Job description and manager
Each agent has a written scope and a named human manager. Its permissions are the overlap of its own grants and the manager's, checked again at the moment the work runs.
Governed AI agents
Each agent in Tosven has a job description, a human manager and an autonomy level you set. It prepares the work, shows its reason and evidence, and nothing runs unless the approval matches the exact data a person saw.
Most AI in business software is a text box. Tosven treats agents as staff. Each one sits on the same roster as your people, with a job description, a named human manager, a permission boundary and an autonomy level. An agent can never see or do more than its manager can. Its accuracy, its cost and every mistake it has made are open on the team screen.
The work arrives as decision cards. A card states what the agent proposes, why, which records it rests on and how confident it is. Your approval is bound to the exact content you looked at. If the invoice, the balance or your own permission changed in the meantime, nothing runs and the card comes back. Every decision is signed into the audit log, by a person or by a published rule.
Agents
Six agents on one roster. Two are in production in the pilot today. Four are shown as designed and are on the way. Each card names the agent, its manager and its current level.
Sample data · agent roster
Each agent has a written scope and a named human manager. Its permissions are the overlap of its own grants and the manager's, checked again at the moment the work runs.
L0 suggests only. L1 acts on exact matches. L2 acts below a threshold you set. L3 has full authority in a defined area. Amount and daily limits apply per type of action.
Every suggestion is a card with the reason, the evidence and the confidence, including what that confidence is based on. A card that lacks any of the three cannot be created.
Your signature covers the exact content and source versions you saw. If either changed before execution, the card turns stale, no command runs and it returns to the queue.
Proposer, decider and executor are three separate records. Decisions are chained per company and cannot be edited or deleted, whether a person or a published rule made them.
Before a level is raised, the last 30 days are replayed at the new level. The raise also needs a second approver and a 72 hour cooling period. Lowering a level is instant.
Accuracy is measured from results verified at the source, such as an invoice that was settled or an entry that was never corrected. Clicking approve does not count as correct.
Each run has a step, time and cost budget enforced in SQL. Model providers are limited by an allowlist that changes only with a product release, and every call is logged.
When an agent lacks history, evidence or a fresh source, it says so with a reason code and the next step for a person. It does not guess.
In practice
A freight invoice from Baltic Freight OÜ arrives and the Ledger agent prepares the entry for the accountant.
A reminder for Nordlys AB is approved in the morning, and the payment lands before the letter leaves.
Elina Koskinen wants small invoices from known suppliers posted under a published rule, without a click.
Sample data
In production today. It codes incoming purchase invoices, proposes matches between bank lines and open items, and watches suspense balances and period-close tasks. It learns a coding rule per supplier only from invoices that people coded, tests the rule on invoices it has not seen and runs it in shadow before anyone can publish it.
In production today. It watches open receivables and proposes payment reminders on a ladder of up to three steps, with minimum intervals per country. It is a rule engine and makes no model calls. It charges no fees and no interest. A pause, a dispute or a payment promise stops it at once, and a partial payment always goes to a person.
Some ceilings are product law. No agent executes a payment. Tax filings and ledger postings stay under human approval unless a documented rule has been published. Every payment, any new or changed IBAN and any entry above 10.000,00 € need two different people. A kill switch stops a single run, one agent or a whole company.
Treasury, Sales, Purchasing and Compliance agents are designed and are not yet in the pilot. Treasury will propose payment runs from open payables and the cash forecast. Sales will raise follow-up cards from deal signals. Purchasing will propose orders from lead times and minimum quantities. Compliance will watch VAT deadlines, filings and period locks.
In the field · Property management company
A property management company keeps the books for dozens of housing companies and commercial buildings. Each building receives the same kinds of invoices every month: energy, cleaning, lift service, waste. Each has tenants and owners who pay late in the same patterns. The work is repetitive, the volume is high, and a wrong VAT code or a reminder sent after a payment becomes a complaint from a board chairman.
In Tosven the Ledger agent codes the recurring supplier invoices for each building and brings the unusual ones forward with a reason. The Collections agent keeps the reminder ladder for every overdue rent or service charge and stops when a tenant disputes the amount or promises a date. The property accountant approves a short queue in the morning. The manager of each agent sees accuracy, cost and errors on the team screen.
Works with
Decision cards land on the morning board, where a person approves or rejects them from the keyboard.
Accounting and financeThe Ledger agent proposes entries. The ledger accepts them only with a signature or a published rule.
Metadata engineA record type defined in Builder becomes a tool an agent can read and propose changes to.
Status
Only under a rule that a person published. By default every entry an agent proposes waits for a human signature. A company can publish a documented posting rule and raise the agent to L2 for that rule, after a replay, a second approver and a cooling period. Each automatic posting is then signed in the name of that rule.
Nothing runs. Your approval is bound to a hash of the card content and to the versions of the records it was computed from. At execution Tosven recomputes both and asks again whether you still hold the permission. If anything differs, the card is marked stale and returns to the queue with the reason.
No. The payment ceiling is zero as a product rule, not as a setting. The planned Treasury agent will propose a payment run and prepare a payment file. Two different people sign it and a person uploads it to the bank. Any new or changed IBAN also needs two signatures.
Agents may call only the model providers on an allowlist, defined per class of data. The list changes only with a product release, never from a settings screen, and every call is recorded with its cost. The Collections agent uses no model at all: its reminder ladder is a deterministic rule.
From outcomes verified at the source, not from approvals. A coding counts as correct when the period locks without the entry being reversed or corrected. A reminder is verified when the invoice is settled. A wrong VAT box, a wrong period and a wrong counterparty count as critical errors and are listed in the open error log.
Early access
Tosven is in pilot. Tell us about your company and the tools you use today in a few sentences, and we will say plainly whether it is a fit.
sales@tosven.com · No card details and no commitment.