Skip to main content

What's new

What we shipped, and what's next.

Reverse-chronological. Live means you can use it today. Coming soon means we know exactly what we're building and roughly when. We don't log internal refactors here — only things you, the customer, can use or wait for.

Latest

2026-05-09 · Tier 6.11

Multi-vendor draft router (Claude + GPT-4o + Cohere) with auto-retire when local catches up

  • live Every customer-facing draft is now drafted by Claude Sonnet 4, GPT-4o, and Cohere Command-R+ in parallel. Sonnet then judges all four candidates (the three plus our local agent) on the same 5-axis rubric — accuracy, register, completeness, cultural fit, voice — and ships the winner.
  • live Loser drafts become DPO training pairs for the local agent — every round produces a 'this would have beaten you by N points on register/cultural-fit' lesson. The local agent learns from production traffic without retraining.
  • live Auto-retire gate: when the local agent wins ≥60% of rounds over rolling-50, the ensemble disables itself and the local model takes over. Tracked live in /bbb-admin → Foundry tab.
  • live Headline KPI swap: dashboard now shows usable-rate (drafts that ship as-is or with light edits) instead of strict-accept (which was structurally throttled by 240+ pre-fix reviews). The metric that actually maps to 'is the agent doing its job' is finally what's surfaced.
Filter

2026-05-09 · Tier 6.11 (also pinned above)

Multi-vendor draft router (Claude + GPT-4o + Cohere) with auto-retire when local catches up

  • live Every customer-facing draft is now drafted by Claude Sonnet 4, GPT-4o, and Cohere Command-R+ in parallel. Sonnet then judges all four candidates (the three plus our local agent) on the same 5-axis rubric — accuracy, register, completeness, cultural fit, voice — and ships the winner.
  • live Loser drafts become DPO training pairs for the local agent — every round produces a 'this would have beaten you by N points on register/cultural-fit' lesson. The local agent learns from production traffic without retraining.
  • live Auto-retire gate: when the local agent wins ≥60% of rounds over rolling-50, the ensemble disables itself and the local model takes over. Tracked live in /bbb-admin → Foundry tab.
  • live Headline KPI swap: dashboard now shows usable-rate (drafts that ship as-is or with light edits) instead of strict-accept (which was structurally throttled by 240+ pre-fix reviews). The metric that actually maps to 'is the agent doing its job' is finally what's surfaced.

2026-05-09 · Tier 6.10

Voice packs — 4 Arabic dialects, 16 languages, per-tenant dialect picker

  • live Arabic now ships with FOUR production dialect voices: MSA (formal regulator/legal/CMA), Khaleeji (Gulf-warm — Saudi, UAE, Qatar, Kuwait, Bahrain, Oman), Egyptian (Cairo media cadence), Levantine (Lebanon/Syria/Jordan/Palestine). Each gets its own ElevenLabs voice ID + persona register.
  • live Voice coverage extended to 16 languages (added Chinese, Tagalog, Malayalam, Bengali). Email drafter and voice agent both use the same canonical (lang × accent × gender) catalog.
  • live Console → Settings → Voice pack: customers pick their default Arabic dialect (or 'auto-detect') and can pin a custom voice ID per (language, accent, gender) combo using their own ElevenLabs cloned voices.
  • live Chat agent system prompt explicitly enumerates the 4 Arabic dialects + their markers + the 'mirror customer's dialect AND honorific stack' rule. No more MSA-cold replies on warm Khaleeji threads.

2026-05-09 · Tier 6.10

Foundry observability — real-time view of all 67 agents

  • live /bbb-admin → Foundry tab now shows every one of the 67 agents in the fleet, not just dept-level aggregates. Per-agent: level (color-coded ≥7 gold, ≥5 ok), 30-day XP, drill count, reliability score (traffic-lighted), recent trend (↑/→/↓), latest milestone.
  • live Sortable by level / XP / drills / name; filterable by department. Lead and promotion-ready agents are flagged inline.
  • live Customer onboarding wizard rewritten — 4-step flow with animated stepper, per-step transitions, live field validation with shake-on-error, save-and-resume, and a 'Workspace ready' success state with three next-action cards.

2026-05-09 · Tier 6

Mobile SDKs, CRM marketplace, KSA region election, 4 active depts

  • live iOS Swift Package + Android Gradle library for in-app voice — wraps Twilio Voice native SDK; one-line integration.
  • live CRM marketplace: Salesforce, HubSpot, Zendesk, Freshdesk. OAuth for the first three; API key for Freshdesk. Tokens AES-256-GCM-encrypted under your tenant DEK.
  • live Per-tenant data residency. Bahrain region available for KSA latency parity (~22ms p50) + PDPL Article 29 alignment. Default Frankfurt (EU).
  • live Customer Care, Legal, Finance, Marketing departments flipped from learning_only → active. Their leads now run live work.
  • live Daily summary email — every conversation across voice/chat/WhatsApp, summarized to the operator inbox at 08:00 UTC.
  • live Chat retention 7 days minimum + per-visitor chat retrieval endpoint (operators can pull any conversation by visitor_id for the last week).
  • coming soon CallKit / ConnectionService handoff for backgrounded mobile calls (Tier 6.6.1, Q3).
  • coming soon Full per-vendor CRM sync — paginated contacts pull + activity push (Tier 6.7.1, Q3).

2026-05-09 · Chat agent posture

The web chat is now an actual customer-care agent

  • live When a customer says "I want a refund," the agent asks for the order number + email instead of reciting policy. It commits to the next concrete step rather than punting to "contact Customer Care AI" — because it IS Customer Care AI.
  • live Multi-turn conversation: the widget sends the last 8 turns to the API on every request, so follow-ups land with full context.
  • live No more [#0]/[#1] citation markers in the rendered chat — citations remain in the audit trail but are hidden from the customer-facing bubble.

2026-05-08 · Tier 5

12 languages, custom partitions, voice cloning, sentiment, cohorts

  • live 12 languages auto-detected on first turn at this tier (en, ar, fr, es, de, hi, ur, fa, tr, pt, it, ru). Arabic with Khaleeji-warm register. (Tier 6.10 adds zh/tl/ml/bn for 16 total + 3 more Arabic dialects.)
  • live Per-tenant voice cloning via ElevenLabs — upload 60 seconds, agent speaks in your brand voice.
  • live Sentiment + intent on every turn; 7/30/90-day cohort metrics in the console.
  • live Custom partitions (define your own employee partitions beyond the default `general`).
  • live HMAC-signed outbound webhooks with exponential backoff retry.

2026-05-06 · Tier 4

Complaint workflow + audit log

  • live T1 → T2 → T3 → T4 complaint pipeline built into the runtime. Every transition audited; the customer-facing reference number is generated at T1 and follows through all 4 tiers.
  • live Write-once audit log. Exportable as CSV per tenant for GDPR Article 15 / PDPL Article 14 (data subject access).
  • live Cost dashboard — per-tenant, per-agent, per-call cost breakdown. STT, TTS, LLM, embeddings, vector queries, rerank — all priced and totaled.

2026-05-04 · Tier 3

Voice on PSTN + WebRTC + WhatsApp

  • live Inbound voice across the GCC via SIP trunks with the local carriers — your existing carrier relationship stays intact.
  • live In-browser WebRTC voice ("Talk to us" button) — no phone number needed.
  • live WhatsApp Business API — inbound + outbound text, images, audio, PDFs.
  • live Per-tenant Twilio subaccount, auto-provisioned at signup.

2026-05-01 · Tier 1+2

Multi-tenant foundation

  • live Postgres + pgvector with RLS per tenant, per-tenant DEK encryption (AES-256-GCM).
  • live Hybrid retrieval: pgvector dense + tsvector sparse + Cohere rerank.
  • live Stripe Checkout + webhook + tenant provisioning. Cloud Starter, Cloud Pro, Pro Local plans live.