[FIELD REPORT · 006]
CLASSIFIED · PRE-LAUNCH · BY INVITATION

An AI layer that turns your team’s knowledge — into every rep’s answer.

Not a copilot you summon. A substrate that sits under every reply your team writes, cites the page it pulled from, and learns from the edit your senior rep just made.

[EXHIBIT D] · MESSAGE → PIPELINE → SENT
MSGCLASSIFY[01]RETRIEVE[02]SUGGEST[03]ACT[04]LEARN[05]SCORE[06]SENT
[01]CAPABILITY MATRIX · DEEPSix movements, one drafting surface
[01] · Classify

Classify

Every inbound thread is tagged the moment it lands — intent, urgency, account tier, channel. The taxonomy isn’t static and it isn’t ours. It’s learned from the patterns your senior reps already mark. No ontology to maintain, no rules engine to configure. The model watches what your best people flag, and it mirrors the shape of their judgment in real time.

INBOUND · TAGGED
INBOUNDINTENT: ORDERURGENCY: HIGHTIER: ENTERPRISE
[02] · Retrieve

Retrieve

The model doesn’t guess — it searches your corpus first. Pricing sheets, engineering specs, return policies, engagement letters. Every sentence in the draft carries a citation back to the page it pulled from. If the source is stale or the confidence is low, the system says so instead of confabulating an answer.

INBOUND
Hey! I want to order the 3-pack bowl set in walnut. Can you ship to Haifa and is the oil food-safe? Also do you take returns if I don't love the grain?
DRAFT · AI

Hey Maya, all good on those three.

The walnut 3-pack ships from our Tel Aviv workshop to Haifa in 2–4 business days.

Cited
[SOURCE:•••••••••••••••••••••••][SOURCE:•••••••••••••••••••••••••••][SOURCE:••••••••••••••]
Classify: order intentRetrieve: catalog + policySuggest: video upsell
[03] · Suggest

Suggest

The first draft appears beside the thread, not in a separate window. It reads like your senior rep wrote it because it was grounded in what your senior rep has approved before. The junior edits a paragraph, not a blank page — and the delta between the draft and what they actually send becomes the next training signal.

INBOUND
Hi — question on the Kravet boucle for the living room sofa. Two dogs and a toddler, is the Martindale rating enough? Can it ship to SF in time for the April 18 install? And you mentioned we could save with a trade swap — pull something in the same warm-ivory?
DRAFT · AI

Sarah, good questions — all three answerable.

Kravet 12345: Martindale is 40K, above the 30K residential heavy-use threshold. Holds up to dogs + toddler per the durability guide.

Cited
[SOURCE:•••••••••••••••••••••••••••][SOURCE:••••••••••••••••••••••••••][SOURCE:•••••••••••••••••••••]
Retrieve: fabric ratingCheck: lead-time windowSuggest: trade swap
[04] · Act

Act

A reply is rarely just a reply. The drafting surface can push a quote to your ERP, update an order status in your commerce layer, or flag a billing adjustment. Write-backs are gated by rep approval — never autonomous. The rep confirms; the system executes the downstream call.

WRITE-BACK · GATED
DRAFTAPPROVEERPCOMMERCEBILLING
[05] · Learn

Learn

Every edit a senior rep makes to a draft is a training signal — scoped to your tenant, not shared with anyone else. Over weeks, the drafts converge toward the voice and judgment your team already has. No prompt engineering. No fine-tuning console. The feedback loop is the product itself.

ON-TENANT LEARNING CURVE
DAY 1Draft acceptance
DAY 90Draft acceptance
[06] · Score

Score

The system watches its own output. Account health, draft quality, topic drift, escalation rate — surfaced as dials your ops lead can read without switching tools. When a score moves, you know before the customer does. These aren’t vanity metrics; they’re the early-warning system your team never built because nobody had the time.

76%
QUALITY
illustrative
[ILLUSTRATIVE]
[02][PIPELINE]How one message moves through
[01] Classify
intent
urgency
confidence
[02] Retrieve
[SOURCE: catalog.md]
[SOURCE: pricing.md]
[SOURCE: policy.md]
[03] Suggest
DRAFT PREVIEW
[04] Act
POST /orders/quote
→ 201 ACCEPTED
[05] Learn
DAY 1DAY 90
[06] Score
76%
[03][POSTURE · ILLUSTRATIVE]The four dials we watch for you
[α]
Ground-truth accuracy
[ILLUSTRATIVE · NOT REAL]
87%
Ground-truth accuracy

How often the draft's cited facts match the source material verbatim.

[β]
First-draft acceptance
[ILLUSTRATIVE · NOT REAL]
74%
First-draft acceptance

Percentage of AI-drafted replies sent by the rep without major edits.

[γ]
Escalation rate
[ILLUSTRATIVE · NOT REAL]
18%
Escalation rate

Threads requiring senior-rep override. Lower is better — this dial inverts.

[δ]
Knowledge freshness
[ILLUSTRATIVE · NOT REAL]
92%
Knowledge freshness

Corpus sync recency, normalized. High means the knowledge base is current.

[04][SCHEMA · CLASSIFIED]How a thread is tagged
intent:
urgency:
account_tier:
channel:whatsapp
language:en
sentiment:
routing_hint:
owner:
sla_bucket:
confidence:0.

Fields visible above are the schema shape. Values are redacted — the taxonomy is yours, not ours.

[ACCESS]REQUEST INVITATION

We're onboarding in cohorts.
Tell us your shape.

Who you sell to, how many channels you run, and one thing your best rep knows that the rest of the team doesn't. We'll reply with a walkthrough grounded in your actual docs — not a generic demo.

CLASSIFIED · PRE-LAUNCH · BY INVITATION
[RESPONSE SLA] 1 business day · [REGION] EU/US · [CAPACITY] Q1 2026
[FORM · W-01]ENCRYPTED IN TRANSIT

By requesting access you agree to our Privacy and Terms.