Optimizing Product Strategy Using Analytics

Use retention, lifetime value, health and customer feedback to decide where to invest in the installed base. Separate who might leave from whom an intervention can help.

Why this matters

Annual renewals reveal churn late. Usage can decay months earlier, while expansion from a few accounts hides losses elsewhere.

Compare contract and usage survival, revenue retention and lifetime value by segment. Link health and sentiment to the people who use, champion and buy the product. These analyses inform pricing, lifecycle decisions and business reviews.

Survival analysis: when customers leave

The survival curve and the hazard

Survival analysis estimates how long customers stay and when they are likely to leave. Two concepts matter:

• The survival curve shows the share of a starting group still with you after each period. It starts at 100% and can only stay flat or fall. One of the most widely used ways to draw one from real data is the estimate Edward Kaplan and Paul Meier published in 1958, built for exactly the situation where many subjects haven’t reached the end of their story yet.
• The hazard is the chance that a customer who has made it to a given period leaves during that period: the risk right now, for customers still here.

The curve shows cumulative survival. The hazard locates risk: a month-twelve peak suggests renewal work, while a month-two peak points to onboarding.

Peter Fader and Bruce Hardie, whose work underpins much practical retention analysis, note that expected tenure is the area under the survival curve. A small lift in early survival raises the curve for every period after it, which is why early retention is worth so much.

Cox’s 1972 regression model estimates how factors scale the hazard. Ask about factor effects, uncertainty and account counts. Missing connectors or lost champions may correlate with churn without causing it.

Censoring: the customers who haven’t left yet

A still-active eight-month customer has an unfinished lifetime. This is censoring. The lifelines documentation, maintained by Cameron Davidson-Pilon, describes the unobserved remainder of that history.

Dropping current customers removes long-lived observations. Treating today as their departure also understates lifetime. Survival methods count observed periods without inventing churn. State this rule in every request.

Contract churn and usage churn

B2B products have two kinds of departure, and each needs its own survival curve.

Contract churn is the customer not renewing. It’s observed and unambiguous, and it happens at renewal dates. Fader and Hardie call this the contractual setting. With annual contracts, the contract hazard sits near zero mid-term and spikes at each renewal.

Usage churn is the customer still paying while use fades: seats go idle, workloads move, the champion leaves and nobody replaces them. It resembles what Fader and Hardie call the noncontractual setting, where you never see a clean exit and must infer from behavior whether a customer is still “alive.”

Usage survival is the leading indicator; contract survival is the lagging, financial result. A PM who watches only contract churn learns about problems once a year, with no time left to act.

For usage-priced products the two collapse into one: usage decay is revenue churn, immediately. If you charge per resolved incident or per unit consumed, a customer whose usage halves has halved their bill, with no renewal date to warn you.

The sorting effect

Retention often rises with tenure. That alone cannot establish increasing loyalty.

Fader and Hardie explain the sorting effect: high-churn-propensity customers leave early, leaving a more loyal population. No individual change is required.

Compare like segments before crediting product changes. Use survival methods rather than fitting aggregate trend lines, which can imply impossible retention or survival.

Cohort retention

A cohort is a group of customers who started in the same month or quarter. The standard view is a retention triangle: one row per cohort, one column per period since start, each cell showing the share retained. Read it three ways:

• Down a column: do newer cohorts retain better at the same age? This shows whether onboarding or product changes worked.
• Along a row: where does decay stop? A curve that flattens shows a durable core getting lasting value. Lenny Rachitsky and Casey Winters, in a 2020 survey of growth leaders, argue that a flat retention curve plus a scalable acquisition strategy is enough to keep a business alive. A curve still decaying toward zero signals weak fit.
• Along a diagonal: the same calendar month across cohorts. A diagonal dip points to an event that hit everyone at once, such as a price change, an outage or a competitor’s launch.

Track three kinds of retention, because they diverge:

In B2B, cohort by segment as well as by date: annual contract value (ACV) band, industry, channel and onboarding path often explain more than signup month. SaaS Capital’s 2025 retention benchmarks found contract value the best single axis for comparing retention, ahead of company age, size or industry.

Revenue retention: NRR and GRR

Definitions

Two numbers summarize a recurring-revenue base. SaaS Capital’s definitions are typical:

• Net revenue retention (NRR): recurring revenue today from customers who were customers twelve months ago, divided by those customers’ recurring revenue twelve months ago. It includes expansion, cross-sell and price increases, net of contraction and churn, and can exceed 100%.
• Gross revenue retention (GRR): the same calculation without expansion, cross-sell and price increases. It can’t exceed 100%, and revenue churn is roughly 100% minus GRR.

NRR asks whether the installed base grows on its own. GRR asks how much of it you keep. You need both, with logo retention beside them.

Decompose before you conclude

Build an NRR bridge for each segment: starting revenue a year ago, plus expansion (seats, environments, new products, price increases), minus contraction (downgrades, lower consumption), minus churn (customers lost entirely).

Strong GRR with modest expansion suggests packaging or cross-sell investment. Weak GRR masked by a few expanding accounts suggests onboarding, adoption or segment-retention work.

Benchmarks and comparability

Benchmarks vary by sample and year. SaaS Capital’s 2025 survey of over 1,000 private B2B SaaS companies found median NRR of 101% and GRR of 91%. ACV above $250,000 had about 95% GRR; multiyear contracts had 103% NRR and 94% GRR. NRR above 110% correlated with above-median growth. Rachitsky and Winters’ 2020 practitioner survey called roughly 110% good and 130% great for enterprise. Keep these reference groups separate.

NRR definitions differ in cohorts, periods and revenue bases. Snowflake’s January 2026 fiscal-year report used consumption revenue and a trailing two-year capacity-contract cohort for its 125% NRR, retaining inactive customers at zero. Compare definitions before figures.

Usage-priced NRR also moves with workloads: a customer can “expand” because they had a bad quarter with more incidents. Separate environment-driven volume from adoption-driven expansion.

Customer lifetime value

What CLV is for

Customer lifetime value (CLV) discounts expected relationship margin to present value. Gupta, Lehmann and Stuart’s “Valuing Customers” (2004) shows how expected discounted customer earnings can support firm valuation, including growing loss-making firms.

Use CLV to compare segments and channels, acquisition limits, retention versus acquisition investment, and pricing effects. Retain the relevant cost-to-serve differences.

Simple and predictive models

A simple CLV model uses periodic margin, constant retention and a discount rate. Gupta and Lehmann’s margin-multiple view shows how strongly value rises as retention approaches 100%.

The weakness is the constant rate, which ignores the sorting effect. Fader and Hardie’s 2010 paper on customer-base valuation shows that one aggregate retention rate gives a downward-biased estimate of the remaining value of the existing base, because the customers still with you are, on average, the loyal ones.

Better models weight each period’s discounted margin by the survival curve for that segment or account. Fader and Hardie’s shifted-beta-geometric model for contracts gives each customer their own constant chance of leaving, varying across customers, which captures the sorting effect simply. Machine-learning models can also predict CLV per account from usage, firmographics and support history.

Use a per-account predictive model when the decision differs by account: where to put customer success capacity, which accounts get an executive sponsor, which renewals get a pricing exception. For segment-level questions, a survival-weighted segment model is usually enough.

Pitfalls

1. Revenue instead of margin. Use contribution margin after the cost to serve. For AI products this matters more than it used to: Bessemer Venture Partners’ State of AI 2025 report found its fastest-scaling AI companies averaging gross margins of about 25%, often negative, and its steadier cohort about 60%. Inference cost belongs in the margin line.
2. One blended retention rate. Enterprise and mid-market curves differ, and averaging them misstates both.
3. Expansion forever. Plug NRR above 100% into a simple formula and value grows without limit. Cap the expansion assumption or let it decay.
4. Ignoring cost to serve. Customer success time, support load, partner margin and onboarding effort vary sharply by segment.
5. False precision. “Between $0.9M and $1.5M, most sensitive to year-two retention” starts the right conversation; “$1.2M” ends it.
6. CLV without payback. A customer worth a lot in year six doesn’t fund this year’s payroll.
7. Ignoring the partner layer. When partners resell or manage your product, model both the partner relationship and each end customer, because partner churn takes many end customers with it.

Health scores and churn prediction

What a health score is

A customer health score combines several signals into one banded rating per account (green, yellow, red) so customer success and product teams can see which accounts need attention. Gainsight’s chief customer officer Kellie Capote described a structure called DEAR:

• Deployment: is the customer set up to use what they bought? Compare entitlements with what’s activated.
• Engagement: are you connected to the right stakeholders, beyond one champion?
• Adoption: how broadly (across users) and deeply (in core workflows) do they use the product?
• ROI: are they reaching the outcomes they bought for, verified against agreed milestones?

DEAR extends health beyond use. Gainsight also emphasizes sentiment, relationships and customer value. Rework scores when churn surprises you. Segment them because direct enterprise and partner customers may have different risks.

Rules or a learned model

A rules-based health score uses chosen signals, weights and thresholds. A learned churn model estimates weights from historical outcomes. Models can rank risk; rules can make actions easier to explain. Many teams combine them. With few annual renewals, keep either approach simple and validate it.

Validate against outcomes

A health score is a prediction, so test it like one. For the last four to eight quarters of renewals, how many churned accounts were green 90 days out, and how many reds renewed happily? If most churned accounts were green, the score is decoration.

Watch for label leakage: signals that appear only after the customer has decided to leave, such as a cancellation request, a bulk data export or a request for contract terms. They make a back-test look accurate and fire too late to help. Watch for score inflation too: if customer success managers set their own sentiment input and are measured on the share of green accounts, the score drifts green.

Risk is not the same as response

Ascarza’s 2018 retention-futility studies distinguish churn risk from response to intervention. Targeting likely responders outperformed targeting the highest-risk customers. In some cases, outreach to likely defectors increased churn.

Randomize an intervention within a risk band. Some customers already decided to leave and others would stay without help. Use pilot results to identify which accounts benefit from outreach.

Sentiment and the voice of the customer

Sentiment shows what customers think and say, which often explains what they do.

NPS and its critics

Fred Reichheld introduced the Net Promoter Score in a 2003 Harvard Business Review article. Customers rate how likely they are to recommend the company from zero to ten. Nines and tens are promoters, sevens and eights passives, zero to six detractors, and NPS is the percentage of promoters minus the percentage of detractors.

NPS prediction claims remain contested. Keiningham and colleagues’ 2007 Norwegian study could not replicate its superiority over satisfaction measures. Sauro’s 2018 replication found the original analysis used past growth. Against later growth, NPS explained about 38% of two-year variation. Treat it as one signal.

In “Net Promoter 3.0” (2021), Reichheld, Darnell and Burns criticized misuse, including bonuses linked to scores. They proposed Earned Growth, using audited revenue from repeat, expanded and referred purchases as an accounting-based counterpart.

In B2B, samples per account are small, the respondent is often not the person who renews, and one detractor at a top-ten account outweighs twenty promoters at small ones. Report NPS by account and role, next to retention.

CSAT and customer effort

Customer satisfaction (CSAT) surveys ask how satisfied a customer was with a specific interaction, usually right after it: a support case, an onboarding session, a release. They spot a broken touchpoint quickly and say little about renewal.

Matthew Dixon, Karen Freeman and Nicholas Toman argued in “Stop Trying to Delight Your Customers” (Harvard Business Review, 2010) that reducing customer effort drives loyalty in service interactions more than delight does, and introduced the Customer Effort Score. For enterprise products, effort is a sharp lens on onboarding, integration and support.

Text analytics with LLMs

Tickets, transcripts, communities, sales notes, surveys and churn interviews provide rich text evidence. LLM analysis has mixed results:

• Zhang and colleagues, testing 13 sentiment tasks across 26 datasets in 2023, found large language models did well on simple polarity, lagged on harder tasks such as sentiment about specific product aspects, and clearly beat small models when labeled examples were scarce.
• Von der Heyde and colleagues, coding open-ended survey answers in 2025, found performance varied widely between models, and uneven accuracy across categories shifted the share of each theme, distorting the overall picture.
• A 2025 pilot by Langer Research Associates had a frontier model code 1,166 survey answers against a human codebook of more than 40 categories. In the first 500 answers, 35% of substantive responses went uncoded or were obviously miscoded, and 42% were acceptable. The model matched on keywords in category names, missed negative meaning in a positively framed question and missed sarcasm. The authors noted the narrow scope and a hard question, and concluded the oversight needed undercut the productivity gain.

Newer models may improve. Validate aspect-level themes, negation, sarcasm and domain language rather than assuming good polarity classification implies reliable coding.

Anthropic’s Clio, later Anthropic Insights, offers a privacy-conscious pattern: extract facets, cluster items, summarize and organize clusters, and suppress small groups that could identify individuals. Adapt it to voice-of-customer analysis.

The voice-of-customer program

A voice-of-customer (VoC) program turns listening into decisions. It has five parts:

• Listening posts: surveys, tickets, call recordings, community forums, sales and CSM notes, churn and win-loss interviews, and product telemetry.
• A shared taxonomy: a stable set of themes (onboarding, reliability, pricing, trust, integrations) that every source maps to, so trends stay comparable.
• Closing the loop: respond to the customer who raised an issue, and separately fix the root cause behind a recurring theme.
• Ownership: each theme has an owner who reports on it.
• Revenue linkage: themes are reported with the retention and revenue of the accounts raising them.

Link themes to affected accounts and upcoming revenue decisions. Mid-market pricing complaints alongside falling GRR imply different priorities from an aggregate complaint count. Own the taxonomy, review proposed AI themes and trace their revenue implications.

From analysis to strategy

Every analysis in this chapter should end in a decision:

How to apply it

Run this method once to build a baseline, then repeat steps 4 to 9 every quarter.

1. Agree definitions with finance and customer success: churn (non-renewal, contraction, inactivity for how long), active, revenue base (contracted, billed or consumed) and cohort start (signature or go-live). Disagreement here is the usual reason two teams report different churn for the same product.
2. Choose three to six segments that matter for decisions, such as ACV band, channel, industry or onboarding path. More, and every cell is too small to read.
3. Inventory the sources (CRM, billing, telemetry, support and success systems, survey tools) and note which your harness reaches through connectors.
4. Draw contract and usage survival curves by segment, treating current customers as censored, and read the hazard to see where risk concentrates.
5. Build cohort triangles for logo, user and revenue retention, and read them down columns, along rows and along diagonals.
6. Decompose NRR and GRR by segment, and reconcile the totals to finance’s figures before showing anyone.
7. Build a CLV model by segment and channel on contribution margin after cost to serve, with survival-weighted retention, capped expansion and ranges. Name the one or two assumptions that move the answer most.
8. Define or revalidate the health score. Back-test it against four to eight quarters of renewals, remove leaking signals and design a randomized pilot for the at-risk band.
9. Run the VoC analysis on six months of text, validate the coding on a human-labeled sample and link themes to the revenue and retention of the accounts raising them.
10. Write the strategy implications. For each finding, state the decision it supports, the size of the prize, your confidence and what would change your mind. Bring two or three decisions to the portfolio review, not twenty findings.
11. File the outputs in the Product 360 as durable artifacts with owners and review dates, and log the decisions.

The AI lens

When the product is AI-powered

• Margin moves. A more engaged customer can be a less profitable one. Model CLV on contribution margin after inference, and recompute it when routing, pricing or usage changes (Chapter 31 covers the cost levers).
• Outcome pricing turns usage churn into revenue churn. Charge per task or per outcome and every drop in usage is an immediate drop in revenue. Usage survival becomes a financial metric, watched weekly.
• Trust events are churn events. One visible AI failure can raise the hazard for a whole account. Put trust signals in the health score: rejected AI recommendations, reversed agent actions, escalations after AI involvement, falling acceptance rates.
• Switching costs may be low. Bessemer’s 2025 report notes that retention in many fast-growing AI products is fragile where switching is easy, so read early AI retention curves skeptically.

Doing the work with AI

A harness with connectors to your CRM, billing, telemetry and support systems does the mechanics. Your work is a precise request, a verification routine and judgment (Chapter 19 covers asking a harness in general).

A precise request names the decision, sources and time window, population and exclusions, every definition, the method in plain words (“survival curve by segment, treating current customers as still active”), the outputs and the checks. It also rules out fitted trend lines, dropped current customers and conclusions from tiny cohorts.

Verify every output:

1. Reconcile totals to a number you trust, such as finance’s ARR or customer count, within a stated tolerance.
2. Echo the definitions. Have the harness restate every definition it used; mismatches surface here.
3. Spot-check records. Ask for five example accounts behind a finding and check them in the source system.
4. Show denominators. Every rate comes with its count. A 40% churn rate on five accounts is an anecdote.
5. Test sensitivity. Rerun with a different inactivity threshold or cohort start. If the conclusion flips, it isn’t one yet.
6. Argue the other side. Ask for the strongest case against the finding, including the sorting effect.

For text, go further, because LLM coding fails quietly:

1. Human-label a random sample (100 to 200 items is practical) before trusting automated coding.
2. Measure agreement and read the disagreements.
3. Compare theme shares between human and machine coding as well as item labels, since the distribution can shift while individual labels look fine.
4. Test negation, sarcasm and question framing on purpose.
5. Require verbatim quotes with source references for every theme.
6. Keep a frozen codebook for trend lines, and have the model propose new themes separately for human approval.
7. Remove personal data and check customer agreements and data residency rules before sending customer text to any model, and use only tools approved for customer data.

The Beacon case: what’s inside 118% NRR

Definitions first. With finance and customer success you agree that churn is non-renewal of the platform contract, active means at least one weekly active responder in a 28-day window, and cohorts start at go-live. You segment by channel (direct enterprise versus mid-market customers sourced through MSP partners such as Leo’s firm) and by Beacon Classic migration.

The NRR bridge. The harness reconciles to finance’s ARR within half a percent, and the blended 118% splits sharply. In this worked example, direct enterprise accounts show NRR of 124% and GRR of 94%, carried by large customers expanding into Beacon Agents. MSP-sourced accounts show NRR of 101% and GRR of 83%. Direct accounts hold most of the revenue, so the blend looks excellent while the partner channel leaks. Logo retention is starker: most logos lost last year were MSP-sourced.

The survival curves. Contract survival for MSP-sourced customers looks acceptable through the first renewal, but weekly active responders start falling about four months after go-live, right after the MSP’s onboarding team hands over, and the gap widens until renewal. Leo confirms it in an interview: his engineers configure Beacon per tenant, nobody owns tuning after go-live, and correlation quality drifts as customer environments change.

Two more findings come from the hazard view. Accounts migrated from Beacon Classic show a usage hazard spike in the two months after migration, which matters for the sunset plan (Chapter 30). And at the retail customer hit by the wrong rollback (Chapter 31), responders kept using Beacon to group alerts but stopped accepting its diagnoses for weeks: usage churn at the feature level, invisible in a logo count.

The health score. The old score is mostly logins and flagged only a minority of last year’s churned accounts as red 90 days out. You rebuild it on DEAR plus a trust component:

• Deployment: connectors live against entitled environments (observability, ticketing, chat, CI/CD).
• Engagement: relationships with Priya (champion), Marcus (economic buyer) and Sam (security approval current).
• Adoption: weekly active responders, share of incidents touched by Beacon, diagnosis acceptance rate.
• ROI: MTTR change against the customer’s baseline, confirmed in a business review with Marcus.
• Trust: rejected diagnoses, reversed agent actions and escalations after agent involvement.

Back-tested on six quarters of renewals, the new score catches far more of the churned accounts, mostly through Deployment and Trust.

The retention pilot. Customer success wants to call every red account; you propose Ascarza’s approach. Of 40 yellow and red accounts, a random 20 get a guided “Agents readiness review” (connector cleanup, tuning and a value session with the economic buyer) and 20 get normal coverage. After a quarter, reviewed accounts with an active champion improved usage clearly, while those whose champion had left barely moved. Customer success now aims the review at at-risk accounts with a live champion and handles champion loss through an executive sponsor program.

CLV by channel. A CLV brief compares channels on contribution margin after inference cost, partner margin and cost to serve. MSP-sourced customers cost far less to acquire, but at current GRR they’re worth less than direct customers in every scenario; at direct-level GRR, the channel would return more per dollar in most scenarios. The most sensitive assumption is year-two retention, exactly what the post-handover tuning gap destroys.

Voice of customer. The harness codes six months of tickets, post-incident reviews and Northlight Live feedback against the VoC taxonomy. On a human-labeled sample of 150 items, polarity agreement is good, but the pricing theme is undercounted and the model coded “Beacon would be great if it stopped paging me at 3 a.m.” as positive. After the codebook gains negation examples, two themes stand out: requests for EU AI Act documentation, mostly from security reviewers like Sam, and worry about unpredictable bills under per-resolved-incident pricing, concentrated in the MSP accounts already showing weak GRR.

The decisions. You bring three to the portfolio review:

1. A retention bet on the partner channel. Fund multi-tenant tuning and drift alerts for MSPs ahead of two Agents features, measured by MSP-sourced GRR each quarter.
2. Pricing guardrails. The per-resolved-incident proposal for Chapter 15 gains a monthly cap and a usage forecast per customer, so bill shock doesn’t drive churn in the channel least able to absorb it.
3. Trust as a retention investment. Eval quality and the confidence display (Chapter 31) join the retention plan, with the health score’s Trust component as the leading indicator.

The board deck keeps the 118% headline and adds a row beneath it with NRR, GRR and logo retention for each channel, so the partner-channel leak is visible every quarter.

Product 360 artifacts

# Retention and churn report spec · [Product] · v[n] · [Date]

**Owner:** [PM]   **Agreed with:** [Finance lead], [Customer success lead]   **Cadence:** Quarterly   **Next review:** [Date]

## Decisions this report supports
1. [For example: where to invest retention effort next quarter]
2. [For example: whether the onboarding change worked]

## Definitions (fixed; change only with a version bump)
| Term | Definition | Source system |
|---|---|---|
| Customer | [Contracting entity / account] | [CRM] |
| Cohort start | [Contract signature / go-live] | |
| Contract churn | [Non-renewal / termination] | |
| Usage churn | [No active users for N days / usage below X% of baseline] | [Telemetry] |
| Active | [Definition] | |
| Revenue base | [Contracted ARR / billed / consumed] | [Billing] |
| NRR | Revenue now from customers active 12 months ago ÷ their revenue 12 months ago | |
| GRR | As NRR, excluding expansion, cross-sell and price increases (max 100%) | |

## Segments
[Three to six: ACV band, channel, industry, region, onboarding path, migration status]
Minimum cohort size for reporting a rate: [N accounts]

## Sections
1. **Headline**: NRR, GRR and logo retention, overall and by segment, with counts.
2. **NRR bridge** by segment: starting revenue, + expansion, − contraction, − churn.
3. **Contract survival curves** by segment, current customers treated as censored.
4. **Usage survival curves** by segment, with the gap to contract survival.
5. **Hazard view**: where in the lifecycle risk concentrates.
6. **Cohort triangles**: logo, user and revenue retention.
7. **Churn reasons**: coded from exit interviews and notes, linked to revenue lost.
8. **What changed since last quarter** and why.
9. **Implications**: decisions proposed, with size of prize and confidence.

## Checks before publishing
- [ ] Totals reconcile to finance ARR within [x]%
- [ ] Definitions echoed back by the analysis and match this spec
- [ ] Five sample accounts spot-checked in source systems
- [ ] No rate shown without its count
- [ ] Sensitivity to the inactivity threshold tested

## Change log
| Version | Date | Change | Approved by |
|---|---|---|---|
| | | | |

# Customer lifetime value model brief · [Product] · [Date]

**Owner:** [PM]   **Finance partner:** [Name]   **Decision it informs:** [For example: channel investment, acquisition spend limits, CS coverage]

## Question
[One sentence. For example: "Is a mid-market customer acquired through partners worth more or less than a direct one, after cost to serve?"]

## Scope
- Segments or channels compared:
- Population and time window:
- Exclusions:

## Model type
Simple (constant retention, margin multiple) / Survival-weighted by segment / Predictive per account
Why this level is enough for the decision:

## Inputs and assumptions
| Input | Value or range | Source | Confidence |
|---|---|---|---|
| Revenue per customer per year | | | |
| Gross margin after inference and hosting | | | |
| Cost to serve (CS, support, onboarding) | | | |
| Partner margin (if any) | | | |
| Retention by year (from survival curve) | | | |
| Expansion assumption (capped or decaying) | | | |
| Discount rate | | Finance | |
| Acquisition cost | | | |
| Horizon | [years] | | |

## Outputs
| Segment | CLV low | CLV base | CLV high | CLV ÷ acquisition cost | Payback period |
|---|---|---|---|---|---|
| | | | | | |

## Sensitivity
The two assumptions that move the answer most:
1.
2.

## What this model doesn't capture
[For example: strategic value of logos, partner-level churn, referral value]

## Recommendation and confidence
[Decision supported, confidence level, what would change it]

# Customer health score definition · [Product] · v[n] · [Date]

**Owner:** [PM or CS lead]   **Purpose (decision it drives):** [For example: which accounts get proactive CS outreach]   **Review date:** [Date]

## Population
[Which accounts; segment variants if the model differs by segment]

## Components
| Component | Signal | Source system | Weight | Green | Yellow | Red | Available how early before renewal? |
|---|---|---|---|---|---|---|---|
| Deployment | [Entitled vs activated] | | | | | | |
| Engagement | [Stakeholder coverage: champion, buyer, gatekeeper] | | | | | | |
| Adoption | [Breadth and depth of use] | | | | | | |
| ROI | [Agreed outcome milestones met] | | | | | | |
| Trust (AI products) | [Rejected or reversed AI actions] | | | | | | |
| Sentiment | [Survey, ticket themes] | | | | | | |

## Overall bands
Green: [rule]   Yellow: [rule]   Red: [rule]
Refresh cadence: [daily / weekly]

## Exclusions (leakage check)
Signals deliberately left out because they appear only after a decision to leave:
-

## Playbook per band
| Band | Owner | Action | Response time |
|---|---|---|---|
| Green | | | |
| Yellow | | | |
| Red | | | |

## Validation
- Back-test window: last [4 to 8] quarters of renewals
- Share of churned accounts that were red 90 days before renewal: [x]%
- Share of churned accounts that were green 90 days before renewal: [x]%
- Share of red accounts that renewed: [x]%
- Retention pilot (randomized) planned or run: [link]

## Governance
Who can change weights: [name]   Manual overrides allowed: Yes / No, logged by [name]

# Retention pilot plan · [Intervention] · [Date]

**Owner:** [PM]   **CS partner:** [Name]   **Decision it informs:** whom to target with [intervention] at scale

## Hypothesis
Offering [intervention] to [risk band or segment] will improve [usage measure / renewal] compared with normal coverage.

## Design
- Eligible population: [accounts, risk band, segment]
- Assignment: random, [n] treatment and [n] control
- Intervention: [what exactly happens, who delivers it]
- Control: normal coverage
- Duration: [weeks]
- Primary measure: [for example weekly active users, renewal intent, GRR]
- Secondary measures:

## What we want to learn
Which kinds of accounts respond (for example: champion present, deployment incomplete, segment)

## Guardrails
- No account is denied help it requests
- Stop rule: [condition]

## Result
| Group | Accounts | Primary measure before | After | Change |
|---|---|---|---|---|
| Treatment | | | | |
| Control | | | | |

Accounts that responded most: [characteristics]
Targeting rule going forward: [rule]

# Voice-of-customer report · [Product] · [Period]

**Owner:** [PM]   **Sources:** [surveys, tickets, calls, community, CSM notes, churn interviews]   **Items analyzed:** [n]   **Codebook version:** [v]

## Headline
[Two or three sentences: the themes that matter most and what they cost or could earn]

## Themes
| Theme | Share of items | Trend vs last period | Accounts raising it | ARR of those accounts | GRR of those accounts | Owner |
|---|---|---|---|---|---|---|
| | | | | | | |

## For each top theme
**[Theme name]**
- What customers say (three verbatims with source references):
  1. "[Quote]" ([source, date])
  2.
  3.
- Who raises it: [segments, roles]
- Link to retention or revenue: [evidence]
- Root cause (if known):
- Action and owner:

## New themes proposed this period (pending approval)
| Proposed theme | Example items | Approve / merge / reject |
|---|---|---|
| | | |

## Sentiment scores (context only)
| Measure | This period | Last period | Responses | By role |
|---|---|---|---|---|
| NPS | | | | |
| CSAT | | | | |
| Effort score | | | | |

## Closing the loop
- Individual follow-ups completed: [n of n]
- Root-cause fixes shipped since last report:

## Method and quality
- Human-labeled validation sample: [n] items; agreement: [measure]
- Theme shares, human vs machine: [comparison]
- Known weaknesses: [for example negation, sarcasm, a theme undercounted]
- Personal data removed before analysis: Yes / No; approved tool used: [tool]

# Sentiment analysis request · [Topic] · [Date]

**Requester:** [PM]   **Decision it informs:** [Decision]   **Due:** [Date]

## Question
[For example: "What are the main sources of frustration among mid-market customers in the last two quarters, and are they linked to lower retention?"]

## Sources and window
| Source | Connector or export | Time window | Approx. volume | Data restrictions |
|---|---|---|---|---|
| | | | | |

## Data handling
- Personal data to remove:
- Customer contract or residency limits checked by: [name]
- Approved tool:

## Codebook
Use codebook v[n] (attached). New themes may be proposed separately, not added.
Sentiment: positive / negative / mixed / neutral, per theme mentioned (aspect level), not only per item.

## Known hard cases to test
- Negation ("would be great if it stopped...")
- Sarcasm
- Positive question framing with negative answers
- Domain terms: [list]

## Validation
- Human-labeled sample: [n] random items, labeled by [names] before seeing machine output
- Report agreement, disagreements and theme shares (human vs machine)

## Outputs
- Theme table with counts, shares and trends
- Three verbatims per theme with source references
- Link to account retention and revenue where the source allows
- Confidence statement and known weaknesses

# Retention strategy decision · [Product] · [Date]

**Owner:** [PM]   **Decision maker:** [Name]   **Decision needed by:** [Date]

## Finding
[One paragraph: what the analysis showed, with the key numbers and their counts]

## Evidence
| Analysis | Result | Confidence | Link |
|---|---|---|---|
| Survival / hazard | | | |
| NRR and GRR decomposition | | | |
| CLV comparison | | | |
| Health score and pilot | | | |
| VoC themes | | | |

## Options
| Option | Type (retention / expansion / pricing / segment exit / channel) | Expected effect | Cost | Risk |
|---|---|---|---|---|
| 1. | | | | |
| 2. | | | | |
| 3. Do nothing | | | | |

## Recommendation
[Option], because [reason].

## Success measure and review
- Leading indicator: [for example usage survival in segment X]
- Lagging indicator: [for example GRR in segment X]
- Review date:
- What would make us reverse this:

Prompts

Goal: See when customers leave, by segment, for both contracts and usage.
Paste in (or connect the harness to): CRM contract records, billing history and product usage data for [product]; your agreed definitions; the segment list. This involves customer data, so use only an approved tool.

Analyze customer survival for [product] over [time window].

Definitions:
- Cohort start: [go-live / signature date]
- Contract churn: [definition]
- Usage churn: [definition, for example no active users for N days]
- Segments: [list]

Method:
1. Build a contract survival curve and a usage survival curve for each segment. Treat customers who haven't churned as still active (censored), never as churned and never dropped.
2. Show the hazard by period: in which months after start is the risk of leaving highest?
3. Show the gap between usage and contract survival for each segment.
4. Don't fit trend lines or project beyond the observed data.

Output: a chart per segment, a table of survival at 6, 12, 24 and 36 months with the number of accounts at risk at each point, and a one-paragraph finding per segment.

Checks: restate every definition you used; reconcile the starting customer count to [trusted number]; flag any segment with fewer than [N] accounts; list five example churned accounts per segment with their dates.

Check the output: confirm the counts at risk shrink over time but never include churned accounts; open two of the example accounts in the CRM and confirm their dates. If any curve rises, something is wrong.

Goal: Find out whether expansion is masking churn.
Paste in (or connect): ARR or billing by account for the last [13+] months, segment attributes, and finance's reported NRR and ARR. Use an approved tool.

For customers active on [date 12 months ago], calculate by segment:
- Starting recurring revenue
- Expansion (seats, environments, new products, price increases), listed separately
- Contraction
- Churn (customers lost entirely)
- NRR, GRR and logo retention

Present an NRR bridge per segment and a summary table with account counts.

Then:
1. Name the accounts contributing most to expansion and to churn in each segment, and what share of the segment's expansion the top five accounts represent.
2. Tell me whether any segment's NRR depends on a few large accounts.
3. Argue the strongest case that the overall NRR is less healthy than it looks.

Checks: reconcile overall NRR and ending ARR to finance's figures [paste] within [x]%; if they don't reconcile, explain the difference before anything else.

Check the output: if the reconciliation is off, stop and resolve definitions with finance. Read the "less healthy than it looks" argument and decide whether it's true.

Goal: Compare what customers are worth across segments after the cost to serve.
Paste in: survival or retention by segment (from your survival analysis), revenue per account, gross margin after inference and hosting, cost-to-serve estimates, partner margins, acquisition cost by channel and finance's discount rate.

Build a customer lifetime value comparison for these segments: [list].

Rules:
- Use contribution margin after inference, hosting, support, customer success and partner margin, not revenue.
- Use segment-specific retention by year from the survival data, not one blended rate.
- Cap expansion at [x]% per year, decaying to zero by year [n].
- Present low, base and high cases for each segment, and say which assumptions define each case.
- Show payback period beside CLV.

Then name the two assumptions that move the comparison most, and tell me whether the ranking of segments changes under any plausible combination of assumptions.

Use supplied inputs and sources. List missing inputs as gaps.

Check the output: recompute one segment's base case by hand at a rough level to confirm the order of magnitude. If the ranking of segments flips across the range, report the comparison as uncertain.

Goal: Find out whether the health score would have warned you in time.
Paste in (or connect): health score history by account, renewal outcomes for the last [4 to 8] quarters, and the score definition. Use an approved tool.

For every renewal in [period], find each account's health score and each component's value 90 days before the renewal date.

Report:
1. A table of outcome (renewed, contracted, churned) against band (green, yellow, red) 90 days before, with counts.
2. Which components best separated churned from renewed accounts.
3. Any signal that appears only shortly before churn (for example cancellation requests, data exports) and may be leaking the outcome.
4. Churned accounts that were green 90 days before: list them and what the data shows about why the score missed them.
5. Suggested changes to weights or components, each with the evidence for it.

Check the output: read three of the "green but churned" accounts yourself in the CRM and CS notes. Reject any suggested signal that only appears after a customer has decided to leave.

Goal: Turn tickets, transcripts and survey comments into themes you can trust.
Paste in (or connect): the text sources for [period], the codebook, and a human-labeled sample of [100 to 200] items. Remove personal data first, check customer agreements, and use only an approved tool.

Code each item against the codebook below. For each item:
- Assign one or more themes from the codebook.
- Give sentiment per theme mentioned (positive, negative, mixed, neutral), not only for the whole item.
- Quote the phrase that supports each label.

Rules:
- Use only codebook themes. Propose new themes in a separate list with five example items each; don't add them to the results.
- Watch for negation ("would be great if it stopped..."), sarcasm and positively framed questions with negative answers.
- If an item is ambiguous, label it "unclear" rather than guessing.

Then compare your labels with the human-labeled sample: report agreement per theme, list every disagreement, and compare theme shares between your coding and the human coding.

[codebook]
[human-labeled sample]

Check the output: read every disagreement. If any theme's share differs materially between human and machine coding, revise the codebook with examples and rerun before reporting trends.

Goal: Find which customer problems cost revenue.
Paste in (or connect): the validated theme coding by account, and ARR, GRR, health band and renewal date by account. Use an approved tool.

For each theme:
1. Count the accounts raising it and the ARR they represent.
2. Compare GRR, health band and churn rate for accounts raising the theme with accounts that don't.
3. Show the ARR up for renewal in the next two quarters among accounts raising it.
4. Give three verbatims with source references.

Rank themes by revenue at risk, not by volume of mentions. Flag any theme where the account count is below [N] as too small to conclude.

Then argue the strongest alternative explanation for the top theme's link to churn (for example, both caused by segment or account size).

Check the output: an association between a theme and churn doesn't prove the theme causes churn. Check the alternative explanation against segment-level data before putting the theme on the roadmap.

Goal: Move from findings to two or three decisions.
Paste in: your retention report, CLV brief, health score validation, pilot results and VoC report.

Act as a skeptical portfolio reviewer. From the material provided for [product]:

1. List the three findings with the largest revenue or margin implications, with the evidence and confidence for each.
2. For each finding, propose options across these types: retention bet, expansion bet, pricing or packaging change, segment or channel change, do nothing.
3. Estimate the size of the prize for each option as a range, showing the arithmetic and assumptions.
4. Recommend one option per finding and state what result in the next two quarters would make us reverse it.
5. Name the finding most likely to be wrong and why.

Use sourced numbers from the material. Label estimates and state their assumptions.

Check the output: trace every number back to the source reports. Cut any recommendation that doesn't name a leading indicator you can actually measure.

Anti-patterns

Checklist

• Churn, active, revenue base and cohort start are defined in writing and agreed with finance and customer success.
• Contract and usage survival curves exist by segment, with current customers treated as censored, and the hazard view shows where risk concentrates.
• Cohort triangles cover logo, user and revenue retention.
• NRR, GRR and logo retention are reported together by segment, with an NRR bridge that reconciles to finance.
• Any NRR comparison with another company states both definitions.
• The CLV model uses contribution margin after inference and cost to serve, segment retention and capped expansion, and shows ranges and payback.
• The health score has components, bands, playbooks and an owner, is free of leaking signals and has been back-tested against at least four quarters of renewals.
• A randomized retention pilot has informed whom customer success targets.
• Sentiment is reported by account and role, beside retention.
• LLM text coding is validated on a human-labeled sample, with theme shares compared and verbatims attached, using approved tools and data stripped of personal details.
• Each analysis ends in a decision with a leading indicator and a review date in the decision log.