Kenyan enterprises already buy customer-experience metrics: the Kenya Bankers Association publishes an industry NPS every year, and Safaricom reports NPS to its board. The weak link is collection — the incumbent channel is SMS, and the best published SMS survey response rate in Kenya is about 12%. This guide builds voice surveys customers actually finish: a post-call CSAT transfer, an outbound one-question survey call with getDigits, keypad-first question design, the per-second cost math against SMS, and a loop-closer that turns every low score into a callback.
The metric itself needs no selling here. The Kenya Bankers Association runs an annual industry customer-satisfaction survey and reported a banking-industry NPS of 42.6 in 2025, down from 44.0 in 2024, even as 84.6% of customers rated their bank's service excellent or good (per the KBA 2025 survey). Safaricom tracks NPS as a board-level KPI in its annual reports. When the industry association and the country's largest company both publish the number, "should we measure NPS?" is a settled question.
How to collect it is not. The market grew up on SMS: mSurvey, founded in 2012, ran SMS-based customer feedback for Safaricom, rebranded to Ajua in 2019, and merged with Rate My Service in August 2025 — the combined entity claims 45 Kenyan enterprise customers and about 80% of Kenya's CX market (per TechCabal, 14 Aug 2025). That consolidation says two things: the market is real, and it is concentrated in one modality. SMS-first collection has two structural problems that no amount of copywriting fixes: most recipients never respond, and the ones who cannot comfortably read the message never could.
The published numbers are not kind to text surveys, even in their best market:
A four-country study in the International Journal of Public Opinion Research (2019) found SMS survey response of roughly 12% in Kenya — the best rate in Africa — versus under 1% in Ghana and Nigeria. Kenya's best case still loses 88 of every 100 people you message.
A mode experiment in Nigeria published in Survey Research Methods found live phone interviews at roughly 15% response, IVR at about 3%, and SMS at about 0.2% — all against cold research panels.
Those cold-panel IVR numbers deserve context. Calling strangers to ask research questions is the hardest possible setting. A post-transaction survey is the opposite setting: you are calling your own customer, minutes after you served them, about the thing that just happened, from a number they may already recognise. Warm-audience, in-the-moment DTMF surveys are a different exercise from panel research, which is exactly why they are worth building.
The second structural point matters as much as the first. An SMS survey assumes the recipient reads, in the language you wrote. A voice prompt in Swahili answered by pressing a key requires neither literacy nor a smartphone. The customers SMS surveys systematically exclude — older, rural, feature-phone, non-literate — are still customers, and a survey channel that never hears from them is quietly biased toward the segment that types.
The classic contact-centre pattern: when the agent call ends, the caller rolls straight into a one-question rating. With Sautikit's voice actions this is one response document, because actions execute top to bottom — when the dialed agent leg ends and the caller is still on the line, the call continues to the next action:
npm install @sautikit/node@0.2.0
import { sauti } from "@sautikit/node";// res.json(...) serializes the chain to the { actions: [...] } envelope.res.json( sauti .dial({ number: "+254720000100", timeout: 30 }) .say( "Before you go, one question about the call you just had. Swali moja tu kabla hujaenda.", { language: "en-KE" }, ) .getDigits({ timeout: 10, numDigits: 1 }, [ // seconds { say: { text: "On a scale of one to five, five meaning excellent, how satisfied are you " + "with the help you received? Bonyeza namba moja hadi tano.", language: "en-KE", }, }, ]) .say("Asante sana. Kwaheri.", { language: "en-KE" }) .hangup(),);
One implementation detail: when the digits arrive, the platform POSTs back to the same voice_callback_url that serves your inbound line. Mark the session as "in survey" (keyed by the callback's session id) when you return the survey actions, so your handler dispatches the follow-up digits to the survey branch rather than treating them as an IVR menu choice.
The more robust pattern is a dedicated survey number that dials the customer after the interaction — a closed ticket, a completed delivery, a resolved dispute. Your system triggers one call per event:
// Fired by your ticketing system when a ticket closes.import { SautikitClient } from "@sautikit/node";const client = new SautikitClient({ apiKey: process.env.SAUTIKIT_API_KEY });async function triggerSurveyCall(ticket) { const { call_id, session_id } = await client.calls.create({ from: process.env.SURVEY_NUMBER, // your dedicated survey caller ID to: [ticket.customerPhone], // "+2547..." // One survey attempt per ticket — even if the job re-runs. idempotencyKey: `csat:${ticket.id}`, }); // Store the mapping so the voice callback can find the ticket. await db.pendingSurveys.insert({ sessionId: session_id, callId: call_id, ticketId: ticket.id, });}
Two things carry this flow. First, the idempotency key (csat: plus the ticket id) means a retried job or a re-run backfill never dials the same customer twice — see idempotency on the calls API. Second, call routing lives on the number, not the call: point the survey number's voice_callback_url at your survey handler with PUT /v1/numbers/{id}/routing, and correlate each callback to its ticket by matching the callback's session id against the session_id you stored when you placed the call.
getDigits does not take a follow-up URL. When the caller presses a key, the platform POSTs back to your voice_callback_url with the Digits field populated, and your handler dispatches on it. One handler serves the whole survey:
import express from "express";import { sauti } from "@sautikit/node";const app = express();app.use(express.json()); // voice callbacks arrive as JSONconst surveyQuestion = () => sauti .getDigits({ timeout: 10, numDigits: 1 }, [ // seconds { say: { text: "Hello, this is Acme Bank with one quick question about " + "the service you received today. On a scale of zero to " + "nine, where nine means excellent, how likely are you to " + "recommend us? Kwa kipimo cha sifuri hadi tisa, tisa " + "ikiwa bora kabisa, bonyeza namba kwenye simu yako.", language: "en-KE", }, }, ]) .say("Hatukupokea jibu. Thank you for your time. Kwaheri.", { language: "en-KE" }) .hangup();app.post("/survey", async (req, res) => { const digits = req.body.Digits || ""; const sessionId = req.body.sessionId; if (!digits) { // First hit: the customer answered. Ask the one question. return res.json(surveyQuestion()); } // Follow-up hit: Digits is populated. Record and branch. const score = Number(digits); const pending = await db.pendingSurveys.findBySession(sessionId); await db.surveyResponses.insert({ ticketId: pending?.ticketId, score, respondedAt: new Date(), }); if (score <= 6) { // Detractor: flag for a human callback. await db.callbacks.flag(pending?.ticketId, { reason: "low_score", score }); return res.json( sauti .say( "Pole kwa hayo. We want to make this right. A member of our " + "team will call you back within one business day. Asante.", { language: "en-KE" }, ) .hangup(), ); } return res.json( sauti.say("Asante sana. Thank you for your feedback. Kwaheri.", { language: "en-KE" }).hangup(), );});
Note the nested block: the question plays inside the digit collection, so a caller who already knows their answer can press it mid-sentence and skip the rest of the prompt. timeout is in seconds — ten seconds of silence after the prompt ends, then the fallback thank-you runs and the call hangs up cleanly. Never leave dead air.
The strongest lever in voice surveys is restraint. Every additional question sheds respondents, and on a phone call there is no scrolling back to check what question three said. One question answered by most of your callers produces better data than five questions answered by a determined few. Pick the single number you will act on this quarter and ask only that.
CSAT fits a keypad natively. "One to five" is one keypress, no ambiguity. If you run CSAT, use it as-is.
NPS needs a remap. The canonical NPS scale is 0–10, and "10" is two keypresses — which forces multi-digit collection and a finish key, and multiplies input errors on the networks covered in our DTMF reliability guide. The practical fix is a 0–9 scale, collected with numDigits: 1 as above, with 9 anchored as "extremely likely". Score it with the same promoter/passive/detractor cut-offs shifted down one (9 promoter, 7–8 passive, 0–6 detractor). One caveat: a 0–9 keypad NPS is not directly comparable to a 0–10 SMS NPS, so treat it as your own trend line, not a cross-benchmark against figures like KBA's.
Three script rules finish the job:
Identify yourself in the first five seconds. "Hello, this is Acme Bank" before anything else; unexplained robocalls get hung up on.
Anchor the scale before the question. "Where nine means excellent" must land before "how likely"; callers cannot re-read the scale.
Keep the whole prompt under twenty seconds. Intro, scale, question. If the script needs a second breath, it is too long.
Timing beats wording. A survey call placed within minutes of the interaction reaches a customer who still remembers it; the same call two days later is an interruption. Trigger from the closing event, and as best practice keep automated calls inside daytime hours.
The say verb takes a language key — use en-KE, and write bilingual scripts the way Kenyan service lines actually speak: English framing with Swahili reinforcement of the action ("bonyeza namba moja hadi tano"). Anchor the top of the scale in both languages ("tisa ikiwa bora kabisa") so the meaning of the number never depends on which language the customer processed.
This is also where voice earns its structural advantage. The customer who cannot read your SMS hears the Swahili prompt, understands it, and presses a key. For customer bases with vernacular-first segments, a recorded human prompt in the local language via the play verb — same getDigits capture — outperforms any TTS. Keep one consistent voice across all your survey calls; the survey line becoming recognisable is part of the response rate.
Sautikit pricing (as of 2026-06-30): outbound calls at KES 3/min — KES 0.05 per second, billed per second after connect. See /pricing for current rates. Per-second billing is what makes one-question surveys cheap: a tight script runs about 20 seconds of prompt plus a keypress and a short thank-you — call it 30 seconds, or KES 1.50 per answered call. A call that is never answered never connects, so it costs nothing.
Worked model — the percentages are assumptions to plug your own numbers into, not published statistics. Suppose 100 closed tickets, 60 customers answer the call, and 80% of those who answer press a rating:
60 answered calls × KES 1.50 ≈ KES 90 total spend
48 completed responses → ≈ KES 1.90 per completed response
Now run the same arithmetic on SMS: at Kenya's best published response rate of ~12% (per the 2019 IJPOR study), cost per completed response is your per-message price divided by 0.12 — roughly eight times whatever one SMS costs you — and the responses skew toward literate smartphone users. Voice is not automatically cheaper per attempt; it is dramatically cheaper per completed response from a representative customer, which is the unit a CX programme actually buys.
SMS survey
Outbound DTMF call
Post-call transfer
Best published Kenya response
~12% (2019 IJPOR study)
No published Kenya benchmark; warm-audience completion is the driver
A survey programme that only produces a dashboard number is a cost. The handler above already does the valuable part: any score of 6 or below writes a callback flag with the ticket id and score attached, and the customer hears a commitment — "we will call you back within one business day" — before the call ends. That promise converts a detractor moment into a recovery opportunity, and your team's callback queue each morning is the survey's ROI.
Two extensions when you outgrow the basics. For open-ended feedback, follow a low score with the record verb — "after the tone, tell us what went wrong" — and the recording lands on your webhook when it completes. For completion tracking, set an events_url on the survey number and consume call.completed events alongside your response table: answered-but-silent calls tell you the script is losing people, unanswered calls tell you the timing is wrong. And because survey calls are ordinary Sautikit calls, you can pull yesterday's outcomes conversationally — connect the Sautikit MCP server and ask Claude to cross-check list_calls against your response table for completion rate per script variant.
When the callback queue needs real ticketing, agent assignment, and SMS or WhatsApp follow-ups alongside voice, Helloduty — the multi-channel CX platform Sautikit is part of — picks up where the flag in your database leaves off.