AI Loop

Retention, a customer kept every week.

One agent that works with the customers you already have. Every Monday it reads who came back, who went quiet and whose payment failed, sees what competitors do to keep theirs, and listens to why people left. Then it writes the week's messages, the save offer and the referral ask, and stops. Nothing is sent until you've said yes.

Step 01 · The numbers

We read who stayed
and who went quiet

Every Monday the agent pulls last week from billing, the product and the inbox. Not just how many left — why. A customer who cancelled and a customer whose card failed look the same in the total, and need completely different messages. The agent keeps them apart, and puts a number on what's still recoverable.

  • Customers in, out and back, vs last week and last year
  • Cancelled by choice vs a card that failed — kept apart
  • Revenue at risk this month, and how much is recoverable
  • Who went quiet: no login in 14 days, usage down by half
  • Referred customers: how many, and how long they stay

Last week

Customers8,412
Left96
Came back41
FetchingDone

Why they left

Cancelled52
Card failed44
Involuntary46%
Reading billingDone

At risk now

Health under 60318
No login 14 d204
Card expires71
ScoringDone

Revenue at risk

This month94,000
Recoverable~60%
Referred last week23
ComparingDone
Step 02 · Competitors

We see how
the others keep theirs

Then the agent reads the competition where it matters for this loop: what happens when their customer tries to leave. It walks their cancel flow, counts the clicks, reads what they offer to stay, what a lapsed customer gets a month later, and what a friend gets for bringing a friend.

  • Their cancel flow: how many clicks, what they offer to stay
  • Win-back: what a lapsed customer gets, and when
  • Referral: who gets what, and at which moment
  • Loyalty and annual plans — what's bundled to stay
  • The retention move nobody in the market has made

How they keep customers

Competitor #1pause instead of cancel
Competitor #2annual plan: 2 months free
Competitor #3no win-back at all
Competitor #4refer a friend: both get a month
Competitor #5cancel takes six clicks
Reading competitor 1 of 5 Analysis done
Step 03 · Why they left

We listen to
the people who left

Numbers say how many. The people say why. The agent reads every exit survey, the last support tickets before a cancel, and the calls from the week, and counts the reasons by theme. "Too expensive" and "stopped using it" are two different problems, and only one of them is fixed by an offer.

  • Exit survey answers, counted by theme, not by feeling
  • "Too expensive" vs "stopped using" — different fixes
  • Support tickets in the last 30 days before they left
  • The words they used for what we didn't do
  • Which reasons a save offer can fix, and which it can't

Exit reasons · 96 left

Why they leftSaid itSaveGone
Too expensive right now386436
Stopped using it294159
Forgot it existed217228
Missing a feature173565
Switched to a rival91288
A bad support moment68020
Cancel survey · support · 12 callsSave = % an offer can fix
Step 04 · Risk watch

We catch them before
they click cancel

The best save happens before anyone reaches the cancel button. Every customer gets a health score from five signals: how often they log in, whether they still use the thing they pay for, what support looks like, whether the card is fine, and whether they open what we send. Under 60 gets a message. Under 40 gets a person.

  • A health score per customer, 0 to 100, from five signals
  • Login down by half for two weeks: high risk
  • Data export started: critical, someone calls today
  • Card expiring in 30 days: fixed before it fails
  • Nobody contacted twice inside the cooldown window

Risk list · this week

Account · signalScore
Account 2190Data export started · yesterday22
Account 4471Logins down 62% in two weeks34
Account 7732Three seats removed on Friday41
Account 1058Billing page visited four times48
Account 9902Rated us 4 of 10 last week57
Account 3315Card expires in 18 days71
318 under 60 · 41 new this week12 get a call
Step 05 · The decision

Everything weighed
into three moves

Here everything meets. The numbers, the competitors, the exit reasons and the risk list are laid next to last week's messages, and the agent draws one conclusion. Not a campaign to everyone — three moves, each with a segment, a reason and a number to judge it by next Monday.

  • Where last week's messages worked, and where they didn't
  • Three moves this week, no more
  • Who gets a call, who gets an email, who gets nothing
  • One thing we stop sending, named
  • A number per move, judged next Monday

Week 38's decision

The numbers
Competitors
Exit reasons
The risk list
Three moves
  • Call the 12 accounts that started an export or dropped seats
  • "Pause instead" goes into the cancel flow, before any discount
  • Win-back to 214 lapsed 60–90 days — no discount in email one
  • Judge by: saves 18% → 25%, card recovery 61% → 70%
Step 06 · The lifecycle

One message, one job,
one moment

Before any copy, the agent settles the moments. A customer hears from us six times in a normal year: when they arrive, after the first week, after the first month, before renewal, when they've gone quiet, and when a card fails. Each moment has one message with one job. Between them: silence, on purpose.

  • Welcome, first week, first month, renewal, lapsed, failed
  • Each message has one job and one button
  • Value before the ask: what they got, then what's next
  • Nothing sent between the moments — silence is fine
  • The renewal recap: what it was worth, in their numbers

Six moments

Day 0Welcome · the first step, one buttonemail
Day 7Check-in · did it work? Reply to a personemail
Day 30Your first month · what you usedemail
Day 335Renewal in 30 days · what it was worthemail
Quiet 60Win-back · we noticed, no discountemail
FailedCard didn't go through · one-click fixemail + sms
Setting the moments
Step 07 · The save offer

The offer matches
the reason

When someone does click cancel, the agent asks why first, and answers with the one offer that fits. "Too expensive" gets a pause or a smaller plan. "Not using it" gets a person and a setup, not a coupon. Nobody gets 50% off just for reaching the button, because that trains everyone to reach it.

  • The offer matches the reason, not one discount for all
  • "Too expensive": pause or a smaller plan, never 50% off
  • "Not using it": a call and a setup, not a coupon
  • Cancel stays two clicks — an offer, not a maze
  • Every save is tagged so it can be judged in 90 days

Save offers · by reason

Too expensive0SEK
  • Pause 1–3 months
  • Data kept
Most pick149SEK
  • Smaller plan
  • + same data
  • + upgrade any day
Not using it0SEK
  • 15 min with a person
  • Setup done for them
  • Check-in at day 14
This week's movepause before any discount
Expectedsaves 18% → 25% of cancels
Instead of20% off for everyone
Step 08 · Payment recovery

The card that failed
gets fixed, not blamed

Failed cards can be a third to half of all churn, and they are the easiest churn to get back. The agent runs the whole ladder: a note before a card expires, a retry on the day that worked last time, four plain-text emails over ten days, and a way back in after the cancel. No blame in any of them.

  • Failed cards: up to half of churn, and the easiest to fix
  • Card expiring: a note 30, 15 and 7 days before
  • Soft decline: retry on the day that worked last time
  • Hard decline: no retry — ask for a new card, once
  • Four plain-text emails over ten days, then a way back

44 cards failed · Monday

Day 0Email 1 · "It didn't go through. Update here."+9 recovered
Day 1Retry 1 · soft declines only+7 recovered
Day 3Email 2 + retry · "Quick reminder"+8 recovered
Day 5Retry 3 · on the day that worked before+4 recovered
Day 7Email 3 · "Paused in 3 days" + SMS+3 recovered
Day 10Email 4 · last chance, support number+1 recovered
Day 11Paused · one link back in, data kept 90 days12 paused
32 of 44 back · 73% · was 61%
Step 09 · Referral

The happy ones
bring the next one

A kept customer is worth more when they bring another. The agent finds the ones who just had a win — hit a milestone, rated us 9 or 10, renewed — and asks them then, not at signup. Both sides get a month, the message is written, and sharing is one click. Referred customers stay longer, which is why the month is worth it.

  • Ask right after a win, never at signup
  • Both sides get something — one month each
  • One click to share, the message pre-written
  • Referred customers churn 18–37% less — worth the month
  • Limits: verified, capped, never cash

Who to ask this week

Ask now after a win
Hit a milestone · 118 Rated us 9–10 · 64 Renewed · 37
The offer both sides
A month each · 8.1 Gift card · 6.4 Tiered: 3 friends, 3 months · 5.9
Parked with reason
Cash reward · fraud Ask at signup · too early
219 to ask · 23 referred last week
Step 10 · AI agents

The right agents
are put on the job

Now the right agents are pulled in. One writes the sequences, one reads the risk and runs the cancel flow, one builds the referral ask, one keeps every text message short and legal — and one holds the state: who was contacted, when, and who is inside the cooldown and must be left alone.

  • Which agents the week needs — not all, every time
  • One writes the sequences, one reads the risk
  • One builds the referral ask, one keeps SMS legal
  • One holds the state: who was contacted, when, cooldown
  • Every message is tagged with the move it belongs to

Agents on the job

emails
StartingWorking
churn-prevention
StartingWorking
referrals
StartingWorking
sms
StartingWorking
marketing-loops · analytics · customer-research
StartingWorking
Step 11 · The copy

Then the messages
get written

This is where the moves become words. The win-back that doesn't beg, the card email that doesn't blame, the text message that fits in one segment, the save offer, the referral ask and the check-in for the account that went quiet. Short, plain, and nothing the company can't stand behind.

  • Plain text for card emails — it beats designed email
  • One email, one job, one button
  • SMS under 160 characters, sender first, STOP at the end
  • No blame: "it didn't go through", not "you failed to pay"
  • Every line tagged with the move, so it can be judged

Writing the week

Win-back · email 1 · day 60
Win-back Card SMS Save Referral Check-in
Step 12 · Approval

Here everything stops
and waits for you

The week is written. The win-back, the card emails, the text message, the save offer, the referral ask and the call list sit ready, exactly as they will go out — nine drafts. Nothing is sent until someone says yes.

Showing 1–3 of 9 drafts
H Win-back · email 1
214 lapsed 60–90 days · draft
Did something change?
No offer in this one — just a question
You haven't been in since June. Nothing's wrong on our side — we just wondered if something changed on yours. If it did, reply and tell us. If it didn't, everything is where you left it. Email 2 in 14 days: what's new. Email 3: the offer. Then we stop.
Pick up where I left off
H Card failed · email 1
day 0 · plain text · draft
Your payment didn't go through
Takes 20 seconds to fix, no login needed
It happens — a card expires, a bank blocks a charge. Nothing has changed in your account and nothing will for ten days. Update your card here and that's the end of it. If it's easier, reply and a person will help. Everything stays as it is. Your data, your settings, your team.
Update card
HText message
card expiring · 71 customers · draft
From Nord · to opted-in only
From Nord: Your card expires this month. Update it in 20 seconds so nothing pauses: nord.se/card. Reply STOP to opt out.
134 characters · 1 segmentTuesday 10:00Not the same day as email
HCancel flow
two clicks, one offer · draft
Now20% off for three months, shown to everyone who clicks cancel
ProposedAsk why first. "Too expensive" → pause up to 3 months. "Not using it" → 15 min with a person. Then cancel, one click.
Judged in 90 daysTarget: saves 18% → 25%
HCall list
12 accounts · Monday · draft
22Account 2190 — started a data export. Call today, ask what they need it for
34Account 4471 — logins down 62%. Ask if the champion left
41Account 7732 — three seats removed. Offer the smaller plan before they ask
48Account 1058 — billing page four times. Send the annual option, no call
RuleNobody on this list was contacted in the last 30 days
H Referral ask
after a milestone · 118 customers · draft
You just hit 100. Know someone who'd want the same?
A month on us, for both of you
A hundred orders through Nord this week. That's yours, not ours. If you know someone who'd get the same out of it, send them your link. When they start, you both get a month free. One click. The message is already written — change it if you like.
Share my link
HCheck-in · at risk
no login 14 days · 204 customers · draft
NNordfrom a person, not a no-reply
We noticed you haven't used reports since August
Usually that means it wasn't set up the way you wanted. Want 15 minutes with someone who'll do it for you? Pick a time, or reply "no thanks" and we'll leave it.
One button: pick a timeNo offer, no discount
H Hubso · memo
To the owners · Monday · draft
Week 38: 96 left, 44 of them by accident
Three moves, one stop · 3 minutes to read
Last week 96 customers left. 44 were failed cards, and we got 61% of those back — the ladder should take it to 70%. Of the 52 who chose to leave, 38 said price. The cancel flow gives everyone 20% off; from this week it asks why first and offers a pause. Twelve accounts get a call on Monday. 214 lapsed get a question, not a discount.
Read the week
HStop list
what we no longer send · draft
Stop20% off to everyone who clicks cancel — trains them to click
StopThe monthly newsletter to lapsed customers — 0.4% open
StopDiscount in win-back email one — moved to email three
StopRetrying hard declines — ask for a new card instead
FreedSEK 31,000 a month in discounts nobody needed
Step 13 · Send and track

Each message goes
at its own moment

After your yes, nothing goes out in one Monday blast. Each message leaves at its moment: the calls first, the text on Tuesday, the check-in midweek, the referral ask on Thursday when people are in a good mood, the card emails on their own clock. Every send is logged, and nobody hears from us twice in thirty days.

  • Each message goes at its moment, not in one Monday blast
  • Email and SMS never land on the same day
  • Every send logged: who, when, which move
  • Cooldown: 30 days before the same person hears again
  • Saves, recoveries and referrals counted per message

Week 38 · going out

Mon12 calls · export, seats, loginsdone
MonWin-back 1 · 214 · no discountsent
TueCard-expiry SMS · 71sent
WedCheck-in · 204 quiet · from a personsent
ThuReferral ask · 219 · after a winsent
DailyCard emails · on each card's own clockrunning
SunRead: saved, recovered, referred23:00
Sending
Step 14 · And then it starts again

Next Monday
it reads the results

Next Monday step 1 reads exactly what this week promised: how many were saved, how many cards came back, how many friends were brought in. What worked is kept. What didn't is changed once, with the reason saved. After a year the company has a retention playbook built on its own customers, not on someone else's best practice.

  • Every message judged against the number it carried
  • What saved gets kept; what didn't gets one change
  • Failed hypotheses saved with the reason why
  • The exit reasons rewrite next month's offers
  • After a year: a playbook built on your own customers
Week 39 And again 14 steps · every week