←BaldanAll skillsLet's talk
← Dashboards · 04 Customer analytics

Is the home-internet base getting healthier?

Fixed Wireless Access (FWA) is home internet over the mobile network: a router at home with a SIM, on 4G or 5G. This is the whole job of looking after such a base, from the raw tables to the campaign results, and the customer surveys that explain why people leave. All on data made up for this page, so every row can be shown.

4,702 customers22 months, Jan 25 – Oct 267 raw tables4,002 survey answers102,112 rows5 campaigns with control groups
Open the playground

The job, answered

Six parts of the role, six answers from the data.

01Base monitoring3,085 active customers at the end of Oct 26, +151 in the month (277 joined, 126 left). Monthly churn is 4.29%: 73 chose to leave and 53 were lost to non-payment.02Risk segmentation861 active customers sit in the three riskiest segments; in the past, 13.7% of 'overdue or lapsing' customers left within 60 days, against 4.6% of healthy ones. The churn model ranks well on months it never saw (AUC 0.714): calling the top 20% would reach 48.3% of the customers who then left.03Retention executionFive campaigns, each with a random control group. 3 show a proven lift (5G upgrade +14.8 pts, rebuy reminder +5.8 pts, save offer +5.2 pts); payment recovery looks positive but needs more customers to be sure; early-life welcome changed nothing.045G adoption349 customers have 5G coverage at home but are still on 4G. After upgrading, usage goes from 247.5 to 498.5 GB a month.05Lifecycle requestsA relocation fixed within its SLA loses 8.2% of customers in 60 days; a late fix loses 18.9%.06Performance feedback17.2% of new customers leave within 3 months. It rises to 28.2% when first-month usage is under 35% of the plan, and to 23.2% for customers sold through online. Three-month survival moved from 81.4% (the Jan 25 cohort) to 83.2% (Aug 26).+ the whyCustomer voiceDetractors leave 2.3× as often as promoters within 90 days (14.1% vs 6.1%), and customers who never answer leave at 11.5%. Of leavers logged as 'didn't pay', 37.8% said the service itself made them stop. The survey turns "who" into "why": on the decision board it narrows one action, rebuilds another and adds one no dashboard suggested.

01 · The raw data

What the team has, down to the last row.

Seven tables a telco would really hold on an FWA base, customer surveys included. Read how they join, what every column means and how they were made, then browse or download every row.

  1. ml/fwa_generate.pysimulates 22 months of the base→
  2. 7 CSV files102,112 rows, below→
  3. ml/fwa_analyse.pycohorts, survival, segments, model, campaign tests, surveys→
  4. fwa-dashboard.jsonwhat the charts read→
  5. this pagecharts, playground, agent

Seven tables, joined on customer_id

What a telco would hold on a home-internet base: the CRM record, monthly usage from the network, billing, service tickets, campaign logs and customer surveys.

usage_monthlykey: customer_id, monthbillingkey: customer_id, cycle_monthcustomerskey: customer_idservice_requestskey: customer_idcampaign_contactskey: customer_id, campaign_idcampaignskey: campaign_idsurveyskey: customer_id, survey_month

How the data was made

Made up by a script with known rules, so the analysis can be checked against the truth. No real customer is in it.

  1. Who joins. New customers every month for 22 months (Jan 2025 – Oct 2026), more each month, with bumps at Ramadan and year-end. Each gets a region, area type, sales channel, plan, way to pay and maybe a 12-month bundle.
  2. 5G arrives. Coverage reaches homes over time, metros first. The 5G plan launches in July 2025 and needs coverage at home; covered 4G customers can upgrade.
  3. Hidden traits. Every customer has an engagement level that drifts month to month, and an affordability level. Neither is in the data, just like in real life: you only see their effects.
  4. Usage. Data used depends on the plan, engagement and season, starts low in the first month, and jumps after a 5G upgrade.
  5. Paying. On-time rebuys and payments depend on affordability, engagement, the payment method (autodebit is best, retail counters worst) and habit: people who were late tend to be late again.
  6. Leaving. Churn is likelier in the first three months, with low or falling usage, after late payments, on congested 4G in metros, when a relocation or device fault is fixed late, or when a bundle ends without renewal. Bundles and 5G keep people.
  7. Surveys. Customers are asked after their first month, every quarter, after a repair and when they leave. Answers follow the same hidden state that drives churn. Online and agent sales skip the signal check at the address, so more of those homes have weak indoor signal. Engaged customers answer more, so the quiet ones are heard least.
  8. Campaigns. Five campaigns pick customers by a rule and keep a random control group. Their true effects are set here (one of them does nothing), so the measured lifts can be checked.

Data dictionary

Every column of every table: its type and what it means.

Customers · One row per customer (CRM master) · key: customer_id

ColumnTypeMeaning
customer_idtextMade-up ID, the key every other table joins on
activation_datedateFirst day of service; its month is the customer's cohort
region / area_typetextWhere the home is: 8 regions; metro, urban, suburban or rural
sales_channeltextRetail outlet, online, field sales or partner agent
plan_at_activation / plan_currenttextHome 4G 20 Mbps, Home 4G 50 Mbps or Home 5G 100 Mbps
devicetext4G CPE or 5G CPE (the home router)
monthly_fee_idrnumberPrice of the current plan per month
pay_type / payment_methodtextMonthly package to rebuy, or postpaid invoice; e-wallet, VA, autodebit, retail counter
bundle_12mY/NSigned a 12-month device bundle
home_5g_coverage_frommonthWhen 5G reached the home (blank: not yet)
status / churn_datetext, dateActive or churned, and when
churn_type / churn_reasontextVoluntary or involuntary (non-payment), and the reason given

Raw data explorer

The seven CSV files exactly as the analysis reads them. Search, sort by any column, page through, or download.

customers.csv

Loading…

Loading the file…
page 1 / 1

02 · Base monitoring

Size, flows, churn, usage, payments.

The monthly view of the base: who joined and left, why they left, how much they use and how they pay.

Active customers

3,085

+151 vs Sep 26

Joined / left

277 / 126

net +151

Monthly churn

4.29%

Sep 26: 4.09%

Voluntary / non-payment

73 / 53

of this month's leavers

ARPU

Rp 252k

billed per active customer

Data per customer

290.3 GB

13.4% use under half the plan

Rebought on time

83.4%

2.8% did not rebuy

On a 5G plan

31.6%

1,323 homes have 5G

Base and flows

Active customers at month end, and who joined and left each month

0100200300400Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26
Active base (right scale ÷10)JoinedLeft, by choiceLeft, non-payment

Read it:The base grows every month because joiners outnumber leavers, but leavers grow too: 126 left in Oct 26 against 58 a year earlier. Bars are the base divided by ten, to share the scale.

Churn rate

Leavers ÷ customers at the start of the month, split by kind

0.0%2.5%5.0%7.5%10.0%Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26
TotalBy choiceNon-payment

Read it:Early months are noisy because the base is young; from mid-2025 churn settles around 4% a month, and almost half of it is customers who simply stop paying or rebuying.

Payment health

Share of packages rebought on time, and of invoices paid by the due date

0%25%50%75%100%Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26
Package rebought on timeInvoice paid on timePackage not rebought

Usage

Average data per active customer, and the share using less than half their plan's typical

0200400600Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26
GB per customerLow users (% × 10)

Leaving in the first 3 months

17.2% of new customers overall, by first-month usage

Under 35% of plan28.2%n=962
35-60% of plan17.0%n=1488
60-90% of plan12.5%n=1011
90%+ of plan9.6%n=705

Read it:The first month tells you a lot: customers who barely use the service in month one are the likeliest to be gone by month three.

…by sales channel

Same early churn, by where the customer was sold

Retail outlet16.0%n=1601
Online23.2%n=908
Field sales12.7%n=1041
Partner agent19.3%n=616

Read it:Online and partner sales bring in customers who leave sooner; field sales the most loyal ones. A welcome check matters most for the first two.

Why people left

Reason recorded at churn, all 22 months

Did not rebuy705
Speed or coverage198
Price177
Low use, not needed141
Other125
Switched to fibre123
Payment default90
Device problem32
Moved house26

03 · Cohorts & survival

Does each new group of customers stay longer?

Customers grouped by the month they joined, followed month by month. Filter by plan, channel, way to pay, area or region.

4,702 customers

Retention by cohort

Share of each month's new customers still active after N months (0 = end of the first month)

Cohortn0123456789101112131415161718192021
Jan 2514594868179747372706764595959575452515049484847
Feb 25138939086807472696462595757515148454443414038
Mar 252049488838177747371696864626159565554525149
Apr 2517290837776747169656362595855534948474644
May 25161918884817978736765636260555350504948
Jun 251869387827876706360595856555452494746
Jul 2517392878482777472706866646058585755
Aug 25188908580787472707068666463615958
Sep 252049489838176737270696765636059
Oct 2518793898785827976757272706968
Nov 25222918582797774726664625957
Dec 252329286838077757269676563
Jan 2620789858278757369696663
Feb 26249918884787672696765
Mar 262769387847975726765
Apr 2622587817772706866
May 26222928986838279
Jun 262539288848379
Jul 2624291878379
Aug 26280938783
Sep 262599388
Oct 2627791

Read it:Read across a row to follow one cohort as it ages, and down a column to compare cohorts at the same age. The biggest drop is always in the first three columns: that is where retention work pays off most.

Survival curves

Kaplan–Meier: share still a customer after each month of service, using every customer's tenure

30%50%70%90%110%1471013161922
Retail outletOnlineField salesPartner agent

x axis: months of service. Curves stop when fewer than 20 customers are left to follow.

Cohort survivability

Share of each cohort still active after 3, 6 and 12 months: are newer cohorts doing better?

40%60%80%100%Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Aug 26
After 3 monthsAfter 6 monthsAfter 12 months

Read it:Three-month survival moved from 81.4% (the Jan 25 cohort) to 83.2% (Aug 26). The welcome campaign started in Sep 2025; if it worked, the 3-month line would climb after that. It barely does, which the campaign results below confirm.

04 · Risk segments & machine learning

Who needs attention, and in what order.

Rule-based segments the team can act on, and a churn model that ranks every active customer by their chance of leaving in the next two months.

Risk segments, today

Every active customer falls in the first segment whose rule fits. The last column checks each rule on the past: of the customers who fitted it at the end of Jun 26, how many left within 60 days?

SegmentRuleActive nowFees / monthLeft within 60 days (past check)
Overdue or lapsingLate last month, late twice in the last 3 months, or still owing487Rp 123.7m
13.7%
New, low useFirst 3 months of service and using under 60% of the plan's typical data187Rp 46.4m
10.9%
Low or falling useUsing under 45% of the plan's typical data, or down 40% from the two months before187Rp 48.8m
9.8%
Service issueRelocation or device fault in the last 2 months, or a 4G line in a congested metro363Rp 79.3m
5.8%
5G-ready on 4GHome now has 5G coverage but the line is still 4G: an upgrade opportunity118Rp 25.3m
4.3%
HealthyNone of the above1,743Rp 462.1m
4.6%

Read it:861 active customers sit in the three riskiest segments; in the past, 13.7% of 'overdue or lapsing' customers left within 60 days, against 4.6% of healthy ones.

Churn model, tested on the future

Leaves in the next 2 months. Trained on Jun 25 – Apr 26 month-ends (16,579 rows), tested on May 26 – Aug 26 month-ends (10,247 rows) it never saw.

Logistic regression

AUC 0.714

exported, runs in your browser

Gradient boosting

AUC 0.72

for comparison

Base rate

6.8%

leave within 2 months

Top 20% catches

48.3%

of leavers, 16.5% hit rate

Read it:The simple, explainable model ranks about as well as the boosted one, so it is the one used: every score can be explained by its drivers.

How much of churn you catch

Call the riskiest X% (by model score): what share of the customers who then left were in that group?

0%25%50%75%100%020406080100
LogisticBoostingRandom calls

x axis: % of the base called, riskiest first

What drives the score

Model weights on standardised features: right pushes risk up, left pulls it down

Late payments, last 3 months
Days past due
Sold online
Paid late last month
In the first 3 months
Lives in a metro area
Pays at a retail counter
Relocation request, last 2 months
Lives in a rural area
On Home 4G 50 Mbps
Inside a 12-month bundle
Days online last month
On Home 5G 100 Mbps
Data used vs plan's typical
Pays by postpaid invoice

Who to call this week

The 25 active customers with the highest model risk today, with the two features pushing their score up most (made-up IDs)

CustomerPlanMonthsSegmentRisk, 60 daysWhy
FWA-0027064G 20 Mbps7Overdue or lapsing52.3%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0014024G 20 Mbps13Overdue or lapsing51.2%Late payments, last 3 months · Days past due
FWA-0036384G 20 Mbps4Overdue or lapsing48.9%Days past due · Late payments, last 3 months
FWA-0020495G 100 Mbps10Overdue or lapsing48.4%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0041644G 20 Mbps2Overdue or lapsing47.9%Late payments, last 3 months · Days past due
FWA-0021954G 50 Mbps10Overdue or lapsing47.9%Late payments, last 3 months · Relocation request, last 2 months
FWA-0031974G 20 Mbps5Overdue or lapsing46.5%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0031554G 20 Mbps6Overdue or lapsing43.2%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0035054G 50 Mbps4Overdue or lapsing41.0%Days past due · Late payments, last 3 months
FWA-0038115G 100 Mbps3Overdue or lapsing39.3%Days past due · Late payments, last 3 months
FWA-0035864G 20 Mbps4Overdue or lapsing39.1%Late payments, last 3 months · Days past due
FWA-0019354G 20 Mbps11Overdue or lapsing38.2%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0024644G 50 Mbps8Overdue or lapsing37.7%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0034554G 50 Mbps4Overdue or lapsing37.5%Days past due · Late payments, last 3 months
FWA-0034854G 50 Mbps4Overdue or lapsing37.3%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0021554G 50 Mbps10Overdue or lapsing37.3%Late payments, last 3 months · Days online last month (low)
FWA-0033554G 20 Mbps5Overdue or lapsing36.1%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0040854G 20 Mbps2Overdue or lapsing35.8%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0033574G 20 Mbps5Overdue or lapsing35.3%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0013644G 20 Mbps14Overdue or lapsing35.1%Days past due · Late payments, last 3 months
FWA-0039234G 20 Mbps2New, low use34.7%Days online last month (low) · Data used vs plan's typical (low)
FWA-0024805G 100 Mbps8Overdue or lapsing34.7%Days past due · Late payments, last 3 months
FWA-0040024G 20 Mbps2Overdue or lapsing34.6%Late payments, last 3 months · Pays by postpaid invoice: no
FWA-0047014G 50 Mbps0Overdue or lapsing34.2%Data used vs plan's typical (low) · Pays by postpaid invoice: no
FWA-0035394G 20 Mbps4Overdue or lapsing34.1%Days past due · Late payments, last 3 months

05 · Payment recovery

Late rebuys and unpaid invoices.

How on time customers rebuy or pay, by method, how fast late payers come back, and what is owed today.

How late packages get rebought

Every monthly-package cycle, by days after expiry

On time80.1%
1-3 days6.2%
4-7 days6.4%
8-14 days3.1%
15-30 days0.7%
Not rebought3.5%
Pays withOn timeNever rebought
E-wallet84.5%2.6%
Bank transfer (VA)80.4%3.6%
Retail counter72.5%5.0%

Invoices recovered over time

Postpaid: share of invoices paid by each day after the due date (day 0 = on time)

70%80%90%100%0612182430364245
AutodebitBank transfer (VA)E-wallet

x axis: days after the due date

Read it:Autodebit customers are almost always on time. Everyone else recovers mostly in the first two weeks; what isn't paid by day 30 rarely comes back. Owing right now: 76 invoices, Rp 18.9m.

What the nudges did

Measured against a random control group that got nothing

Rebuy reminder

Rebought on time that month

76.5%

contacted (n=1794)

70.8%

control (n=462)

+5.8 pts

significant

Payment recovery

Paid that month's invoice

80.3%

contacted (n=239)

70.0%

control (n=50)

+10.3 pts

not proven (p=0.1055)

Read it:The reminder works and costs almost nothing. The recovery call looks strong but its control group is tiny, so the result could still be luck: hold out more people next time.

06 · 5G adoption

Moving covered homes onto 5G.

Where 5G reaches, who has moved, what upgrading does to usage and churn, and who is ready now.

5G-ready, still on 4G

349

covered homes to upgrade

Covered homes on 5G

73.6%

974 of 1,323

Data before → after upgrade

248 → 499 GB

314 upgrades measured

Monthly churn, covered 4G

4.32%

Monthly churn, 5G

2.52%

Coverage and adoption

Active customers whose home has 5G, how many are on a 5G plan, and upgrades each month

05001,0001,500Jul 25Sep 25Nov 25Jan 26Mar 26May 26Jul 26Sep 26Oct 26
Upgrades in the monthHomes with 5GOn a 5G plan

The upgrade campaign

Free 5G device swap with a 5G plan. Who: 4G customers whose home now has 5G coverage, not contacted in the last 3 months, up to 150 per month.

Contacted25.4%
Control10.6%

+14.8 points more moved to 5G within two months (95% range +10.5 to +19.1), about 174 extra upgrades from 1,175 contacts.

Read it:349 customers have 5G coverage at home but are still on 4G. After upgrading, usage goes from 247.5 to 498.5 GB a month. Upgraded customers also leave less, so the campaign pays twice: more data and a stickier base.

07 · Lifecycle requests

Renewals, relocations and device faults.

Every request, how fast it was handled against its target, and what a slow fix costs in customers.

Relocation

233 requests · target 7 days

Median days

6

Within target

59%

Left within 60 days: 8.2% if fixed in time, 18.9% if late.

Device fault

313 requests · target 3 days

Median days

2

Within target

60%

Left within 60 days: 10.2% if fixed in time, 20.8% if late.

Renewal

478 requests · target 5 days

Median days

2

Within target

100%

59.2% renewed. Left within 60 days: 2.5% after renewing, 16.4% after declining.

5G upgrade

464 requests · target 3 days

Median days

1

Within target

100%

Done in a day; the bottleneck is reaching eligible customers, not the swap.

Requests per month

Every lifecycle request, by type

0204060Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26Oct 26
RelocationDevice faultRenewal5G upgrade

Read it:A relocation fixed within its SLA loses 8.2% of customers in 60 days; a late fix loses 18.9%. Today 42% of relocations and 40% of device faults miss their target: fixing that is one of the cheapest ways to keep customers.

08 · Campaign results

What actually worked, against a control group.

Each campaign kept a random group that got nothing. The lift is the difference between the two, with its 95% range and cost per result.

Campaign scoreboard

Every campaign kept a random control group. The bar is the measured lift with its 95% range: if the range crosses zero, the effect isn't proven.

CampaignMeasured outcomeContactedControlLift (points) and 95% rangeExtraCostPer extra

Early-life welcome

Setup tips and a usage check call

Still a customer 90 days later85.0%n=1,81984.8%n=461

+0.2 [-3.4, +3.9] · p=0.9012

+4Rp 10.9m–

Rebuy reminder

Reminder 3 days before the package expires

Rebought on time that month76.5%n=1,79470.8%n=462

+5.8 [+1.2, +10.3] · p=0.0105

+104Rp 718kRp 7k

Payment recovery

Reminder call and a payment link

Paid that month's invoice80.3%n=23970.0%n=50

+10.3 [-3.3, +24.0] · p=0.1055

+25Rp 1.2m–

Save offer

20% off for 3 months

Still a customer 90 days later87.6%n=66082.3%n=317

+5.2 [+0.3, +10.1] · p=0.0276

+34Rp 90.6mRp 2.6m

5G upgrade

Free 5G device swap with a 5G plan

Moved to 5G within 2 months25.4%n=1,17510.6%n=293

+14.8 [+10.5, +19.1] · p=0

+174Rp 3.5mRp 20k

Read it:Keep the 5G upgrade and the rebuy reminder: clear, cheap wins. The save offer works, but see who it works for below. The recovery call needs a bigger control group to prove itself, and the early-life welcome shows no effect: redesign it or stop it.

Who the save offer actually helps

Same campaign, split by how the customer had been paying when contacted

Paid on time

+14.7 pts

stayed: 93.2% vs 78.4% · n=117/51 · significant

Late once

+4.7 pts

stayed: 89.7% vs 85.1% · n=419/194 · not proven

Late 2+ times

-2.8 pts

stayed: 75.0% vs 77.8% · n=124/72 · not proven

Read it:A discount keeps people who were happy to pay but tempted to leave. For customers who keep paying late, money is the problem and the discount doesn't fix it: send them to payment recovery instead.

09 · Customer voice · surveys

Why do they really leave? Ask them.

The numbers above say what happened and to whom. Surveys say why: after the first month, every quarter, after a repair and when a customer leaves. Here is what they add, and how they change the decisions.

Why the survey matters

Behaviour shows what customers do. Only asking them shows why they leave.

Usage and billing can say a customer stopped paying; they can't say whether it was money, a slow evening connection or a router that never got signal by the window. Fix the wrong cause and the money is wasted. The survey closes that gap, and it checks every other conclusion on this page.

  1. 01 · behaviourWhat happenedusage, billing, requests
  2. 02 · modelWho is nextthe churn model
  3. 03 · surveyWhywhat customers tell us
  4. 04 · experimentWhat workscampaigns with control groups

Decisions are safest where all four point the same way. The board at the end of this section shows where they do.

Surveys sent

14,834

4 moments in the customer's life

Answered

4,002

27.0% answer rate

Comments coded

2,220

into 11 themes

NPS, last 6 months

+38

1,524 answers

Exit surveys answered

414

why leavers say they left

Detractors vs promoters

2.3×

as likely to leave in 90 days

Onboarding

3,838 sent · 1,281 answered (33%)

Relationship

8,880 sent · 2,124 answered (24%)

After service request

504 sent · 183 answered (36%)

Exit

1,612 sent · 414 answered (26%)

What the system logged vs what leavers told us

Exit survey, 414 leavers who answered. Each bar is one reason in the CRM; the colours are what those customers said themselves.

Did not rebuy n=1712842
Speed or coverage n=552951
Price n=481558
Switched to fibre n=382453
Other n=36332514
Low use, not needed n=311932131323
Payment default n=17242929
Device problem n=1080
Moved house n=813131363
Slow at busy hoursWeak signal indoorsRouter or device problemsToo expensive for meForgot to top up, then gave upMoved house, couldn't take it alongSwitched to fibreDidn't use it enough

Red and orange: the service itself. Violet: money. Blue, teal, grey: life changes and competition.

Read it:The CRM says 188 of these leavers simply didn't pay. Asked, 38% of them blamed the service (slow evenings, weak signal, the router), not money (52%). "Low use, not needed" hides the same thing: 52% of those 31 pointed at the service. A payment reminder can't fix a connection.

Unhappy customers leave; silent ones too

Share who left within 90 days of a regular survey, by how they answered (surveys with 90 days observed)

Promoters6.1%n=1,352
Passives7.8%n=705
Detractors14.1%n=524
Didn't answer11.5%n=7,008

Read it:Detractors leave 2.3× as often as promoters (14.1% vs 6.1%). And the 7,008 times nobody answered, 11.5% left: nearer the detractors than the promoters. Silence is a warning sign, so the survey never replaces behaviour; the two work together.

After a repair

Survey sent when a relocation or device fault is closed

Fixed within target

4.4 / 5

NPS 30 · n=108

Fixed late

2.1 / 5

NPS -21 · n=75

Read it:Same customers, same kind of problem: only the speed of the fix differs, and the support rating halves. It agrees with the churn numbers in the lifecycle section.

What drives the score, and how we do on it

Across: how strongly a rating moves with the recommend score (correlation). Up: share satisfied (4-5 out of 5).

30%40%50%60%70%0.00.20.40.60.8Fix firstKeep it upWatchSpeed at busy hoursr 0.54 · 52% satisfiedSignal indoorsr 0.71 · 61% satisfiedValue for moneyr 0.21 · 42% satisfiedSupport and repairsr 0.32 · 57% satisfied

Speed at busy hours

left in 90 days: 11.0% if rated 1-2, 7.2% if 4-5

Signal indoors

left in 90 days: 12.2% if rated 1-2, 6.4% if 4-5

Value for money

left in 90 days: 12.1% if rated 1-2, 5.8% if 4-5

Read it:Signal indoors and speed at busy hours move the score most, and speed sits in "fix first". Price barely moves the score (r 0.21), though it has the fewest satisfied (42%). Yet people who rate value 1-2 leave 2.1× as often, and 66% of those leavers simply stop paying (38% for everyone else): price is a payment problem, not a satisfaction problem. That is why the discount only works for customers who pay on time.

Where it hurts

Share rating 1-2 out of 5, regular surveys. The recommend score (NPS) next to each group.

Unhappy with speed at busy hours, by line

4G in a metro55%NPS 6
4G elsewhere10%NPS 34
5G1%NPS 60

Unhappy with signal indoors, by how they were sold

Retail outlet11%NPS 35
Online17%NPS 20
Field sales9%NPS 45
Partner agent20%NPS 25

Read it:55% of metro 4G customers are unhappy with evening speed, against 1% on 5G: congestion, which a 5G move fixes. Online and agent sales leave 17% and 20% unhappy with indoor signal, against 9% for field sales, who check the signal at the address before selling. That one step is missing.

In their own words

2,220 comments, coded into themes. Bars: detractors' comments by theme. Click a theme to read real lines (as written, with a translation).

“Di kamar belakang tidak dapat sinyal sama sekali. Harusnya bisa lebih baik.”
No signal at all in the back room. It should be better.

Onboarding · NPS 4 · 4G 20 Mbps · Suburban · Retail outlet

“Di lantai dua lumayan, di bawah hampir tidak bisa”
Upstairs is OK, downstairs barely works

Relationship · NPS 4 · 4G 20 Mbps · Metro · Online

“Sinyal naik turun, sering putus sendiri. Kecewa sih.”
Signal goes up and down and keeps dropping. Pretty disappointing.

Onboarding · NPS 4 · 4G 20 Mbps · Urban · Partner agent

“Sinyal di dalam rumah lemah, modem harus ditaruh dekat jendela. Kalau begini terus mau berhenti.”
Weak signal inside, the router has to sit by the window. If it stays like this I'll quit.

Relationship · NPS 5 · 4G 20 Mbps · Rural · Retail outlet

Read it:Comments carry what a score can't: "nobody checked the signal at my house" points straight at the sales process. In a real pipeline the coding is done by keyword rules or an LLM, then spot-checked by a person; the raw comments stay in the surveys table for anyone to re-read.

Why the welcome call didn't work

The campaign results say it did nothing. The onboarding survey says why.

What the call offered

Setup tips and a usage check call

Measured effect: +0.2 pts (95% range -3.4 to +3.9). No proven effect.

What unhappy new customers wrote

Weak signal indoors40%
Slow at busy hours37%
Price21%
Router problems1%

Read it:New customers who are unhappy after month one write about weak indoor signal (40%) and slow evenings, not about setting up. Tips on a call can't move a router to where the signal is. Redesign it as a signal check with router placement (or an outdoor antenna), then test it against a control group again.

Does a survey answer sharpen the model?

2,821 survey answers linked to the same month's behaviour; who left in the next 2 months (157). Cross-validated, 5 folds × 10.

Ranking quality (AUC)

Behaviour only0.666
Survey answers only0.612
Both together0.664

Leavers caught in the riskiest 20%

Behaviour only45%
Survey answers only34%
Both together45%

Read it:Four questions alone rank customers well above chance (AUC 0.61; a coin flip is 0.5), but months of behaviour rank better (0.67), and adding the answers changes almost nothing. An honest result: the survey isn't the tool for finding who will leave; behaviour already does that. Its value is the why, which changes what we do for them.

Decision board: three kinds of evidence, one call each

✓ supports it · ✗ argues against · – mixed or not tested. The survey column is where several decisions changed.

  • Move covered 4G homes to 5G

    Go: all three agree
    Behaviour
    Covered 4G lines churn 4.3% a month, 5G 2.5%
    Survey
    55% of metro 4G unhappy with evening speed; NPS 60 on 5G
    Experiment
    Upgrade campaign +14.8 pts against its control group
  • Fix relocations and faults within target

    Go, and track it weekly
    Behaviour
    A late relocation fix loses 18.9% in 60 days vs 8.2%
    Survey
    Support rated 2.1/5 after a late fix, 4.4 on time
    Experiment
    Not an experiment: a service standard to track
  • Signal check before selling online and via agents

    Pilot it with a control group
    Behaviour
    23.2% of online sales leave within 3 months vs 12.7% field sales
    Survey
    17% of online customers unhappy with indoor signal vs 9%
    Experiment
    Not tested yet
  • Save offer (20% off) for everyone at risk

    Narrow it to on-time payers
    Behaviour
    Risk segments leave 2-3× as often as healthy customers
    Survey
    Price barely moves the score (r 0.21); low value ratings end in non-payment
    Experiment
    On-time payers +14.7 pts; late 2+ times -2.8 pts
  • Welcome call with setup tips

    Redesign around signal, re-test
    Behaviour
    17.2% of new customers leave within 3 months
    Survey
    Unhappy new customers write about signal (40%) and evenings, not setup
    Experiment
    +0.2 pts: no proven effect
  • Rebuy reminders before the package expires

    Keep: cheap and it works
    Behaviour
    Late rebuys repeat: habit drives the next one
    Survey
    38% of 'didn't pay' leavers blamed the service, which a reminder can't fix
    Experiment
    +5.8 pts on-time rebuys, at Rp 400 a message

Read it:Without the survey, the discount would go to everyone at risk and the welcome call would just be dropped. With it, the discount is narrowed, the welcome call is rebuilt around the real problem, and a new action appears that no dashboard suggested: checking the signal before an online sale.

Reading surveys with care

Who answers

About 1 in 4 to 1 in 3 answer, and engaged customers answer most. The quiet ones, who leave most, are heard least: always read answers next to behaviour.

Small groups

Some splits rest on few answers (e.g. 31 'low use' leavers). The page shows n next to every split.

Said is not always why

People give the easiest reason. Exit answers are checked against what usage, billing and tickets show.

Ask at the right moments

After month one, every quarter, after a repair, at exit: when the experience is fresh and a fix is still possible.

10 · Playground

Try it yourself.

Cut your own segment from the live base, read every survey answer, score a customer you build, and plan next month's campaign with the measured lifts.

Segment builder

Set the rules; the segment is cut live from today's 3,000-odd active customers, with each one's model risk.

Customers

597

19.4% of active

Fees per month

Rp 152.7m

Average risk

10.4%

base 6.6%

Expected to leave

62

within 60 days, by the model

By plan

Home 4G 20 Mbps208
Home 4G 50 Mbps199
Home 5G 100 Mbps190

By channel

Retail outlet232
Online132
Field sales134
Partner agent99

The same segment as a query on the raw tables

SELECT c.customer_id, c.plan_current, c.sales_channel
FROM customers c
JOIN usage_monthly u ON u.customer_id = c.customer_id AND u.month = '2026-10'
JOIN (SELECT customer_id, COUNT(*) AS late_3m FROM billing
      WHERE cycle_month >= '2026-08' AND status IN ('Late','Not rebought','Paid late','Unpaid')
      GROUP BY customer_id) b ON b.customer_id = c.customer_id
WHERE c.status = 'Active'
  AND u.data_gb < 0.6 * typical_gb(c.plan_current)   -- usage under 60% of the plan

Survey explorer

Every survey answer from the raw file, cut any way you like. Pick a theme to read what those customers wrote.

Loading the survey file…

Model what-if

Build one customer and watch the exported model (the same weights as scikit-learn's) score them.

Chance of leaving in the next 2 months

35.6%

average customer: 6.8% · this one is 5.2× the average

What moves this score (log-odds contribution)

Pays by postpaid invoice: no+0.41
Late payments, last 3 months+0.37
Data used vs plan's typical (low)+0.33
Paid late last month+0.23
Days online last month (low)+0.22
Sold online+0.20
Pays at a retail counter+0.13
In the first 3 months+0.10

Read it:Try a new online customer with low usage who was late last month, then give them a 12-month bundle or move them to 5G: the drivers that push risk up are exactly what the campaigns target.

Campaign planner

Pick a campaign, how many people the team can contact this month and who to pick. The numbers use the lifts measured against control groups above, not guesses.

Who gets the offer

Contacted

150

Result

+5.3

customers kept

Cost

Rp 15.9m

contacts + 3 months of discount

Cost per result

Rp 3.0m

value Rp 13.5m

Without the offer, about 40.7 of these 150 would leave within 60 days.

Read it:Each picked customer's own model risk × the offer's measured effect for their payment band: 69% fewer leavers if they pay on time, 31% if late once, none if late twice or more (84 of this pick). Value = a year of their fee. Compare the three ways to pick: the riskiest customers are mostly late payers, whom a discount doesn't move, so the riskiest on-time payers keep the most customers per rupiah.

11 · Agentic AI

Where an AI agent fits in this job.

The weekly loop of monitor, segment, act and measure is a good fit for an agent with read-only tools and a person approving every action. Step through one run.

An agent that runs the retention loop

A simulation on this page's data: the numbers are real, the reasoning is scripted, and nothing is sent. The agent's tools only read data and draft; a person approves offers, budgets and wording.

  1. agentCheck the base

    Every Monday the agent starts by reading the base-monitoring numbers, like an analyst would.

    monitor_base(as_of='2026-10-31')
  2. tool resultBase health

    3,085 active, churn 4.29% (73 by choice, 53 non-payment), 83.4% rebuys on time. Churn is above the 4% target.

  3. agentFind who needs attention

    Churn is above target, so it pulls the risk segments and scores the active base with the churn model.

    segments(); score_customers(model='fwa-churn-lr')

Where AI helps

Reading the numbers every week, spotting who needs what, drafting personal messages at scale, and remembering what worked.

Where people stay

Offers, budgets, wording, and anything that touches a customer's money or contract.

Guardrails

Read-only data tools, approved templates, a control group on every action, and a log of every step.

About the data. Every customer, payment and campaign here was generated by a script with fixed rules and a fixed random seed, so the numbers are reproducible and no real person is in them. Plans, prices and channels are typical of the Indonesian market, not any one operator.