Talk to your books. Nothing posts until you say so.
Kate answers from real queries against your own ledger and drafts the record you asked for as an approve or discard card. Type to her or call her, it is the same assistant either way, and the same rule underneath: she proposes, you approve, deterministic code writes.
Kate is live on this page. Tap the button in the corner and ask her about Caytava, out loud.
Every number she gives you came from a query.
Your question is routed to one lane and answered by the same engines that produce your statements, so Kate and your reports can never disagree. Every answer names the query it came from, and most carry a link to the screen that figure lives on. When something genuinely cannot be computed, she says so instead of producing a number.
She fills in the form. You decide whether it exists.
244 intents in her drafting router today. Ask in a sentence and Kate comes back with a real record in draft status, or the exact change you described, held behind an approve or discard card in the pane with a pointer to where the same thing lives in the app.
A draft is a real record, not a suggestion
Kate never writes to the ledger. She fills in a structured proposal, and the same validated function the manual screen uses does the write once you approve.
- Double entry, a valid chart of accounts and the period lock are enforced by the write path, not by the model
- A locked period refuses the entry no matter who asked for it
- Approve, edit or discard from the card, and the approver’s name goes on it
Documents and records
The paperwork half of the day, drafted by sentence.
- Invoices, bills, credit notes, vendor credits, quotes, purchase orders, journal entries
- Customers and vendors, GL accounts, products and services, projects, workers
- Customer deposits, applying money held on account, writing off an invoice you will not collect
- Time entries, leave requests, shifts, and the journal entry behind a finished payroll register
- CRM notes, activities, deals and deal lines, prospects, people, and a support ticket to us
On that last one, the boundary matters: Kate reads a finished run, yours or your provider’s, into a balanced dimensioned entry, and a second person has to approve it. She does not calculate gross to net, withhold, or file. Native payroll lands in 2027.
Reports, dashboards and settings
The things people normally file a ticket for.
- Describe a report and she builds it, then saves it as a view, emails it, or shares it with a named colleague
- “Set my home page up for an AP clerk”, and she proposes the widget set, drawn only from the real catalog
- Draft a budget or forecast, and set up an allocation rule that splits a cost across departments
- Turn a module or a setting on or off, requiring a purchase order on bills, for instance
Where she cannot go yet, in writing
Every human action in the product carries a row in our capability registry with Kate’s status against it, and a shipped action with no Kate path is unfinished work with a ticket number on it. Parity is a program we run, not a state we have reached, so here are the edges.
- Releasing a payment run is human only. Kate can tell you what is due, she cannot move the money
- Confirming, fulfilling and invoicing a sales order are human only for now
- A foreign currency invoice is drafted on the form, not by Kate
- She does not build alerts. Reports, dashboards and notifications are what ships
One Kate, typed or spoken.
Kate Live is a real-time voice line into the same assistant, held to the same capabilities on purpose. A capability wired into only one of the two surfaces is treated as a defect and does not ship.
The call is the same conversation
Start typing at your desk, finish the thought out loud in the car. A live call continues the thread you were already in, and what she works out lands in the same history.
- Interruptible. Barge in as soon as you have heard enough
- A draft made by voice renders the same approve card on your screen
- A chat drops out of your list after 30 days unless you save it, and what it actually did stays in the audit trail
Your browser never holds a Google credential
The voice relay runs on our servers. The browser is a microphone and a speaker and nothing else, and every tool call executes inside your own authenticated, tenant-scoped session.
- 433 tools behind the voice path, each one a read or a propose-only draft
- Every tool that carries company data is tagged to a capability module, and the ones your badges do not cover are never offered to the model in the first place
- Voice is a per-company switch and stays off until an admin turns it on
Agents that leave you an inbox, not a surprise.
Kate is not only the chat pane. The same propose-then-approve shape runs on a schedule, over your documents, and on the bank lines nothing else could explain. The scheduled agents and the analyst loop are per-company switches, off until an admin turns them on.
Four scheduled agents
A close sentinel that runs the 27-rule exception scan daily. An AR chaser that ranks who owes you most and proposes a chase. An amortization autopilot that drafts the month’s schedule entries. A flux analyst that explains last month’s misses the moment the month turns.
The AI Actions inbox
Every run writes proposed actions you review. Approving one executes the same deterministic path a person would take. A re-scan replaces its own stale item instead of stacking duplicates, so the queue shrinks as you work it.
Coding proposals on bank lines
The bottom rung of the coding ladder, under your decisions, your rules and a model trained only on your own books. Always queued, never posted. It reasons in words, never a score, and never a count from another company’s books.
Documents read into drafts
A bill or a receipt becomes a drafted payable. A loan agreement and a lease become their schedules. A supplier quote becomes a purchase order. A remittance advice becomes a proposed cash application across the invoices it names.
The analyst loop
For a strategic question, Kate plans which real figures she needs, pulls each one through the deterministic tools, then writes the analysis over verified numbers. If the loop fails she falls back to the ordinary grounded answer, never a guess.
Words, drafted for you to edit
Variance causes and the levers worth pulling, GAAP disclosure notes computed from your balances, the business profile behind your forecast, a pre-meeting brief built from real CRM history. You edit before anything is issued.
Why we can let AI near a ledger.
These are not policies written down somewhere. They are properties of the code path, which is why they hold on a bad day.
- The model never writes. It proposes a structured call. Validated ledger functions do every write and enforce the accounting rules.
- Nothing posts unattended. There is no autonomous mode, deliberately. A human tap commits, and the approval is recorded.
- Your badges decide, not the pane. Every action is checked on the server. An AP clerk is never offered payroll tools, so there is nothing for Kate to decline.
- Answers cannot cross tenants. Q&A is read-only inside a row-level-security-scoped session, on a database where a company-scoped table ships with its own tenant policy or fails the build. The dozen deliberate exceptions are named, with reasons, in the test that enforces it.
- External AI clients get a narrow door. A separate server whose tools only read or draft, reached with per-company tokens, hashed at rest, expiring, revocable, and never wider than the person who minted them.
She is on this page. Ask her yourself.
The “Talk to Kate” button in the corner opens a live voice line running the same voice stack. Here she is grounded in a curated knowledge base about the product rather than anyone’s books, and no customer data is reachable from this page. Ask what Caytava replaces, what is included, how your data is isolated. Out loud.
Kate is in every plan, from Solo at $59 a month, and chat is unlimited. Live voice includes an hour a month, then $0.85 a minute.
Why this is the pitch and not a gimmick: the voice pipeline you are about to use is the one that runs on real books, and so is the rule underneath it, that she looks a fact up and phrases what comes back rather than recalling it. What is missing here is the half that needs a ledger. This Kate has no database session and no product tools, so there is nothing for her to draft.
Everything she answers from is a module here.
Talk to your books. Literally.
Thirty minutes on data that looks like yours: a question answered from the ledger, a record drafted, an approval, and the entry that follows.
Let’s begin