Apexcloud Retention Ops

ApexCloud Retention Ops

This skill helps solve structured ApexCloud retention-operations tasks. The work is mostly data reconciliation: gather the requested records from the public API, apply strict exclusions and date windows, then emit JSON that exactly matches the prompt or template.

First Moves

1. Read the user prompt and any input/payloads/answer_template.json before querying data. The template is the schema source of truth when it exists, including fields such as policy_codes that may not be repeated in the prose.
2. Use the environment entrypoint from environment_access.md when present. If a prompt says http://127.0.0.1:8074 but the evaluation supplies a remote GDPEVO_ENV_BASE_URL, replace the local URL with the remote base URL and keep the same paths.
3. Query only the endpoint families named by the task, plus directly needed lookup endpoints such as /api/accounts.
4. Filter to the exact account IDs, regions, dates, quarters, events, and candidate IDs named in the prompt. Do not expand the portfolio unless the prompt says all regions or all accounts.
5. Return JSON only. No Markdown, explanations, comments, trailing commas, or extra top-level keys.

Endpoint Habits

Use curl, a short script, or another structured client. Parse JSON/CSV as data rather than copying terminal text by hand.

Common endpoints:

• GET /api/accounts returns {accounts, count} for all CRM accounts.
• GET /api/accounts/<account_id> returns one account profile.
• GET /api/accounts/<account_id>/metrics?start=YYYY-MM&end=YYYY-MM returns monthly account metrics.
• GET /api/accounts/<account_id>/tickets?start=YYYY-MM-DD&end=YYYY-MM-DD returns ticket rows.
• GET /api/accounts/<account_id>/nps?start=YYYY-MM-DD&end=YYYY-MM-DD returns NPS survey responses.
• GET /api/billing/snapshots?account_id=<id> returns billing snapshots; filter client-side to the requested as_of date or latest posted snapshot at or before the assessment date.
• GET /api/finance/ar-aging?as_of=YYYY-MM-DD returns A/R aging rows.
• GET /api/opportunities?start=YYYY-MM-DD&end=YYYY-MM-DD returns opportunities; verify close_date and state client-side.
• GET /api/hr/summary?quarter=YYYY-QN returns HR rows by region.
• GET /api/events/performance?event=<event_id>&quarter=YYYY-QN returns event context.
• GET /exports/churn/train.csv, /exports/churn/validation.csv, and /exports/churn/candidates.csv return churn-model CSVs.

The API may return broader data than the query implies. Always re-filter locally on the relevant field (month, created_date, response_date, close_date, as_of, quarter, region, account_id, or customer_id).

Source Rules

Account profile:

• Use billing_arr_current, crm_arr, contract_tenure_months, renewal_date, segment, region, legal_name, display_name, and account_aliases for account context.
• When a task asks for current ARR or revenue exposure, prefer posted billing snapshot ARR for the requested as-of date. Fall back to account billing_arr_current only if no matching snapshot exists.
• Treat lower tenure as higher churn risk; in model-check fields this means tenure risk direction is usually negative unless a task explicitly defines the opposite.

Metrics:

• Monthly QBR revenue comes from recognized_revenue in account metrics.
• Monthly support count comes from support_ticket_count unless the task explicitly asks for cleaned ticket rows.
• SLA percentage comes from sla_compliance.
• Metrics nps_score is useful for month-by-month tables, but latest sentiment should come from NPS responses when the task asks for latest NPS.
• Usage trend is the latest product_usage minus the first product_usage in the requested period.

Tickets:

• A clean ticket excludes rows where is_duplicate is true or is_spam is true.
• SLA health should consider both first_response_sla_met and resolution_sla_met; any false value is a service-risk signal.
• Count tickets only inside the requested date range, inclusive.

NPS:

• Exclude retracted responses.
• Latest NPS is the valid response with the greatest response_date in the period.
• For NPS trend/reason codes, compare latest valid score to the earliest valid score in the period.
• Use null where the schema allows null and there is no valid response; do not invent a score.

Billing and A/R:

• Posted billing snapshots are the most reliable ARR source for as-of reporting.
• A/R current is not overdue. Generic overdue balance is 1_30 + 31_60 + 61_90 + 90_plus.
• If the prompt says "older aging buckets", use 61_90 + 90_plus as the trigger/scope for follow-up inclusion, while still reporting the overdue balance requested by the schema.
• Link A/R rows to CRM accounts through the account ID embedded in aging_id when present (AR-<account_id>-...), otherwise match customer_name to account legal_name, display_name, or aliases.
• Mark unmatched A/R customers as unlinked and use account_id: null if the schema asks for link status.

Opportunities:

• Filter by close_date within the task window.
• Open pipeline includes state: "open" opportunities.
• Won revenue/count uses closed opportunities with stage: "Closed Won" or equivalent won state.
• Lost count uses stage: "Closed Lost" or equivalent lost state.
• Top open product line is the product line with the greatest summed open amount; use a deterministic tie-break such as alphabetical order.

HR and events:

• If the task says all regions, sum numeric HR totals across regions for counts and money fields. Average only when the prompt asks for a rate.
• For event context, use the exact requested event_id and quarter, then report event_orders and event_revenue unless the schema asks for a different field.

Churn exports:

• Train the model on train.csv, validate on validation.csv, and score candidates.csv.
• Drop identifier columns (customer_id) and the target (Churn) from features. Count raw predictor columns unless the task explicitly asks for encoded dimensionality.
• Encode categorical fields consistently across train, validation, and candidates; keep numeric fields numeric.
• Report validation row counts from the CSVs, accuracy as a percentage to one decimal, and probability outputs on the 0-to-1 scale to three decimals.
• Rank only the candidate/customer IDs named in the prompt; ignore other candidates.

Risk and Action Workflow

For renewal-risk queues and action boards, build one per-account record first, then rank. Useful fields include current ARR, days to renewal, latest NPS and NPS change, clean ticket count, SLA misses, usage change, overdue balance, old-bucket overdue balance, tenure, lifecycle status, and open expansion pipeline.

Reason-code guidance:

• overdue_receivable: any overdue balance, especially old-bucket A/R.
• low_tenure_high_churn: short contract tenure, typically around 18 months or less, especially with other risk signals.
• sla_degradation: SLA percentage declines, low SLA, or ticket rows with missed first-response/resolution SLA.
• nps_drop: latest NPS is low or materially lower than the first valid NPS in the period.
• usage_decline: product usage drops across the period or the export shows negative usage trend.
• renewal_window: renewal is near the assessment date, commonly within the next 90 days or recently passed.
• expansion_offset: meaningful open expansion pipeline offsets commercial risk.
• clean_billings: no meaningful overdue A/R and billing is clean.

Primary-action guidance:

• Use collections_followup when overdue A/R, especially old-bucket A/R, is the dominant issue.
• Use technical_recovery when SLA, tickets, or usage are the dominant issues.
• Use renewal_save when renewal timing plus churn/sentiment signals dominate.
• Use executive_qbr for strategic/high-value accounts needing executive alignment, especially with expansion or mixed signals.
• Use nurture_monitor for lower-risk accounts that still need monitoring.
• Use no_action only if the enum allows it and no material risk is present.

Risk ranking should be deterministic. Sort by severity first (critical, high, medium, low), then by primary action urgency, ARR/exposure, renewal urgency, and finally account_id as a stable tie-break. For board tasks that mention a "standard retention board order", use this risk/action ordering rather than alphabetical order.

Output Conventions

• Preserve every key from the prompt/template and omit keys that are not requested.
• Use exact controlled enum strings from the prompt/template. Do not create synonyms.
• Use integers for counts, ranks, and risk scores.
• Use numeric JSON values for money, percentages, probabilities, and counts, not formatted strings.
• Round currency to 2 decimals, percentages to 1 decimal, and churn probabilities to 3 decimals unless the prompt says otherwise.
• Sort arrays exactly as requested: top-N rankings by risk/probability, overdue follow-ups by customer_name when asked, QBR metrics in month order, and action boards in retention board order.
• Dates must be ISO strings from the prompt or source data.
• For summary fields, compute from the returned/reviewed population, not from the entire API dataset unless the prompt says all accounts/regions.

Policy-code blocks are methodology fields. If a codebook is not exposed by the API, choose the allowed code that corresponds to the workflow you actually used, keep it inside the provided enum, and use it consistently. Never leave placeholder strings such as RS-2|RS-6|RS-9.

Common Pitfalls

• Do not use local filesystem environment data or source code when the task provides a public API entrypoint.
• Do not trust the prompt's localhost host in a remote evaluation; swap only the host/base URL, not the API path.
• Do not count duplicate or spam tickets as clean tickets.
• Do not include retracted NPS responses.
• Do not treat A/R current as overdue.
• Do not report CRM ARR when the task asks for billing ARR/current revenue exposure and a billing snapshot exists.
• Do not include accounts, candidates, regions, or customers outside the requested scope.
• Do not let templates remain as placeholders; all enum-pipe strings and sample IDs must be replaced.
• Do not add explanatory text around the JSON; many graders parse the response directly.