AI
Choose the model that answers for your business, keep personal data out of it, decide how long conversations live, and cap what a month can cost.
Navigate to: Settings → AI (Super Administrator)
Trabalance's help assistant, bank-line categorisation, the integration builder and AI assistants connected over MCP all reach a model through one door on the Trabalance API. This page is where a business governs that door.
Your books are never sent to a model. Only the question is. The assistant picks a data tool, Trabalance runs the query with the asking user's own permissions, and the answer is formatted on the server.
Status
Three cards summarise the door for this business:
| Card | What it shows |
|---|---|
| Status | On or Off, and which provider and model currently answer — and whether it is your own key |
| This month | Metered spend so far, against the monthly budget when one is set |
| Privacy shield | Whether personal data is masked before any model sees a prompt |
Model
| Setting | Meaning |
|---|---|
| AI is on for this business | Off means nothing leaves the server: the assistant, bank categorisation and connected AI assistants all answer that AI is turned off |
| Provider | Trabalance default, or your own — Anthropic, or any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, Gemini, Groq, Mistral, a self-hosted model) |
| Model | Leave blank for the provider's default, or name one |
| API key | Your provider key. Encrypted at rest, write-only — it is never shown again. Enter a new one to replace it |
| Base URL | For OpenAI-compatible endpoints: the /v1 root of the API. HTTPS only |
Choose the provider, paste the key (and base URL for a compatible endpoint), optionally name a model, and save. The status card confirms the change and shows your key.
Choose Trabalance default and save. Your key is removed.
Privacy and limits
These are enforced on the server before a call is made — they are not instructions to the model.
| Setting | Meaning |
|---|---|
| Privacy shield | Emails, phone numbers, account and card numbers and the names of your customers, vendors and employees are replaced with tokens in every prompt and restored in the answer. On by default |
| Keep conversations for | 7 days to 1 year. Older assistant conversations are deleted nightly |
| Monthly budget (USD) | When the metered spend reaches it, AI answers that the budget is used up until the 1st of next month. Leave blank for no limit |
Usage this month
Every call is metered: surface (Assistant, AI assistants over MCP, Bank categorisation, Integration builder), model, calls, tokens in and out, and cost. The figure on the This month card is the total of this table.
Every tool an assistant runs — and every one it is refused — is written to the security log with the user, the surface and the tool, never the rows. A user erasure empties that person's conversations.
Who can change this
Only a Super Administrator sees this page. What the assistant may answer for a given user is decided by that user's own role and features on every question, not by anything here.
Agents
An agent is a named helper that works on your behalf with a narrowed set of your own rights. It can never do what you cannot, and it never posts anything itself — it proposes, and a person approves.
| Setting | Meaning |
|---|---|
| Skills | What the agent may propose: draft an invoice, record a customer receipt, classify bank lines, chase overdue invoices |
| Features | Derived from the skills — the only features the agent holds. Payroll, reporting and everything else stay out of its reach |
| Autonomy | Propose (a person approves every action) · Suggest (findings only) · Act within a limit (applies its own proposals up to an amount; above it, it proposes) |
| Monthly budget | The agent's own model spend cap, separate from the business budget |
| Runs | Manual (Run now), daily or hourly |
Every action an agent takes is written to the audit trail as the agent, with its owner attributed. Deleting an agent is instant; its pending proposals stay in the inbox for a decision.
Proposals
Navigate to: AI → Proposals (or Proposals inbox on this page)
A proposal is exactly what would be done — the invoice lines, the receipt allocations, the bank-line categorisations, the reminder messages — together with the evidence behind it and where each figure came from.
Open a proposal to see the lines, allocations or messages and the evidence. A receipt shows how the money is applied: opening balance first, then the oldest invoice date, each settled in full before the next.
Approval runs the same steps you would take yourself, as you: an invoice is posted and numbered, a receipt is recorded and the invoices it settles are updated, bank lines are categorised, reminders are sent through your messaging connection. Approving needs the same right as doing it by hand.
Nothing happens. Add a note so the requester knows why.
The assistant can propose too. Ask it to "invoice Acme for 3 widgets at 50" or say "Acme paid 1,200 today" and it creates a proposal for you to approve here — it never posts on its own. The same is true for AI assistants connected over MCP: their propose_* tools need a key minted with the create right.
Proposals expire after 14 days; the figures may have moved, so ask for a fresh one.
Agents that run on their own
Besides Run now, an agent runs on a schedule (hourly or daily) and on events you tick when creating it — when statement lines arrive on a bank account, when a customer pays, when an invoice is posted. Either way it ends in proposals; an agent with autonomy Act within a limit applies its own proposals up to the amount you set and proposes above it.
Three presets start you off: Collections (daily reminders), Reconciliation (categorise statement lines as they arrive), Month close (a checklist for your signature: uncategorised lines, overdue receivables, drafts and proposals still waiting — signing records the review; the period itself closes under Finance).
Test connection on this page sends one word through the door and reports the provider, model and latency — use it after changing the provider or key.
What the AI saw
Every assistant answer carries its trace: which tool ran, how many rows came back, which model answered, how many personal-data tokens were masked, and exactly what was sent — the question and the last few turns, never your rows. Open it under the answer.