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.
The job, answered
Six parts of the role, six answers from the data.
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.
- ml/fwa_generate.pysimulates 22 months of the base→
- 7 CSV files102,112 rows, below→
- ml/fwa_analyse.pycohorts, survival, segments, model, campaign tests, surveys→
- fwa-dashboard.jsonwhat the charts read→
- 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.
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.
- 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.
- 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.
- 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.
- Usage. Data used depends on the plan, engagement and season, starts low in the first month, and jumps after a 5G upgrade.
- 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.
- 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.
- 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.
- 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
| Column | Type | Meaning |
|---|---|---|
| customer_id | text | Made-up ID, the key every other table joins on |
| activation_date | date | First day of service; its month is the customer's cohort |
| region / area_type | text | Where the home is: 8 regions; metro, urban, suburban or rural |
| sales_channel | text | Retail outlet, online, field sales or partner agent |
| plan_at_activation / plan_current | text | Home 4G 20 Mbps, Home 4G 50 Mbps or Home 5G 100 Mbps |
| device | text | 4G CPE or 5G CPE (the home router) |
| monthly_fee_idr | number | Price of the current plan per month |
| pay_type / payment_method | text | Monthly package to rebuy, or postpaid invoice; e-wallet, VA, autodebit, retail counter |
| bundle_12m | Y/N | Signed a 12-month device bundle |
| home_5g_coverage_from | month | When 5G reached the home (blank: not yet) |
| status / churn_date | text, date | Active or churned, and when |
| churn_type / churn_reason | text | Voluntary 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.
Loading…
| Loading the file… |
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
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
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
Usage
Average data per active customer, and the share using less than half their plan's typical
Leaving in the first 3 months
17.2% of new customers overall, by first-month usage
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
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
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)
| Cohort | n | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Jan 25 | 145 | 94 | 86 | 81 | 79 | 74 | 73 | 72 | 70 | 67 | 64 | 59 | 59 | 59 | 57 | 54 | 52 | 51 | 50 | 49 | 48 | 48 | 47 |
| Feb 25 | 138 | 93 | 90 | 86 | 80 | 74 | 72 | 69 | 64 | 62 | 59 | 57 | 57 | 51 | 51 | 48 | 45 | 44 | 43 | 41 | 40 | 38 | |
| Mar 25 | 204 | 94 | 88 | 83 | 81 | 77 | 74 | 73 | 71 | 69 | 68 | 64 | 62 | 61 | 59 | 56 | 55 | 54 | 52 | 51 | 49 | ||
| Apr 25 | 172 | 90 | 83 | 77 | 76 | 74 | 71 | 69 | 65 | 63 | 62 | 59 | 58 | 55 | 53 | 49 | 48 | 47 | 46 | 44 | |||
| May 25 | 161 | 91 | 88 | 84 | 81 | 79 | 78 | 73 | 67 | 65 | 63 | 62 | 60 | 55 | 53 | 50 | 50 | 49 | 48 | ||||
| Jun 25 | 186 | 93 | 87 | 82 | 78 | 76 | 70 | 63 | 60 | 59 | 58 | 56 | 55 | 54 | 52 | 49 | 47 | 46 | |||||
| Jul 25 | 173 | 92 | 87 | 84 | 82 | 77 | 74 | 72 | 70 | 68 | 66 | 64 | 60 | 58 | 58 | 57 | 55 | ||||||
| Aug 25 | 188 | 90 | 85 | 80 | 78 | 74 | 72 | 70 | 70 | 68 | 66 | 64 | 63 | 61 | 59 | 58 | |||||||
| Sep 25 | 204 | 94 | 89 | 83 | 81 | 76 | 73 | 72 | 70 | 69 | 67 | 65 | 63 | 60 | 59 | ||||||||
| Oct 25 | 187 | 93 | 89 | 87 | 85 | 82 | 79 | 76 | 75 | 72 | 72 | 70 | 69 | 68 | |||||||||
| Nov 25 | 222 | 91 | 85 | 82 | 79 | 77 | 74 | 72 | 66 | 64 | 62 | 59 | 57 | ||||||||||
| Dec 25 | 232 | 92 | 86 | 83 | 80 | 77 | 75 | 72 | 69 | 67 | 65 | 63 | |||||||||||
| Jan 26 | 207 | 89 | 85 | 82 | 78 | 75 | 73 | 69 | 69 | 66 | 63 | ||||||||||||
| Feb 26 | 249 | 91 | 88 | 84 | 78 | 76 | 72 | 69 | 67 | 65 | |||||||||||||
| Mar 26 | 276 | 93 | 87 | 84 | 79 | 75 | 72 | 67 | 65 | ||||||||||||||
| Apr 26 | 225 | 87 | 81 | 77 | 72 | 70 | 68 | 66 | |||||||||||||||
| May 26 | 222 | 92 | 89 | 86 | 83 | 82 | 79 | ||||||||||||||||
| Jun 26 | 253 | 92 | 88 | 84 | 83 | 79 | |||||||||||||||||
| Jul 26 | 242 | 91 | 87 | 83 | 79 | ||||||||||||||||||
| Aug 26 | 280 | 93 | 87 | 83 | |||||||||||||||||||
| Sep 26 | 259 | 93 | 88 | ||||||||||||||||||||
| Oct 26 | 277 | 91 |
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
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?
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?
| Segment | Rule | Active now | Fees / month | Left within 60 days (past check) |
|---|---|---|---|---|
| Overdue or lapsing | Late last month, late twice in the last 3 months, or still owing | 487 | Rp 123.7m | 13.7% |
| New, low use | First 3 months of service and using under 60% of the plan's typical data | 187 | Rp 46.4m | 10.9% |
| Low or falling use | Using under 45% of the plan's typical data, or down 40% from the two months before | 187 | Rp 48.8m | 9.8% |
| Service issue | Relocation or device fault in the last 2 months, or a 4G line in a congested metro | 363 | Rp 79.3m | 5.8% |
| 5G-ready on 4G | Home now has 5G coverage but the line is still 4G: an upgrade opportunity | 118 | Rp 25.3m | 4.3% |
| Healthy | None of the above | 1,743 | Rp 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?
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
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)
| Customer | Plan | Months | Segment | Risk, 60 days | Why |
|---|---|---|---|---|---|
| FWA-002706 | 4G 20 Mbps | 7 | Overdue or lapsing | 52.3% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-001402 | 4G 20 Mbps | 13 | Overdue or lapsing | 51.2% | Late payments, last 3 months · Days past due |
| FWA-003638 | 4G 20 Mbps | 4 | Overdue or lapsing | 48.9% | Days past due · Late payments, last 3 months |
| FWA-002049 | 5G 100 Mbps | 10 | Overdue or lapsing | 48.4% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-004164 | 4G 20 Mbps | 2 | Overdue or lapsing | 47.9% | Late payments, last 3 months · Days past due |
| FWA-002195 | 4G 50 Mbps | 10 | Overdue or lapsing | 47.9% | Late payments, last 3 months · Relocation request, last 2 months |
| FWA-003197 | 4G 20 Mbps | 5 | Overdue or lapsing | 46.5% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-003155 | 4G 20 Mbps | 6 | Overdue or lapsing | 43.2% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-003505 | 4G 50 Mbps | 4 | Overdue or lapsing | 41.0% | Days past due · Late payments, last 3 months |
| FWA-003811 | 5G 100 Mbps | 3 | Overdue or lapsing | 39.3% | Days past due · Late payments, last 3 months |
| FWA-003586 | 4G 20 Mbps | 4 | Overdue or lapsing | 39.1% | Late payments, last 3 months · Days past due |
| FWA-001935 | 4G 20 Mbps | 11 | Overdue or lapsing | 38.2% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-002464 | 4G 50 Mbps | 8 | Overdue or lapsing | 37.7% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-003455 | 4G 50 Mbps | 4 | Overdue or lapsing | 37.5% | Days past due · Late payments, last 3 months |
| FWA-003485 | 4G 50 Mbps | 4 | Overdue or lapsing | 37.3% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-002155 | 4G 50 Mbps | 10 | Overdue or lapsing | 37.3% | Late payments, last 3 months · Days online last month (low) |
| FWA-003355 | 4G 20 Mbps | 5 | Overdue or lapsing | 36.1% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-004085 | 4G 20 Mbps | 2 | Overdue or lapsing | 35.8% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-003357 | 4G 20 Mbps | 5 | Overdue or lapsing | 35.3% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-001364 | 4G 20 Mbps | 14 | Overdue or lapsing | 35.1% | Days past due · Late payments, last 3 months |
| FWA-003923 | 4G 20 Mbps | 2 | New, low use | 34.7% | Days online last month (low) · Data used vs plan's typical (low) |
| FWA-002480 | 5G 100 Mbps | 8 | Overdue or lapsing | 34.7% | Days past due · Late payments, last 3 months |
| FWA-004002 | 4G 20 Mbps | 2 | Overdue or lapsing | 34.6% | Late payments, last 3 months · Pays by postpaid invoice: no |
| FWA-004701 | 4G 50 Mbps | 0 | Overdue or lapsing | 34.2% | Data used vs plan's typical (low) · Pays by postpaid invoice: no |
| FWA-003539 | 4G 20 Mbps | 4 | Overdue or lapsing | 34.1% | 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
| Pays with | On time | Never rebought |
|---|---|---|
| E-wallet | 84.5% | 2.6% |
| Bank transfer (VA) | 80.4% | 3.6% |
| Retail counter | 72.5% | 5.0% |
Invoices recovered over time
Postpaid: share of invoices paid by each day after the due date (day 0 = on time)
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
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.
+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
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.
| Campaign | Measured outcome | Contacted | Control | Lift (points) and 95% range | Extra | Cost | Per extra |
|---|---|---|---|---|---|---|---|
Early-life welcome Setup tips and a usage check call | Still a customer 90 days later | 85.0%n=1,819 | 84.8%n=461 | +0.2 [-3.4, +3.9] · p=0.9012 | +4 | Rp 10.9m | – |
Rebuy reminder Reminder 3 days before the package expires | Rebought on time that month | 76.5%n=1,794 | 70.8%n=462 | +5.8 [+1.2, +10.3] · p=0.0105 | +104 | Rp 718k | Rp 7k |
Payment recovery Reminder call and a payment link | Paid that month's invoice | 80.3%n=239 | 70.0%n=50 | +10.3 [-3.3, +24.0] · p=0.1055 | +25 | Rp 1.2m | – |
Save offer 20% off for 3 months | Still a customer 90 days later | 87.6%n=660 | 82.3%n=317 | +5.2 [+0.3, +10.1] · p=0.0276 | +34 | Rp 90.6m | Rp 2.6m |
5G upgrade Free 5G device swap with a 5G plan | Moved to 5G within 2 months | 25.4%n=1,175 | 10.6%n=293 | +14.8 [+10.5, +19.1] · p=0 | +174 | Rp 3.5m | Rp 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.
- 01 · behaviourWhat happenedusage, billing, requests
- 02 · modelWho is nextthe churn model
- 03 · surveyWhywhat customers tell us
- 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.
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)
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).
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
Unhappy with signal indoors, by how they were sold
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.”
Onboarding · NPS 4 · 4G 20 Mbps · Suburban · Retail outlet
“Di lantai dua lumayan, di bawah hampir tidak bisa”
Relationship · NPS 4 · 4G 20 Mbps · Metro · Online
“Sinyal naik turun, sering putus sendiri. Kecewa sih.”
Onboarding · NPS 4 · 4G 20 Mbps · Urban · Partner agent
“Sinyal di dalam rumah lemah, modem harus ditaruh dekat jendela. Kalau begini terus mau berhenti.”
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
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)
Leavers caught in the riskiest 20%
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
| Decision | Behaviour | Survey | Experiment | Verdict |
|---|---|---|---|---|
| Move covered 4G homes to 5G | Covered 4G lines churn 4.3% a month, 5G 2.5% | 55% of metro 4G unhappy with evening speed; NPS 60 on 5G | Upgrade campaign +14.8 pts against its control group | Go: all three agree |
| Fix relocations and faults within target | A late relocation fix loses 18.9% in 60 days vs 8.2% | Support rated 2.1/5 after a late fix, 4.4 on time | Not an experiment: a service standard to track | Go, and track it weekly |
| Signal check before selling online and via agents | 23.2% of online sales leave within 3 months vs 12.7% field sales | 17% of online customers unhappy with indoor signal vs 9% | Not tested yet | Pilot it with a control group |
| Save offer (20% off) for everyone at risk | Risk segments leave 2-3× as often as healthy customers | Price barely moves the score (r 0.21); low value ratings end in non-payment | On-time payers +14.7 pts; late 2+ times -2.8 pts | Narrow it to on-time payers |
| Welcome call with setup tips | 17.2% of new customers leave within 3 months | Unhappy new customers write about signal (40%) and evenings, not setup | +0.2 pts: no proven effect | Redesign around signal, re-test |
| Rebuy reminders before the package expires | Late rebuys repeat: habit drives the next one | 38% of 'didn't pay' leavers blamed the service, which a reminder can't fix | +5.8 pts on-time rebuys, at Rp 400 a message | Keep: cheap and it works |
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
By channel
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 planSurvey 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)
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.
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')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.
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.