Main site

Media AI agent · Retention

Churn-Prediction & Retention-Offer AI Agent

Score accounts for cancellation risk, assemble the evidence behind each score and propose a pre-approved offer — with pricing held by a named subscription owner and the cancel path left alone.

Typical delivery
4–6 weeks
Deployment
Your stack
Offer release
Named owner
After launch
Agent Care

What this agent does

Scores the account, proposes an approved offer

In

  1. 01

    Take the subscription record — plan, tenure, billing state and the promotions already applied to that account.

  2. 02

    Read what the account actually watched: sessions, titles started, devices and the gaps between them.

Reason

  1. 03

    Score cancellation risk against the model in force, and assemble the reasons that moved that particular score.

  2. 04

    Read the offer catalogue your subscription owner approved, and which of its offers this account is eligible for.

  3. 05

    Test the proposal against the written basis for offering different terms, so eligibility rests on a rule, not a score.

Decide

  1. 06

    Detect an account already cancelled, opted out or failing payment, and hold it out of the retention list.

  2. 07

    Route anything the catalogue does not carry — a new price, a credit, a changed term — to the person who releases it.

Out

  1. 08

    Present the score, its reasons and the eligible offer to a named owner, who decides whether it is made at all.

  2. 09

    Retain the score, the model version, the reasons, the offer proposed and what the owner released or refused.

Product statement

The agent scores and proposes. A named subscription owner approves the catalogue and releases each offer, and nothing on this path sits between a subscriber and cancelling.

Example workflow

One account, end to end

  1. 1Account readPlan, tenure, billing state, promotions already applied and the contact preferences held against the account
  2. 2Viewing assembledSessions, titles started, devices and support contacts across the window agreed in discovery
  3. 3Risk scored with reasonsA score from the model version in force, and the account-level reasons that moved it
  4. 4Eligibility and suppressionCatalogue eligibility, the written basis for differing terms, consent and preferences, and who has already cancelled

No human action required

Stages 1 to 4 run without a person in the loop — the score, its reasons and the suppression checks are finished before anyone is asked to look. Nothing has been said to a subscriber at that point.

5DecisionSplits on eligibility and the approved catalogue
Eligible, inside the catalogue

Goes to the owner as a proposal.

Outside it, or suppressed

Goes to the person who may release it, or nowhere.

Subscription owner and retention lead

Reads the score, the reasons behind it and the terms the catalogue attaches to that offer, then decides whether it is made, changed or not made at all — and is the only route to anything the catalogue does not already carry.

Release · Change · Do not offer

Released — handed back

  1. 6Offer released or refusedOnly from the catalogue, only by a named owner, and only on a channel the account has consented to
  2. 7Outcome evaluatedCalibration against what the renewal did, offers refused, complaints raised, and how long cancelling takes

Refusals

Every proposal an owner declines is counted against the score.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Setting a price, a discount or a term.
    • Applying a credit, a refund or a plan change.
    • Making an offer the catalogue does not carry.
    • Placing anything in front of the cancel button.

  • Automation boundary

    Agent acts unaided

    • ✓Score cancellation risk against the model version in force.
    • ✓Assemble the account-level reasons that moved a score.
    • ✓Check catalogue eligibility, consent and contact preferences before proposing.
    • ✓Hold cancelled, opted-out and failed-payment accounts out of the list.

    Write actions run only inside the approval boundaries agreed during implementation. Cancelling is not a path it steps into.

    • Contacting an account that has already cancelled.
    • Messaging outside the consent and preference record.
    • Deciding that one cohort gets worse terms.
  • Changing the model, thresholds or the catalogue.

Example output

One account, annotated

Everything the agent proposes is attached to the account it came from.

Retention output · single account

Illustrative example
Plan
Annual, renews in 21 days
Signal
No session in six weeks
Tenure
31 months
Proposed offer
Pause — catalogue entry
Confidence
78%
Price
Not set by the agent

As received

The subscription and the sessions on it, as billing and playback hold them — nothing here is inferred.

Evidence used

  • Six weeks, no session
  • Renewal in 21 days
  • Pause offer, eligible

Why a pause

The approved catalogue puts pause first for a lapsed annual account, not the agent.

What the score decides

Whether this account reaches an owner at all, not what it is charged or whether it may cancel.

Action

ReleaseChangeDo not offer

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every account, scored

Subscription, billing and playback records

03Catalogue & basis

Choose from a written catalogue

The offers, the eligibility rules and the basis for differing terms are yours.

01Approved path

Hand the owner a reason, not a rank

The account-level reasons arrive with the score, so the decision is about that account rather than its position in a list.

02Human review

Keep price and cancelling out of it

Anything touching a price, a term or the cancellation journey leaves for the person who is allowed to decide it.

04

Build an evidence trail

Retain the score, the model version, the reasons, the consent read, the offer proposed and what the owner did — on both paths.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

  • Subscription & billing

    • Recurly
    • Zuora
    • Chargebee
    • Stripe Billing
    • in-house billing

  • Identity & entitlement

    • Account service
    • plan catalogue
    • Entitlements
    • device registry

  • Playback & engagement

    • Player telemetry
    • viewing history
    • Watchlists
    • app analytics

Agent

Churn prediction & retention offers

  • Scores the account
  • Assembles the reasons
  • Proposes from catalogue

  • Offer & contact

    • Braze
    • Iterable
    • Consent and preference centre
    • CRM

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the score and the offer

Each control wraps the one inside it. A score clears every layer before it becomes a proposal, and the price sits outside all six.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeReturn scoring to your retention team if calibration or fairness signals degrade.Roll back
L5TraceabilityRecord the score, model version, reasons, consent read, offer and decision.Record
L4Cancellation parityThe cancel journey is measured against sign-up and reported when it lengthens.Measure
L3Owner releaseA named subscription owner decides whether an offer is made at all.Gate
L2Suppression and consentCancelled, opted-out and failed-payment accounts are held out.Suppress
L1Catalogue restrictionProposals are limited to offers your subscription owner has approved.Restrict
Model coreScore proposed — the risk, the reasons that moved it, the eligible offer and confidence
L1 – L2
Decide whether a proposal may stand
L3
Decides who may release an offer
L4 – L5
Keep cancelling easy and the trail intact
L6
Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the offer that was proposed, not only the score.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the score and offer an owner reads

  1. E1Final-output evaluationWas the proposed offer one the catalogue actually allows?
  2. E2Step-level evaluationDid the reasons given match what moved the score?
  3. E3Tool evaluationDid it read the right account, plan and consent record?
  4. E4Calibration and upliftDo higher scores cancel more, and did being contacted change anything?
  5. E5Slice evaluationHow do score quality and offer terms differ across account cohorts?
  6. E6Business outcomeWhat did an owner refuse, and how long does cancelling take?

Floor — what the offer changed, and what it did not

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle. An offer made on a wrong reading is honoured or withdrawn by the subscription owner, and the account told.

Agent lifecycle

Direction of processing →

  1. 01 · Account read

    1 mode

    • CP-04

      Failed card read as exit

      An expired card is scored as a decision to leave.

    Stage gathers

    Plan, tenure, billing state and the sessions on it

  2. 02 · Scoring

    2 modes

    • CP-05

      Wrong reasons attached

      Plausible reasons that are not the model's.

    • CP-06

      Settled account woken

      An account that was staying is proposed for an offer.

    Stage scores

    Cancellation risk and the reasons that moved it

  3. 03 · Eligibility

    2 modes

    • CP-02

      Cancelled, then chased

      An account that already left is scored and contacted.

    • CP-03

      Worse terms by cohort

      An eligibility rule gives one group less.

    Stage checks

    The approved catalogue, consent and suppression

  4. 04 · Proposal

    1 mode

    • CP-01

      Offer in the cancel path

      A proposal lands where a subscriber was leaving.

    Stage proposes

    The offer an owner reads, and the channel it would use

  5. 05 · Change / Version

    1 mode

    • CP-07

      Silent threshold drift

      A model or catalogue change widens who gets offered.

    Stage tracks

    Model, threshold, catalogue and cancel-flow changes

  • Sev-1 · the agent steps where it must not
  • Sev-2 · an owner decides on something untrue
  • Sev-3 · scores drift and proposals are refused
See the complete Media failure-mode catalogue

Affected slices

A score is weakest where there is least to read

New accounts, legacy prices and shared household profiles carry most of what a renewal later contradicts and most of what an owner refuses. Nestack reports performance by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
First-month new subscribers5.1%3.0×Review
Accounts on legacy prices3.9%2.3×Review
Shared household accounts3.1%1.8×Watch
Long-tenure monthly accounts1.2%0.7×Normal

Bar: mis-scored account rate lift vs. long-tenure monthly accounts · scale 0–4.0× · tick marks the 2.0× threshold

2 of 4 slices over threshold

Evidence-linked improvement

The label arrives after the renewal, not with the score

A score is a claim about something that has not happened yet, so nothing about it is settled until that renewal date passes and the account either stays or does not.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Refused proposals or contradicted scores move in one cohort.

  2. 02

    Diagnose

    Traced to the account read, a feature, a rule or the catalogue.

  3. 03

    Improve

    The rule, threshold or catalogue entry is changed with the owner.

  4. 04

    Verify

    Re-run against renewals that have since resolved, offers included.

  5. 05

    Learn

    The refused proposal is kept beside the reason the owner gave.

  6. Learn → Detect

    The return edge. Each cycle re-scores renewals that have now resolved — a score that looked right in March is settled only once that account renewed.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, the catalogue and suppression rules, scoring, evaluation, integration, then production validation and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

  1. 01Workflow discovery and automation-boundary definition.Week 1
  2. 02Subscription and billing system assessment.Week 1, Week 2
  3. 03Offer catalogue and eligibility-basis mapping.Week 1, Week 2
  4. 04Suppression, consent and contact-preference rules.Week 2
  5. 05Account and engagement feature assembly.Week 2, Week 3
  6. 06Risk scoring and reason assembly.Week 3
  7. 07Owner release workflow and refusal capture.Week 3, Week 4
  8. 08Cancellation-path parity measurement.Week 4
  9. 09Evaluation suite, slices and calibration cases.Week 4, Week 5
  10. 10Fairness testing across account cohorts.Week 5
  11. 11Billing and campaign-platform integration.Week 5, Week 6
  12. 12Observability, deployment and Agent Care handover.Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tierPilotOne subscriber cohortProductionProduction billing integrationAdvancedMulti-market / multi-brand
Introduced at Pilot
Cancellation-risk scoring with its reasons✓✓✓
Offers only from your approved catalogue✓✓✓
A named owner releases every offer✓✓✓
Cancelled and opted-out accounts suppressed✓✓✓
Cancel path measured against sign-up✓✓✓
Baseline evaluation, including cohort fairness✓✓✓
Introduced at Production
Billing and engagement data integration—✓✓
Owner release workflow and refusal capture—✓✓
Observability and slice evaluation—✓✓
Introduced at Advanced
Multi-market and multi-currency catalogues——✓
Multi-brand and enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on billing and engagement systems in scope, catalogue complexity, subscriber volume, markets and currencies, approval controls and deployment requirements.
Separate from build
Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

  1. 01The offer catalogue you already approve, and who owns it

    Offer catalogue and eligibility-basis mapping

    Week 1

  2. 02Your subscription and billing systems, and how plans are held

    Subscription and billing system assessment

    Week 1

  3. 03Your consent, contact-preference and suppression rules

    Suppression, consent and contact-preference rules

    Week 2

  4. 04Renewals and cancellations from a period already closed

    Risk scoring, reason assembly and the calibration cases

    Week 3

  5. 05The cancel journey a subscriber actually walks today

    Cancellation-path parity measurement

    Week 4

  6. 06The written basis on which offers may differ between accounts

    Fairness testing across account cohorts

    Week 5

  7. 07A named subscription owner and a retention lead

    Owner release workflow, then supervised offers and validation

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases are drawn over the weeks they actually occupy. Week 5 carries both the fairness slices and the first offers a subscriber is actually made.

Phase

W1

W2

W3

W4

W5

W6

  1. DiscoveryW1
  2. BuildW2 – W3
  3. EvaluateW4 – W5
  4. Pilot & LaunchW5 – W6

Week focus

W1The catalogue, its owner and the eligibility basis

W2Billing and engagement data, suppression and consent wired in

W3Risk scoring, reason assembly and the owner release workflow

W4Cancel-path parity, evaluation suite, slices and calibration

W5Fairness testing, integration and the first supervised offers

W6Offers released inside the boundary, then Agent Care starts

Reading the band
Suppression and consent are built in week 2, before a score is assembled in week 3. Nothing already cancelled or opted out reaches a proposal.
At the end of W6
Offers have been released from your own catalogue by a named subscription owner, and the cancel path has been counted against sign-up, then Agent Care takes over monitoring.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Media AI agent

Build a retention agent around your offer catalogue and your subscription owner.

Show us how your plans and billing are held, the offer catalogue you already approve, and the cancel journey a subscriber walks today. We'll score a period that has already closed against what those renewals did, and count the steps from your cancel link to the confirmation.

Typical delivery
4–6 weeks
Deployment
Your subscription and billing stack
Offer release
A named subscription owner
After launch
Agent Care available

Nestack Agents · Churn prediction & retention offers

AGT-ME-11 · Agent Care available after launch

Build a retention agent

Four questions so the call starts from your actual setup — then pick a time that suits you.

About you
Your setup

No obligation. We’ll map the workflow, name the automation boundary and tell you if an agent is the wrong answer.