Executive Summary
We are a ~$1.3M-revenue multi-channel retailer/wholesaler running on a heavily automated NetSuite estate. The diagnostic — validated and independently reviewed since v1 — points to one dominant conclusion: our biggest AI opportunity is in decisions, not data entry. Inventory turns below 1.0, bills paid a month early, a service queue worked by one person, and master data too thin for any model to reason with. These are decision failures that AI, with a human approving each action, can fix in months.
The strategy in one paragraph
Three waves in 12 months. Foundation + Wave 1 (0–90 days): purge test fixtures from production reporting, back-fill the four gating master-data fields with AI-proposed / human-approved updates, and ship three narrow assistants — replenishment recommendations for stocked positions, an AP payment-timing & anomaly screen, and a service triage copilot. Wave 2 (3–6 months): demand forecasting, inventory rebalancing from stores to the LA DC, generative product content, and conversational analytics for managers. Wave 3 (6–12 months): churn scoring, sales next-best-action, and supervised finance agents — only after Waves 1–2 have proven acceptance ≥ 90%.
Value at stake — measured, with formulas in §Value bridge
| Lever | One-time cash | Annual P&L |
|---|---|---|
| Inventory to 1.5 turns (from ~0.6) | $600–770K | $30–40K carrying cost |
| Pay vendors on terms, not 28 days early | $250K | $20K cost of capital |
| Dead stock liquidation (47 items) | $20K | $10K carrying cost |
| Service triage & drafting | — | ~150 h / yr + key-person risk |
| Master-data back-fill via AI | — | 53 h saved once; enables all models |
What we are asking the ELT to decide (§Decision page)
- Endorse the ten objectives and Wave 1 scope.
- Approve a six-week data-foundation sprint as a hard prerequisite.
- Adopt the AI operating model: human-in-the-loop for all writes, privacy by default, audited actions.
- Name business owners: Supply Chain, Controller/AP, Head of Support.
- Authorise removal of the 30 test-fixture invoices from production reporting.
What Changed Since v1
| Item | v1 said | v2 says | Why |
|---|---|---|---|
| Open A/R >90 days | 51% · $473K | 1 real invoice · $53K · 69 d | 31 open invoices validated as test fixtures (future createddate, no source SO) |
| 12-month revenue | $2.29M | $1.30M (967 docs) | 30 test docs worth $797K removed |
| Monthly volatility | 3.3× ($81K–$265K) | 2.05× ($60K–$123K, full months) | Spikes were test fixtures; Sep-26 partial excluded from range |
| Store share of inventory value | "two-thirds" (unsourced) | 65% · $722K (computed) | Independent reviewer flagged an unsupported figure; recomputed |
| Chicago DC "89% stocked out" | Framed as service failure | Dormant location — 0 sales/shipments in 12 mo | Reviewer challenge; verified by query |
| Vendor payment timing | Not measured | Median 28 days early, 995/1,005 payments | New measurement — a $250K float |
| "252 duplicate-pattern bill pairs" | Implied duplicates | Recurring standard-priced orders at 7 vendors; 1 same-day pair | Deeper look; detector needs vendor-specific baselines |
| Manual-entry load | Not measured | Business UI edits negligible; estate is script/workflow-fed | Reviewer: systemnote measures config edits, not transaction entry — conclusion reframed as "pursue decision automation" |
| Objective ranking | #1 Collections | #1 Replenishment, #2 AP timing, #3 Service; Collections → watch item | Follows from the above |
| Independent review | None | 10 claims audited by a second model; 6 confirmed, 4 challenged and corrected | Appendix D |
Method & Data Caveats
How the diagnostic was run
We interrogated the system of record directly: three parallel read-only research passes (order-to-cash; procure-to-pay & inventory; master data, service & automation) followed by ~15 targeted queries to validate anomalies and close gaps. Cycle times were measured through actual document links, not status fields. Master-data completeness is a field-by-field null count on active records. All aggregation ran server-side; no personal data left the system beyond aggregate counts.
Every headline claim was then packaged with its source data and formulas and sent to a different AI model acting as adversarial reviewer. Its verdicts, and what we changed in response, are in Appendix D.
What to trust
| Signal | Reliability | Why |
|---|---|---|
| Inventory positions, reorder coverage, value by location | High | Live inventoryitemlocations; no subsidiary-4 rows |
| Vendor payment timing vs due date | High | 1,005 bill→payment links with due dates |
| Master-data completeness | High | Direct null counts |
| Case aging, assignment, escalation | High | Real dates; n=87 is small — directional |
| Validated revenue & A/R | High | Test fixtures identified by 3 independent markers |
| Inventory turns / DIO | Medium | COGS estimated from revenue at 40–60% margin (range shown) |
| Process cycle times (0–3 day medians) | Low | Reflect bulk loading, not operations — not used |
Diagnostic Findings
Six areas, in order of the working capital and risk they represent after validation.
1 · Inventory & Replenishment — the largest lever
Fig 2 · On-hand value by location ($1,116,806) — hover for detail
| Fact | Value |
|---|---|
| Estimated turns (GM 40 / 50 / 60%) | 0.70 / 0.58 / 0.46 |
| Days inventory outstanding | 524 / 629 / 786 |
| Positions with reorder point / lead time / ever counted | 12 / 0 / 7 of 1,924 |
| Active items that sold in 12 months | 148 of 337 (44%) |
| Dead stock (on hand, no sale 12 mo) | 47 items · $39.9K |
| Stock-out positions at active sites (LA, SF, NY, Miami) | 98 of 636 (15%) |
| Backordered | 160 units · 5 items · $3.1K cost |
| Work orders / assembly items | 24 / 9 |
2 · Procure-to-Pay — paying early, controlling loosely
Fig 3 · Document flow, all-time (bills → payments) — hover for detail
| Fact | Value |
|---|---|
| PO volume, 12 months | 400 · $1.23M · avg $3,066 |
| Top-10 vendor share of PO spend | 98.3% (top 4 = 82%) |
| Vendors with PO activity / active / with terms | 16 / 76 / 52 |
| Discount terms configured | 2 |
| Working capital float from early payment | ≈ $250K (one-time) |
| Same vendor + amount pairs ≤30 d | 252 · $656K · 1 same-day |
| Open POs / open A/P | 35 · $41K / 13 · $184K (98% current) |
3 · Customer Service — slow, concentrated, unclassified
Fig 4 · Open cases by age (days), n=40 · closed-case time-to-close, n=47
| Fact | Value |
|---|---|
| Cases total / open / closed | 87 / 40 / 47 |
| Escalated | 21 (24%) |
| One agent's share | 54% (47 of 87) |
| Unassigned | 11 (13%) |
| No category / no origin | 34% / 36% |
| Mix | Returns & exchanges 25 · Amazon 21 · Problem/Concern 10 |
4 · Sales & Revenue — validated baseline
Fig 1 · Validated monthly revenue ($K) with 3-month average — hover for detail
| CRM fact | Value |
|---|---|
| Customers with no sales rep | 207 of 273 (76%) |
| Open opportunities · win rate (closed) | 10 · $15.3K · 58% |
| Estimates converted / expired | 1 of 10 / 5 |
| Renewal-risk / churn-date populated | 0 / 0 of 273 |
5 · Master Data & Automation
| Gap (active records) | Missing |
|---|---|
| Customer · payment terms | 80% (218/273) |
| Customer · sales rep | 76% (207/273) |
| Customer · invoice e-mail routing | 100% |
| Item · class | 91% (306/337) |
| Item · UPC | 99.7% |
| Item · image / description | 53% / 28% |
| Vendor · terms / e-mail | 32% / 30% |
| Automation | Value |
|---|---|
| Script deployments active / total | 1,857 / 1,928 |
| Workflows released / total | 10 / 17 |
| Script errors 90 d (excl. AI agent) | 47 · 36 from one bundle |
| Admin-role share of login users | 29% (9 of 31) |
| Business-user UI edits, 90 d | negligible — estate is script/workflow-fed |
6 · Governance finding — test fixtures in production reporting
| Marker | Test-fixture invoices (31 open / 30 in 12-mo sales) | Real invoices |
|---|---|---|
| Memo | "TEST — <customer>" | blank or business text |
| Created date | 14 Sep 2026 — after analysis date | 9–17 Sep 2026, ≤ trandate + days |
| Source sales order link | none (0 of 31) | present (10–28 links) |
| Lines per invoice | 1–2 | 6–15 |
| Impact on reported open A/R | $790K of $928K (85%) | $138K |
| Impact on reported 12-mo revenue | $797K (38%) | $1.30M |
Value Bridge — with the arithmetic shown
Every figure below is derivable from the data in §Diagnostic. Assumptions are stated so they can be contested; ranges are given where an input is estimated.
| # | Lever | Formula | Low | Base | High | Type | Confidence |
|---|---|---|---|---|---|---|---|
| V1 | Inventory to 1.5 turns | On-hand − COGS / 1.5, COGS = $1.297M × (1 − GM), GM 40/50/60% | $598K | $685K | $771K | One-time cash | Medium — GM assumed |
| V2 | Carrying cost avoided on V1 | V1 × 5% cost of capital + storage/shrink (5%) | $30K | $34K | $39K | Annual | Medium |
| V3 | Pay to terms, not 28 days early | $3.25M paid × 28 / 365 (all-time links; ~12-mo equivalent ≈ $1.2M × 28/365 = $94K) | $94K | $250K | $250K | One-time cash | High on timing; base uses all-time volume |
| V4 | Cost of capital on V3 | V3 × 8% | $7.5K | $20K | Annual | High | |
| V5 | Early-pay discounts (alternative to V3, not additive) | 1% × annual paid, only where discount terms negotiated (2 exist today) | $0 | $12K | $32K | Annual | Low — terms must be negotiated |
| V6 | Dead-stock liquidation | $39.9K × 50% recovery; carrying 25%/yr avoided | $10K | $20K | $28K | One-time + $10K/yr | High |
| V7 | Service triage & drafting | ~261 cases/yr (annualised from erratic 4 months) × 35 min saved | 90 h | 152 h | 250 h | Annual hours | Medium — volume erratic |
| V8 | Master-data back-fill | 1,287 field values × 3 min manual vs 0.5 min AI-proposed/approved | — | 53 h saved | — | Enabler | High |
| V9 | Product content generation | 96 descriptions × 20 min + 180 image-gap triage × 5 min | — | 47 h + conversion uplift (unmeasured) | — | Annual | Medium |
| V10 | Real collections follow-up | 1 invoice $53,424 × 69 d × 8% | — | $0.8K | — | Watch | High |
Data Readiness by Objective
Which fields each AI use case needs, and how populated they are today. Green ≥ 90%, amber 50–89%, red < 50%. Red cells are the foundation-sprint backlog.
Strategic Objectives — re-ranked after validation
Scored 1–5 on Value, Feasibility (data readiness · effort · change load) and Confidence in the evidence.
Data Foundation & AI Governance Prerequisite · weeks 0–6
- Flag/isolate 30 test-fixture invoices; exclude from reports and model inputs
- Back-fill: terms (218), sales rep (207), item class (306), reorder point + lead time (556 stocked positions) — AI proposes from history, human approves in bulk
- Case intake: category and origin mandatory
- Retire the plain-text credential field on vendor records
- AI policy ratified: human approval on every write, dry-run diff, read-only analytics by default
- Privacy mode default-on for entity analysis
- Audit trail reviewed monthly; acceptance rate published
- Terms / rep / class ≥ 95%
- Reorder point on 100% of stocked positions at active sites
- Owners named; policy signed
Intelligent Replenishment Wave 1 → 2
Two stages. Wave 1: statistical reorder-point and preferred-stock recommendations for the 556 stocked positions at four active sites, from 12-month sell-through; rebalancing proposals from stores to LA DC; dead-stock markdown list. Wave 2: demand forecasting by SKU × location with seasonality once a second year of history exists; Chicago DC decision (decommission or restock).
- Sales history for 148 items — thin for 189 others (treat as slow movers)
- Lead times: none — capture from top 4 vendors (82% of spend) in the sprint
- Item class needed for category-level pooling
- Turns 0.6 → 1.0 by month 6, 1.5 by month 12
- Active-site stock-outs 98 → <40 positions
- Dead stock $40K → <$15K
- Forecast MAPE tracked from Wave 2
AP Payment Timing & Anomaly Screen Wave 1
Schedule payments to due date (or to a negotiated discount date) instead of 3 days after receipt; screen every bill pre-approval for duplicates using vendor-specific baselines, missing PO linkage and price drift vs PO.
- Complete: bills, POs, links, due dates all present
- Vendor terms on 52 of 76 — fill the 24 in the sprint
- Median days-early 28 → ≤ 3
- Non-PO bill share 31% → <10%
- 100% of bills screened; 0 confirmed duplicate payments
- Discount terms negotiated with ≥ 2 of top-4 vendors
Service Triage & Response Copilot Wave 1
Classify, prioritise and route every case at intake; draft policy-grounded first responses for returns and marketplace disputes; summarise threads; re-rank the backlog by escalation likelihood.
- 87 cases, 47 closed — enough for few-shot with policy documents, not for training
- Needs returns policy and Amazon SOP as reference texts
- Median close 33 d → <10 d · escalation 24% → <10%
- Largest agent share 54% → <40% · unassigned >24 h → 0
Generative Product Content Wave 2
Channel-specific descriptions, attribute tags and image-gap worklists for the 96 items without descriptions and 180 without images; matrix-aware (20 parents, 76 children). Feeds the eCommerce channel and the Item Catalog interface already built on the account.
Conversational Analytics for Managers Wave 2
Roll the in-NetSuite AI assistant out to department heads in a read-only Analyst posture with account-specific skills, and automate the monthly management pack from the same source. Target ≥ 8 weekly active managers; −50% ad-hoc report tickets.
Churn & Renewal-Risk Scoring Wave 3
Populate the dormant renewal-risk / churn-date fields from purchase cadence, payment behaviour and service history; alert reps to accounts going quiet. Requires rep coverage (Objective 0) and two years of cadence.
Sales Next-Best-Action & Quote Assist Wave 3
Reorder and cross-sell prompts from basket affinity; quote drafting with margin guardrails; follow-up nudges before estimates expire (5 of 10 expired, 1 converted). Pipeline data is too thin to justify earlier sequencing.
Supervised Finance Agents Wave 3
Month-end close pre-work, bank reconciliation matching and continuous controls monitoring — process definitions exist already. Graduates from recommender to agent only after 6+ months of audit trail at ≥ 90% acceptance. Every journal remains human-approved.
Vendor Intelligence Wave 3
Price benchmarking, lead-time learning and OTIF scoring for the four vendors that are 82% of spend; negotiation briefs ahead of renewals (including the discount terms Objective 2 needs).
Collections — watch item, not an objective Monitor
After validation, real open A/R is $138K with one invoice 69 days past due ($53K). The Controller follows up this month; an A/R aging alert (any invoice >30 days or >$25K past due) is added to the scoreboard. If the real tail grows, Collections Intelligence from v1 is ready to activate.
Prioritization
Fig 5 · Value × feasibility — bubble size = confidence; click a bubble to jump to the objective
| # | Objective | Value | Feas. | Conf. | Value at stake | Wave |
|---|---|---|---|---|---|---|
| 0 | Data Foundation & Governance | — | 5 | 5 | Enables all; removes $800K of fiction from reporting | Pre |
| 1 | Intelligent Replenishment | 5 | 3 | 4 | $600–770K cash · $30–40K/yr | 1→2 |
| 2 | AP Payment Timing & Anomaly Screen | 4 | 5 | 5 | $94–250K cash · $8–20K/yr | 1 |
| 3 | Service Triage Copilot | 3 | 4 | 4 | ~150 h/yr · key-person risk | 1 |
| 4 | Generative Product Content | 3 | 5 | 5 | 47 h · conversion uplift | 2 |
| 5 | Conversational Analytics | 3 | 5 | 4 | Manager self-service | 2 |
| 6 | Churn & Renewal-Risk | 4 | 3 | 3 | Retention on concentrated base | 3 |
| 7 | Sales Next-Best-Action | 3 | 3 | 2 | Conversion 10% → ? | 3 |
| 8 | Supervised Finance Agents | 4 | 2 | 3 | Close effort · controls | 3 |
| 9 | Vendor Intelligence | 2 | 3 | 3 | Price / terms on 82% of spend | 3 |
Use-Case One-Pagers — Wave 1
Current state → future state for the three Wave 1 objectives. Blue = AI does; amber = human decides; green = written to NetSuite after approval.
1 · Intelligent Replenishment
Today
With AI
2 · AP Payment Timing & Anomaly Screen
Today
With AI
3 · Service Triage & Response Copilot
Today
With AI
Delivery Roadmap
Fig 6 · 12-month plan — diamonds are wave-exit reviews
Foundation Sprint
- Test fixtures flagged and excluded
- AI-proposed back-fill approved in bulk: terms, reps, class, ROP/lead time
- Lead times from top-4 vendors
- Case intake fields mandatory
- AI policy signed; owners named
Wave 1 · Assist
- AP screen live week 3; payment-to-terms with top-4 vendors week 6
- Replenishment recommendations (no writes) week 4; writes week 8
- Service classification week 4; reply drafts week 8
- Weekly acceptance review
Wave 2 · Optimise
- Forecasting pilot top-50 SKUs → all positions
- Store→DC rebalancing; Chicago decision
- Product content for eCommerce
- Analytics rollout to managers; monthly pack automated
Wave 3 · Predict & Delegate
- Churn scores populate CRM fields
- Next-best-action for reps
- Supervised finance agents (close, bank rec, controls)
- Vendor briefs ahead of renewals
Fig 7 · AI maturity — today vs 12-month target
Resourcing
| Role | Wave 1 | Wave 2–3 |
|---|---|---|
| Head of AI | 60% | 50% |
| NetSuite developer / analyst | 50% | 50% |
| Data steward (Finance Ops) | 50% | 20% |
| Business owners (3 → 6) | 2 h/week each | 2 h/week each |
| Platform | Existing embedded AI agent + native NetSuite; cost is model usage. Wave 3 agents = Scheduled/Map-Reduce scripts we already govern. | |
What we will not do in FY2027
- No customer-facing chatbot or AI-sent customer e-mail
- No unattended writes to the ledger, customers or items
- No model training on customer personal data; analytics run tokenised
- No new AI platform purchase before Wave 2 results
- No AI use case without a named business owner and a baseline KPI
Stakeholder Impact Map
| Role | What changes | What they gain | Enablement | Wave |
|---|---|---|---|---|
| Inventory planner / buyers | Weekly recommendation review replaces ad-hoc PO judgement; reorder points become the system of record | Fewer stock-outs to fight; reasoning shown for every proposal | 2-hour workshop; 30-day recommendation-only period | 1→2 |
| AP clerk / Controller | Bills screened before approval; payment runs scheduled to terms | Duplicate risk removed; $250K float recovered; less rework | 1-hour walkthrough; hold-review checklist | 1 |
| Support agents (3) | Cases arrive classified and routed; reply drafts available | Balanced load; faster first response; less escalation firefighting | Policy documents curated; draft-acceptance feedback loop | 1 |
| Store managers | Store stock rebalanced toward DC; ask-the-data access | Right stock, fewer transfers by phone | Analyst-profile access; 30-min intro | 2 |
| Merchandising / eCommerce | Content drafts generated; image gaps prioritised by revenue | Full catalogue coverage in weeks | Style guide as reference skill | 2 |
| Sales reps | Assigned to all accounts; risk and next-best-action alerts | Prioritised book; early warning on quiet accounts | CRM hygiene expectations | 0, 3 |
| Finance leadership | Test fixtures out of reporting; monthly pack automated; agent proposals to review | Trustworthy numbers; close effort down | Steering committee monthly | 0, 2, 3 |
| IT / NetSuite admin | Guardrail ownership; admin-role reduction paired with permissions audit | Audit trail for every AI action | Runbook; monthly telemetry review | All |
Operating Model & Governance
1 · Human in the loop
No AI action writes to the ledger, a customer, an item or a payment without a human approving a before→after diff. Deletions need an explicit click-through that cannot be scripted. Analytics run read-only.
2 · Privacy by default
Customer, vendor and employee identities are tokenised before any model sees them for analysis. Content generation uses no personal data.
3 · Auditable
Every AI action is logged: who, what, tool, outcome, purpose. Monthly review of acceptance and drift; results published to the steering committee.
4 · Least privilege
AI runs under the invoking user's role, never a super-user. 29% of login users are Administrators — paired with a permissions audit, because AI amplifies whatever access it has.
5 · Measure before automate
Every objective ships as a recommender first. Graduation to a supervised agent requires ≥ 90% acceptance over 8 weeks and reversible actions only.
6 · Independent review
Material analyses are cross-checked by a second model before reaching the ELT — this document was, and it changed the headline. Reviews and provenance are printed in appendices.
Success Scoreboard
| KPI | Baseline · 10 Sep 2026 | Target | By | Obj. |
|---|---|---|---|---|
| Inventory turns (est.) | ~0.6 | 1.0 → 1.5 | M6 → M12 | 1 |
| Stock positions with reorder point (active sites) | 0.6% | 100% | Foundation | 0/1 |
| Stock-out positions at active sites | 98 | <40 | End Wave 2 | 1 |
| Dead-stock value | $39.9K | <$15K | End Wave 2 | 1 |
| Store share of inventory value | 65% | <50% | End Wave 2 | 1 |
| Vendor payments — median days before due | 28 | ≤ 3 | End Wave 1 | 2 |
| Vendor bills without PO | 31% | <10% | End Wave 1 | 2 |
| Bills screened pre-approval | 0% | 100% | End Wave 1 | 2 |
| Case median time-to-close / escalation | 33 d / 24% | <10 d / <10% | End Wave 1 | 3 |
| Largest agent share / unassigned >24 h | 54% / 11 | <40% / 0 | End Wave 1 | 3 |
| Customers with terms / rep · items with class | 20% / 24% · 9% | ≥95% each | Foundation | 0 |
| Test-fixture documents in production reports | 30 · $797K | 0 | Foundation | 0 |
| Item description / image coverage | 72% / 47% | 100% / 90% (top-100) | End Wave 2 | 4 |
| Weekly active managers on AI analytics | 1 | ≥ 8 | End Wave 2 | 5 |
| Customers with renewal-risk score | 0% | 100% | End Wave 3 | 6 |
| Real A/R past due >30 d (watch) | 1 · $53K | 0 | Monthly | W |
| AI recommendation acceptance (all) | n/a | ≥ 90% before any automation | Continuous | Gov. |
Decision Page
Motions for the Executive Leadership Team — 10 September 2026
- Endorse the AI Implementation Strategy v2 — ten objectives in three waves, with Intelligent Replenishment, AP Payment Timing & Anomaly Screen, and Service Triage Copilot as Wave 1.
- Approve a six-week Data Foundation Sprint (test-fixture isolation; back-fill of terms, sales rep, item class, reorder point/lead time; mandatory case intake fields) as a hard prerequisite, staffed at 0.5 FTE data steward + 0.5 FTE developer.
- Adopt the AI Operating Model: human approval on every write, privacy by default, audit trail reviewed monthly, ≥ 90% acceptance before any automation, independent review of material analyses.
- Appoint business owners: VP Supply Chain (Obj. 1), Controller (Obj. 2), Head of Support (Obj. 3); and a monthly AI Steering review (CFO, COO, Head of Support, Head of AI).
- Authorise Finance to flag and exclude the 30 test-fixture invoices ($797K) from production reporting, and to follow up the one real past-due invoice ($53K) this month.
Wave 1 exit review: ~10 December 2026. Wave 2 exit: ~March 2027. FY2028 strategy refresh: September 2027.
Appendix
A · Data sources
| Area | Tables / fields |
|---|---|
| Sales & A/R | transaction (CustInvc, CashSale; foreignamountunpaid, duedate, memo, createddate), transactionline.location/subsidiary |
| Document links | nexttransactionlinelink (OrdBill, ShipRcpt, Payment) — used for P2P flow, payment timing and fixture validation |
| Inventory | inventoryitemlocations (onhandvaluemli, quantityavailable, quantitybackordered, reorderpoint, leadtime, lastinvtcountdate), location.subsidiary |
| Service | supportcase (status, startdate, enddate, assigned, origin, category) |
| Master data | customer (terms, salesrep, custentity_renewal_risk), item (class, description, upccode, custitem_atlas_item_image), vendor (terms), term.discountpercent |
| Automation | scriptdeployment × script, workflow, scriptnote, systemnote (90 d) |
B · Limitations
- COGS is estimated from validated revenue at 40–60% gross margin; turns/DIO are shown as a range. Actual COGS by subsidiary would tighten V1.
- Payment-timing float (V3 base) uses all-time bill→payment volume; the 12-month-equivalent low case is shown.
- Case volume (87) is small and erratic month to month; service figures are directional.
- The non-PO bill share mixes all-time bills with 12-month POs and may overstate the current gap.
- Cycle-time medians of 0–3 days across all flows reflect bulk loading; not used for conclusions.
- Two item-level custom fields intermittently fail in aggregate SuiteQL; item completeness comes from one verified aggregate query.
C · Existing assets to reuse
Process definitions already in the account: data-quality-sweep, forensic-anomaly-sweep, continuous-controls-monitor, month-end-close, bank-reconciliation, monthly-management-reporting, permissions-audit, customization-retirement, agent-builder. Skills: privacy-analysis-protocol, house-suiteql-style. Interfaces: Item Catalog (SuiteCard).
D · Independent review — verdicts and actions
Ten claims with source data and formulas were sent to a second model (claude-opus-4-8) under an adversarial-reviewer charter. Provenance: primary analysis by the embedded Sonar AI agent; independent review via user-supplied API key; both are LLM-generated — a sibling-model review checks arithmetic, logic and interpretation but is not fully independent and is not a substitute for human review.