Your CRM knows the stage. It doesn't know the deal.
A CRM tells you what stage a deal is in. This maintains an evidence-backed model of what is actually happening inside it — every claim traced to the buyer's own words — and recommends the next action, with reasons, deterministically.
Robinett Industries is the first customer: every real RI prospect conversation runs through the same engine on a private internal instance, so the demo below is the reference implementation of a system in actual use — not a mockup of one.
The same deal, two records
Stage, date, and dollar amount tell you a deal exists. They do not tell you what to do today.
Calloway Manufacturing Stage: Discovery Last contact: 3 days ago Deal value: $32,000
Calloway Manufacturing Goal Eliminate manual order re-keying Primary blocker ROI not established Evidence "I don't know if there's enough volume here to justify a project." Missing information transaction volume minutes per order loaded labor cost Next best action Quantify the current process cost before proposing scope. Do not Send a proposal yet.
The model observes. The engine decides.
A model does the one thing only a model can: read messy human communication and emit observations, each carrying a verbatim quote validated against the source — it cannot invent its evidence. Everything downstream is deterministic and inspectable, which is why the recommendation can always answer “why?” with the buyer's own words.
raw interaction
↓
model: semantic observations the ONLY model step
↓
deterministic normalization quotes validated against source text
↓
FACT assertions
↓
deterministic inference rules
↓
INFERENCE assertions
↓
deal-state projection rebuilt from evidence, never stored
↓
deterministic recommendation versioned rules — your methodology
↓
RECOMMENDATION + DO-NOTsWhat it does
An evidence-backed deal model: every claim attributed to a buyer traces to a validated span of text the buyer actually supplied — 'why do we think integration risk is the blocker?' answers with quoted evidence, not a score
Three reasoning layers kept structurally apart — FACT (what the communication establishes), INFERENCE (what the engine concludes), RECOMMENDATION (what it advises) — never collapsed into one record
A deliberately small decision ontology: eight buyer decision drivers, eight blocker categories, and a deterministic blocker-to-intervention map — the AI interprets language; the rules own the sales methodology
Objective before artifact: the engine recommends what to establish and which information is missing; drafting an email or a proposal is a subordinate extension, never the decision
Explicit do-nots as first-class output — 'do not send a proposal yet' when the economic case is not established and the buyer has not asked
Append-only truth with supersession: interactions, extractions, and assertions are never deleted; deal state rebuilds from source history, so the reasoning engine can be replaced without losing a single deal
Recommendations with lifecycles and structured outcome observations (completed never implies successful) — the longitudinal record that eventually measures whether the engine's advice advances deals
One narrow correction mechanism: confirm, reject, or correct any extracted fact or inference; corrections are append-only human assertions that overrule re-inference and become the extraction-evaluation dataset
Claims, graded
Everything the demo executes is the real engine: evidence-span validation (a model cannot invent its supporting quote), the FACT/INFERENCE/RECOMMENDATION separation, deterministic inference and versioned recommendation rules with explicit do-nots, append-only assertions with supersession, first-class recommendations with lifecycles and structured outcomes, per-opportunity event history with contiguous sequences, and append-only human corrections that outrank re-inference. All of it gate-verified, including a full provenance-chain traversal.
The public demo is deterministic on purpose: five seeded reference deals plus prepared follow-up interactions whose extraction is known. No model call happens on the public tenant, and the demo says so rather than pretending to analyze arbitrary text. The live extraction path — the same engine over freeform prospect conversations — runs on Robinett Industries' private internal instance, where RI is the first customer.
CRM sync (HubSpot, Salesforce), Gmail and calendar ingestion, call recording, drafting artifacts beyond the recommendation, autonomous outreach, dashboards, and multi-user permissions. The architecture treats every one of them as an adapter around the core loop, and none of them ship until the loop itself proves out.
This is not an AI salesperson, not autonomous outreach, and not a model deciding your sales methodology. The model interprets messy human communication; explicit, inspectable rules decide what to do about it — and that division is the product.
Limits, stated plainly
- The public demo world is yours alone and expires after 24 hours; the five reference deals are fictional companies built through the production write path, not hand-authored screenshots.
- Public extraction is fixture-only: content outside the prepared set gets a truthful refusal, never a fake analysis. Live extraction is reserved for the internal instance, under the site's shared daily AI budget.
- When live extraction cannot run — budget exhausted, provider down, unusable output — the interaction is saved, deal state does not change, and the failure keeps its name. There is no silent fallback.
- One operator, one credential on the internal instance. Multi-user permissions are out of scope until the loop earns them.
- The v0 ontology is deliberately small: eight decision drivers, eight blocker categories, one intervention map. It is versioned so it can grow without rewriting history.