does nothing.
Letting the person being rated pick who is called
Sample by rule. If the agent, branch or technician being scored chooses the list, the score measures their selection. Draw the sample from the full event list.
Leading questions and long scripts
Ask one score, then "what was the main reason?". Avoid "how excellent was our service?". Long scripts lower completion and push people to hang up.
Slipping a sales pitch into the call
A survey that sells is a marketing call. Keep offers out, ask for a review only after the survey ends and where allowed, and let a person handle referrals.
Reporting a score without the sample size
Show the number of responses and the response rate beside every score. Responders are not a random sample, so say so, and compare phone and email responders.
13 · Questions

Questions people ask before they start

Is a phone survey better than an email survey?

It depends on who you need to hear from. Email is cheaper and fine for engaged customers. A phone survey reaches people who ignore email, such as older or less digital customers and high-value accounts, and it captures the reason in their own words. Both have low response rates, so compare the two samples and report each.

What is the difference between CSAT, NPS and customer effort score?

CSAT asks how satisfied a customer was with one interaction. NPS asks how likely they are to recommend you, on a 0 to 10 scale, and measures the relationship. Customer effort score asks how much effort it took to get something done, on a 1 to 5 scale. Use one per call, plus one open why.

Can an AI voice agent legally make survey calls?

In the US, the FCC ruled in February 2024 that AI-generated voices are artificial voices under the TCPA, so calls need prior express consent, including to cell phones. Written consent applies if the call includes advertising. Say it is an AI, keep proof of consent, honor stop requests, and ask your counsel.

Does a survey call count as telemarketing?

A call made solely to conduct a survey is exempt from the FTC's Do Not Call provisions. If the caller also tries to sell, it is treated as telemarketing and must follow them. In the UK, the ICO calls selling under the guise of research sugging. Keep offers out of the survey and ask for reviews afterwards, where allowed.

What happens to a customer who gives a low score?

The call routes to a person. The AI offers to bring in an agent now, or sets a same-day call-back. The agent sees the score, the customer's reason in their own words, the transcript and the order or case. This is the inner loop. Trends across customers go to leaders as the outer loop.

What is a good response rate for a phone survey?

No benchmark applies to customer surveys by phone. Pew's cold telephone polls reached 6% in 2018, and online surveys in published research averaged 44.1%, so those numbers are not comparable. Measure your own email baseline, report response rate and sample size with every score, and compare responders with your customer base.

How much does an AI survey call cost?

VoiceGenie charges $0.14 per minute for the AI, billed per second, with the phone line extra. A three-minute survey call is about $0.42 in AI time. Human seats for call-backs start at $29 per person per month on annual billing. There is a free trial credit and no minimum commitment.

Will customers hate being surveyed by an AI?

Some will. Gartner found 87% say a way to reach a human is essential when GenAI is used. So the call says it is an AI in the first sentence, takes about two minutes, and offers a person on request. Respect stop requests, cap how often you call, and do not sell on the call.

Try it on one call type

Hear it handle your calls before you change anything.

Talk to our AI agent in your browser, or start a trial and point one line at it. Your team keeps every call that needs them.

No credit card · No minimum commitment · Keep your carrier

Net Promoter, Net Promoter Score and NPS are registered trademarks of Bain & Company, Inc., NICE Systems, Inc. and Fred Reichheld.

Sources

  1. Pew Research Center, Response rates in telephone surveys have resumed their decline (2019).
  2. Computers in Human Behavior Reports (Wu, Zhao, Fils-Aime), Response rates of online surveys in published research: A meta-analysis (2022).
  3. Harvard Business Review (Dixon, Freeman, Toman), Stop Trying to Delight Your Customers (2010).
  4. Gartner, Gartner Survey Finds 87% of Customers Say Companies Using GenAI Must Provide Access to a Human Agent (2026).
  5. Net Promoter System, Measuring your Net Promoter Score (2026).
  6. Bain & Company, Customer experience tools: Net Promoter System (2026).
  7. Federal Communications Commission, Declaratory Ruling FCC 24-17 on AI-generated voices under the TCPA (2024).
  8. Cornell Legal Information Institute, 47 CFR 64.1200, Delivery restrictions (2026).
  9. Federal Trade Commission, Q&A for telemarketers and sellers about DNC provisions in the TSR (2026).
  10. Cornell Legal Information Institute, 16 CFR 310.4, Abusive telemarketing acts or practices (2026).
  11. Epstein Becker & Green, Eavesdropping by Algorithm: How Courts Are Applying Decades-Old Privacy Laws to AI (2026).
  12. Information Commissioner's Office (UK), Identify direct marketing (2026).
  13. Insights Association, FTC advisory opinion on survey researcher access to the Do Not Call Registry (2004).
  14. Information Commissioner's Office (UK), A guide to lawful basis (2026).
  15. Net Promoter System, About the Net Promoter System (2026).