Questions. Answered.
Organized by theme. Every answer links to the canonical source — ADR, roadmap row, or pricing line.
Frequently asked questions
Q1. What does EXSEL actually do?
A. Three things: (1) cross-platform product intelligence across 9 connectors (Amazon, Shopify, TikTok Shop, Walmart, Etsy, eBay, Pinterest, Meta Ads, Google Ads). (2) Six AI specialists you unlock individually — PPC Automation, Account Health, Inventory Forecast, Review Management, Content Gen, Cross-Channel Repurpose. (3) Your ClientConductor — a merchant-persistent agent that learns your store over time. Full catalog at `/pricing`.
Q2. What's a "ClientConductor"?
A. One persistent AI agent per merchant account. Tenant-isolated memory, per-tenant LoRA adapter after 500 causally-confirmed trajectories, refuses to confabulate below a 0.35 confidence threshold. Architecture at `/agents`. ADR-087.
Q3. What's a "VisitorConcierge"?
A. The cookie-scoped anonymous agent on our marketing pages. 90-day TTL, reflex-class model, read-only on a shared public-knowledge index, never touches tenant data. The chat widget at the bottom-right of this page IS a VisitorConcierge instance. ADR-087.
Q4. Do I need the Companion desktop app?
A. Only if you sell on Walmart (Seller Center has no public API for the writes you need) OR if you want tamper-evident audit mode on any AI specialist. Otherwise, your browser + our orchestrator is enough. See `/companion`. ADR-045.
Q5. Can the Conductor actually take actions, or does it just suggest?
A. Both — you choose per AI specialist. Default on every unlock = **manual** (Conductor drafts, you approve). Opt-in per AI specialist to **2FA-gated autonomous** (Conductor executes, each action cleared through a 2FA pre-confirm) or **tamper-evident audit** (autonomous + hash-chained log, Companion required). ADR-034 ModeGate.
Q6. Which platforms are you on?
A. Today (9 connectors): Amazon, Shopify, TikTok Shop, Walmart, Etsy, eBay, Pinterest, Meta Ads, Google Ads. Roadmap additions tracked in UNIFIED_ROADMAP.md on GitHub.
Q7. Is there an API?
A. Yes — every Core service exposes both a typed TS/Python class AND an HTTP endpoint `POST /api/core/<service>/*`, tenant-scoped via capability token. OpenAPI spec + generated SDKs (TS, Python) on `/api-docs`. G49 rule 3.
Q8. Where are the plans?
A. There aren't any. Start free; add only the capabilities you need. You'll never pay for a tier you don't use. ADR-046-v2.
Q9. How does the free plan work?
A. Forever-free. One connector of your choice, read access + the full dashboard + 50 safety-floor actions/month (review replies under SLA, suppression ACKs, critical appeals). Your ClientConductor is active in reflex mode. No card. ADR-046-v2 §2.1.
Q10. Will I get surprise bills?
A. No. Every unlock has a hard monthly cap by default. Opt-in overage is capped at 2× the base monthly price. You can never be billed beyond what you agreed to. 15-overage-caps.md.
Q11. What's the annual prepay discount?
A. 20% off per unlock SKU. Toggle on the pricing drawer. 7-day grace for cash refund; after that, prorated credit toward other unlocks. 14-annual-prepay.md.
Q12. Is there a bundle discount?
A. 4+ active unlock SKUs = 20% off cart total. Stackable with annual prepay. Worst-case margin still clears 70% (18-§G). Stacking beyond these two requires founder approval.
Q13. What happened to the AS-1 / AS-2 / AS-3 / AS-4 tiers I read about?
A. Retired 2026-04-17 PM. Autopilot is now a per-specialist toggle (Manual / 2FA-gated / Audit) inside ModeGate — not a pricing tier. Unlock an AI specialist, flip the mode switch. ADR-046-v2.
Q14. How do you not lose money on day one?
A. Every unlock clears a ≥40% gross-margin floor at its quota cap, even with annual prepay + bundle stacked. Free plan runs ~$0.14/merchant/month variable cost. Day-1-no-loss model at 18-unit-economics.md.
Q15. Can I cancel?
A. Monthly: anytime. Annual: 7-day grace for cash refund, then prorated credit. No retention dark patterns.
Q16. I'm on a legacy 5-tier plan. What happens?
A. Grandfather migration: 12 months of equivalent-unlock entitlement at no additional charge. 60 / 30 / 7-day communication cadence before any change. No silent downgrades. 12-v2 §5.
Q17. How do you enforce zero-suspension?
A. Architecturally, not as a claim. Every action passes through **ScopeGate** (ADR-054 — per-platform YAML allowlist, enforced at the primitive level) and **ModeGate** (ADR-034 — manual/2FA/audit). Amazon Seller Central UI automation is **forbidden in code**. Meta Business Suite UI automation is **forbidden**. TikTok multi-account Symphony rotation is **deprecated** (ADR-057). Quarterly ToS review enforced in CI.
Q18. How is my tenant isolated from other tenants?
A. Type-system rule, not a policy. `AgenticRAG.recall(tenantId, query)` requires `tenantId`; internal Qdrant queries wrap every call with `filter: {tenant_id: $tenantId}`. Missing tenantId → `TenantIsolationError`, fail-closed, logged to audit. Per-tenant LoRA adapters. Per-tenant causal attribution store. ADR-055 rule 6 + ADR-086 rule 3.
Q19. Do you ever share data between merchants?
A. Only through the `AggregateOnly<T>` wrapper type — a compile-enforced PII-stripping aggregation. Used only for Reasoning Bank priors and board-facing GovernorAI analytics. GDPR Art. 32.
Q20. Is there an audit log?
A. Yes — `companion_audit_log_v2` with **hash-chained** entries (each row's hash incorporates the previous). Weekly signed evidence bundle exportable as a legal artifact. Required for AS-3/AS-4 regulated workloads. ADR-044.
Q21. GDPR / CCPA / SOC-2?
A. GDPR Art. 32 enforced architecturally; SCCs available for EU merchants. CCPA "do not sell or share" honored. SOC-2 Type II in progress (evidence collection active; target 2026-Q4). Interim: penetration-test reports + control inventory under NDA. See `/security`.
Q22. Do you take my platform passwords?
A. Never. We use OAuth where platforms provide it and Companion (your device, your session) for UI-driven actions where they don't. Credentials never leave your device on a Companion flow.
Q23. How do I delete my data?
A. Tenant data: `/settings/delete`, 72h hard delete, written confirmation. Visitor cookie: "Forget me" link in footer, immediate. Audit log: 13-month default, 7y for regulated merchants (legal-hold compatible). See `/security#data-handling`.
Q24. How does continuous learning actually work?
A. Every `BaseAgent.call()` emits a trajectory (state, action, outcome, reward, causal_confirmed bool). Five learning consumers fan out: LoRA trainer (nightly), reranker priors (online), causal DAG weights (weekly), forecast calibration (weekly), safety policy (online), Reasoning Bank priors (weekly). **Only DoWhy-confirmed trajectories** (p < 0.05 across 3 refuters + Cohen's |d| ≥ 0.2) train the LoRA. Raw approvals log but don't train — this makes reward-hacking impossible. ADR-088.
Q25. When does my ClientConductor actually get better?
A. 4-band cold-start (P0.R1): 0 trajectories → base + generic priors. 100 trajectories → reranker priors active. 300 trajectories → safety + forecast calibration. 500+ causally-confirmed → per-tenant LoRA adapter live via SGLang hot-swap. No restart, no downtime.
Q26. What is "AgenticRAG"?
A. Our Core retrieval primitive. 6-step loop: Recall (hybrid BM42 + BGE-M3) → Rerank (Cloudflare BGE-reranker-v2-m3) → **Confidence gate** (top-hit < 0.35 → refuse) → Agentic multi-hop (up to 3) → Synthesis → Write-back. Mandatory for every customer-facing agent. ADR-086.
Q27. Why does the Conductor sometimes say "I don't have confirmed information about that"?
A. Because the retrieval layer's top-hit confidence was below 0.35. We'd rather refuse than confabulate. That refusal is the architecturally-correct answer — it's what "epistemic honesty" looks like in production. ADR-086 rule 4.
Q28. Does my Conductor have feelings?
A. No, but it detects yours — the safety service classifies every inbound turn into one of six emotional contexts (stressed / frustrated / celebrating / uncertain / investigating / routine) and modulates tone. Stressed → leads with action already taken, shorter sentences, offers human escalation. Celebrating → acknowledges once, specifically, then moves on. Routine → terse. See `/agents#heart`.
Q29. What will the Conductor refuse to do?
A. Architectural refusals (in code, not in FAQ): - Amazon Seller Central UI automation — **forbidden**. - Meta Business Suite UI automation — **forbidden**. - TikTok multi-account Symphony rotation — **deprecated**. - Any action not on the AI specialist's scope manifest — `ScopeViolationError`. - Any cross-tenant data read — `TenantIsolationError`, fail-closed. - Any claim where the retrieval confidence gate fails — `"I don't have confirmed information about that."`
Q30. Can I override a refusal?
A. No. The `MODE_GATE_BYPASS=manual` fail-open flag exists only for production incident recovery (board-audited). There is no end-user override for Scope refusals or confidence-gate refusals.
Q31. What if I WANT full autopilot?
A. Fine — per-specialist toggle. Unlock an AI specialist, set its mode to 2FA-gated autonomous, configure 2FA. Each action will be pre-confirmed through 2FA. If you want tamper-evident audit on top, enable Companion and switch to audit mode.
Q32. I already pay for Helium10 / Jungle Scout / FeedbackWhiz / DataHawk. Do I need to cancel?
A. No. Run side-by-side for 30 days on our free plan. Decide after you see your own numbers. See `/vs/helium10`, `/vs/jungle-scout`.
Q33. Can I import my existing data?
A. Yes — common CSV exports (Helium10, Jungle Scout, Sellerboard) are ingestable via connector onboarding. Platform API data imports on connect. Historical depth depends on platform retention (Amazon: 2y, TikTok Shop: 1y typical).
Q34. Do you have integrations with [tool]?
A. If it's a platform (Amazon, Shopify, etc.), see `/platforms`. If it's a tool (Slack, Discord, ClickUp), Zapier / n8n export hooks are on the roadmap. Specific-integration requests tracked publicly.
Q35. I run a VA / agency. Is there a tier?
A. Yes — **Exsel Partner** (EP-S $149, EP-P $299, EP-A $499). Multi-client switching, PartnerConductor, tenant-isolated audit per client. ADR-053. See `/partner`.
Q36. Can I white-label?
A. Co-brand yes (your logo next to Exsel attribution). Full white-label on EP-A tier with CNAME custom domains. True white-label (full brand removal) = enterprise-custom for ≥50 client books.