TRIVISAI ACCOUNTING

"Everyone talks about agents — usually without knowing what the term means. An agent isn't a chatbot that 'knows' the answer. An agent is something that knows how to reach the answer: it doesn't compute a period's KPI change in its head — it knows which tools to call, in order: query the data, aggregate, compare. The model picks the path and gives the command; the numbers come from the system — exact and auditable."

Toomas Riiberg · Trivis product owner
Vocabulary

Tool, skill and agent

The word "agent" gets used a lot, but rarely precisely. It all starts with three layers — and it's their interplay that makes AI trustworthy in bookkeeping.

Tool (tööriist)

A single, exact, deterministic operation: query an account balance, calculate VAT, record a payment. A tool doesn't guess or estimate — it computes. The same input always gives the same result.

get_account_balance("1010")  → €12,480.00

Skill (oskus)

A recipe that says which tools to call, and in what order. For example: "to find a period's KPI change, query period A, query period B, subtract, compare." A skill is encoded know-how — not the model's guess.

trial_balance(A) → trial_balance(B) → diff → compare

Agent (agent)

An actor that's given a goal by a human, then chooses the right skills and tools itself, runs them step by step, checks the results, and knows when the job is done. Steps that change the books always require human approval.

goal → pick skill → call tools → verify → confirm

Compute. Verify. Eliminate error.

Why it matters

Trivis never lets the model do the math

A language model on its own can give a confident, wrong answer. In bookkeeping that isn't a "hallucination" — it's an error in a report, an invoice paid wrong, a liability.

That's why Trivis never asks the model for a number. The model picks the right tool; the system produces the number — with decimal precision and no silent rounding. Then a second tool verifies the result: debit equals credit, period balances reconcile, the tax rate is valid. The error never reaches the books.

Model alone
  • Computes in its head — numbers can be wrong
  • Rounds silently, loses the cents
  • Leaves no trail — errors surface late, or never
  • The liability is yours
Trivis agent
  • Tools compute exactly, with decimal arithmetic
  • A second tool checks every result
  • Every step is logged to an immutable audit trail
  • Any step that changes the books needs your approval
Example

A purchase invoice, end to end

You drag in a file. The agent does the rest — and shows you every step.

1 create_purchase_import AI extracts the line items, amounts and supplier from the invoice. 2 match_account Semantic search finds the right expense account for each line. 3 get_tax_rate Each line gets the valid VAT rate — from the system, not a guess. validate A second tool checks: lines sum up, debit equals credit, taxes reconcile. human approval The agent waits for your consent. Only then is the invoice posted to the books.

You don't enter the numbers. You review and approve.

See how the agent uses tools

Trivis MCP exposes every tool to your AI assistant. Read the documentation or try it right now.

See the MCP tools → Start for free