AI in Conovo
AI reads the contract. Code sends it.
Conovo uses AI to understand a document once, and never to produce one. A model reads the contract your customer already sends, proposes every variable field, and maps those fields to the data in your product. A person confirms it. After that, generating a contract is deterministic code. The same inputs produce the same document, every time, with no model in the path.
Setup
Runs once, per contract type.
Turning a Word file into something your app can fill.
This is where nearly all of the AI lives, and it is the work that would otherwise land on your engineers: reading someone else's contract and deciding what varies.
Finds every blank in a contract you upload
Reads the document your customer already sends and proposes each variable field: typed, labelled in plain English, and anchored to the exact place it appears. Low-confidence detections arrive as suggestions, not as decisions.
Maps those blanks to your data
Against the payload schema you registered, so a field is not just “client name” but the path in your own product that fills it. The proposal shows the resolved value, not only the path, because a mis-map looks fine as a path and obvious as a value.
Spots the arithmetic and writes the expression
Deposit is half the fee; the final payment is the balance. Proposed as a formula and evaluated in code. A model never does the arithmetic.
Turns “a third of the total fee” into a formula
Typed for the user, parse-checked against the template's real fields, and refused with a readable reason when it references something that does not exist.
Reuses what this business already confirmed
A second upload stops re-asking what the first one answered: fields align onto the keys their earlier templates use, and a reused key with a saved value pre-fills.
Carries a template forward when the document changes
Re-uploading last year's agreement with three edited clauses keeps everything already set up and flags only what moved.
Reads the contract and reports gaps
Missing notice provisions, contradictory clauses, placeholder text left in. Findings carry a plain-English explanation and, where useful, a proposed clause you accept or ignore. Not legal advice, and framed that way on the screen.
Says how much will fill automatically, and what is left
The counting is plain code; one model call phrases it. If the phrasing fails, the numbers still arrive.
At send time
Deliberately almost nothing.
Three things run near a send. None of them write the contract.
A second look before an unattended send
Runs only on auto-sends, the path with no human reviewing, and can do exactly one thing: hold the contract as a draft. It cannot alter a word of it. If the check itself fails, the deterministic validator is still the gate.
Pre-fills one long text field from your record
A scope of work, drafted from the project data, into the form. The person sending it reads and edits before submitting, and the pipeline only ever sees what they submit. Refused outright on money, dates and numbers.
Matches spreadsheet columns to fields
Proposes which CSV column is which field, with the reason. Sample cells are masked to their shape before the model sees them, so a column of ID numbers travels as ###-##-####.
Afterwards
For signers, and for your team.
The parts pointed at people rather than documents.
Answers questions about the document in front of them
Grounded in the filled contract, quoting the passage it used, and honest when the document simply does not say. It describes terms; it does not advise whether to sign.
Turns a sample payload into a registered schema
Infers the schema from a record you paste, and proposes the paths to add next based on what businesses keep filling by hand.
Diagnoses an account and explains what it found
The one tool-using agent in the product. Read-only, over content-free surfaces: held contracts, payload gaps, request logs, webhook deliveries. Ask why contracts are being held and it correlates, quantifies and names the fields to fix.
The boundary
The part your counsel will read.
Four things the AI is not allowed to do.
These are enforced in the architecture, not by prompting. They are also why a platform can put contracts in front of its customers without becoming a legal product itself.
It never writes the contract at send time
Not one word of a generated document comes from a model. Fields resolve, formulas evaluate, sections are kept or removed, all in code and all reproducible. This is the invariant everything else is arranged around.
It never does the arithmetic
Money and dates run through a decimal-safe evaluator. Formulas are data; evaluation is code. A language model is the wrong tool for a number that appears on an invoice.
It never converts a document silently
Every AI output that becomes part of a template (fields, bindings, formulas, conditional sections) passes through a human confirming it. Low-confidence results are shown as suggestions and start switched off.
It never gives legal advice
No case-law search, no opinion on whether terms are fair, no telling anyone which agreement fits their situation. When a document genuinely needs judgment, Conovo routes it to a real attorney and records that the review happened.
Questions
Common questions.
Does AI write the contracts Conovo sends?
No. No model output lands in a generated contract at send time. AI discovers the structure of a document once, during setup, and a human confirms it; from then on generation is deterministic. Fields resolve from your data, formulas evaluate in code, and the same inputs always produce the same document. The only model that runs at send time is an optional anomaly check, and it can only hold a contract as a draft, never change it.
Which model does Conovo use?
Claude, called with structured outputs so no free-text model response is ever parsed into product state. Every pipeline is a typed function with a schema on both ends.
What does the AI see of our customers' data?
Documents at setup, and a masked sample when mapping spreadsheet columns. Fields classified as sensitive are redacted before the anomaly check sees them (the real value still fills the contract, only the model's copy is redacted), and pre-drafting refuses sensitive fields outright. Contract content and recipient details never appear in logs or audit payloads.
Can Conovo tell our customers which contract they need?
No, and that is deliberate. Choosing a legal instrument for someone's circumstances is legal advice. Businesses browse the starter templates their platform published, filtered by words the platform wrote, and a real attorney reviews anything that needs judgment.
What happens if the AI is wrong?
Setup results are proposals a human accepts or rejects, so a wrong extraction costs a click. At send time the deterministic validator is the gate: a contract that fails validation becomes a draft for review and is never sent silently. The anomaly check is an extra net, not the gate: if it fails entirely, nothing about the send changes.
See it
Run it on a contract you actually send.
Upload an agreement your business already uses and watch the extraction, the bindings and the generated document. No account, no card, no install.