Language and conversational intelligence.
It presents what the others established. It does not decide.
The language engine is the surface, not the brain. It works out what you are asking, orchestrates typed tools to find out, and explains the answer — using figures the other engines produced, never figures of its own.
A cascade, cheapest first.
The conversational surface has its own discipline: a request walks a cascade and stops at the first layer that resolves it confidently, so most questions never reach a paid model at all.
- Layer 1, deterministic router — a real grammar over your own vocabulary, with slot-filling and conversation focus. Free, and instant.
- Layer 2, on-device classifier — a bag-of-words natural-language model trained at boot, running on your own server. Free.
- Layer 2b, learned replay — a phrase this tenant has asked before, replayed and re-validated live against current permissions. Free.
- Layer 3, model router — a paid model, answering into a strict contract that is structurally incapable of expressing an action.
- Layer 4, extraction and vision — escalation for documents that need it.
Intent, clarification, orchestration, explanation.
Intent routing
Works out what is being asked, about which records, on which page — and refuses rather than guesses when the question is genuinely ambiguous.
Clarification
Asks the one question that would resolve the ambiguity, rather than answering the wrong question confidently.
Typed-tool orchestration
Calls registered, permissioned read tools. It has no ability to compose a query of its own.
Grounded explanation
Puts the answer in your terminology, with the records it read attached and every figure labelled by how it was produced.
No queries of its own, no writes, no invented arithmetic.
Model output is treated as untrusted data with the same standing as a message from a stranger. It crosses into the system only by validating against a schema that cannot express an action; entity identity is always re-resolved against your database, so an identifier the model supplies is a hint and never a reference; and permissions are checked live rather than read from anything the model produced. It may not write a database query, write to the database, invent a calculation, or state a claim it cannot support.
Every surface, one envelope.
The assistant panel, the command palette, the inline ask on list pages and on errors, the page-aware suggestion chips and the inline insight boxes all emit the same normalised request — so the orchestrator treats a typed question and a clicked suggestion identically.
- “Open sales order SO-10407” — resolved and navigated without a model.
- “What needs attention today?” — the working queue, deterministically assembled and put into a sentence.
- “Show me overdue invoices for this customer” — classified, then a typed read tool, then a reply that cites what it read.
- “Why is this on hold?” — the reason from that record, not a help article about holds in general.
- “Draft a credit note for the short delivery” — a genuine unposted draft, prepared for your confirmation.
Frequently asked questions.
Which model does it use?
Providers sit behind an interface, so the model is a deployment choice rather than an architectural one. What matters is that whichever one answers, it answers into a strict contract and its output is validated before anything acts on it.
Can I ask it to change something?
You can ask it to prepare something. It creates a genuine unposted draft through the normal domain service and waits for your confirmation. Posting, approval and payment are outside what it can do at all.
Does it make up numbers?
It is not permitted to produce a calculation of its own. Figures come from the other engines, and each is labelled with how it was produced — calculated, rule-based, extracted or generated.
What happens if it does not understand me?
It asks. A targeted clarifying question is better than a confident answer to a question you did not ask, and refusing is better than guessing about which record you meant.
Does it cost credits every time?
No — most requests are resolved by the deterministic router or the on-device classifier, and both are free. Only the model router and the vision escalation are metered.
See this engine on your own records.
Join the waitlist and ask it something real. Every answer names the engines it used and the records they read.
Join waitlist