Boards and audit committees
Answer who approved each AI action that touched money through Teradius.
Pre-release. Early access by invitation.Governed AI for finance
The approval and evidence layer for AI agents that touch money.
An agent wants to pay, invoice or change payroll. Teradius stops it at a named person, binds approval to the exact effect, re-checks authority and keeps signed evidence.
Built forBoards and audit committeesCFOs and finance directorsFinance operationsThe people who own the agents
A fictional purchase request traced end to end. An AI agent proposes AUD 12,480 across four lines. One exception is held, reviewed by the budget owner and resolved. Finance director Asha Singh approves the exact amount, recipient and version, AUD 12,480 to Sample Supplier (fictional), request version 2, under reference R-1042. Authority is re-checked at execution. The receipt carries the same reference, R-1042, with the decision record and board pack.
- An agent proposes; a person approves the exact effect
- Authority re-checked at the moment of execution
- Signed evidence the board can read
The problem
AI agents are reaching the money. Who approved that?
Boards are being asked to say yes to AI in finance without a way to prove that any single payment, invoice or payroll change was approved by a person.
An agent that can read can soon act. Once it can pay a supplier, raise an invoice or change a payroll record, the question stops being whether the model is good and becomes who approved that, against what, and whether it can be shown later.
Approvals by email
An approval to pay the supplier says nothing about which amount, which recipient or which version of the bill was approved.
Logs are not evidence
A log records that something happened. It does not show that the person who approved it saw the same effect that executed.
The board sees a summary
Board packs arrive late and curated. Goals, decisions and outcomes are not linked in a way the board can verify.
Not a threshold gate
Teradius binds approval to one exact effect, not a limit: change amount, recipient or version and it refuses.
A threshold gate asks one question: is it under the limit? Teradius asks another: is this the exact effect a named person approved, and may it still run?
Read more
| Threshold gate | Effect-bound approval | |
|---|---|---|
| What is approved | Anything under a limit. | One effect: this amount, this recipient, this record version. True now |
| The bill changes after approval | Typically still passes if under the limit. | Refused. The approval no longer applies. True now |
| The approver loses authority before execution | Typically still executes. | Re-checked at execution. Refused. True now |
| The agent is stopped or revoked | Depends on the integration. | Re-checked at execution. Refused. True now |
| The request is retried | May act twice. | Returns the first receipt. Never acts twice. True now |
| Who approved, and what they saw | Often a rule or a role, without the effect they saw. | A named person, the exact effect as reviewed, a stronger sign-in step. True now |
| What remains afterwards | Typically a log entry. | A signed decision record, checkable without relying on Teradius. True now |
Pay supplier bill B-2026-0912Proposed by an AI agentAccounts-payable agent
- ProposedPay AUD 1,250.00 to Example Supplies for bill B-2026-0912, due 14 Oct 2026. Reason: bill matched to order and receipt.
- ReviewedBound to the exact effect: AUD 1,250.00 · Example Supplies · bill version 3
- ApprovalNeeds the Finance manager, with a stronger sign-in step
- ExecuteAuthority re-checked at execution · receipt issued · evidence recorded
Approve AUD 1,250.00 to Example Supplies, due 14 October?ApproveRefuse
If the amount, the recipient or the bill version changes after review, this approval no longer applies and the action is refused.
Approval binding, the re-check at execution, receipts and decision records: true now. The approval screen a person uses: in development.
One governed action, end to end
From a board decision to one exact approval, and back to proof.
Every money action that passes through Teradius takes the same path, whether an agent or a person proposed it. Teradius is not an accounting system you must replace. It governs the approval; payments are made by your bank or by the accounting system you already run.
Propose. Review. Approve. Prove.
Propose
Review
Approve
Execute and prove
Reduce debtor days to 45
- Accountable: CFO
- Guardrail: credit notes 0.8% of sales
- Health from published numbers
- 01
Set the intent
A directive, one accountable executive and its guardrails, frozen at issue.
- 02
Propose one effect
A person or an AI agent proposes one typed action, with its reason.
What the person saw
Issue invoice INV-2026-1187Proposed by an AI agent- Amount
- AUD 12,480.00
- Recipient
- Sample Customer (fictional)
- Source
- Invoice INV-2026-1187 · version 7
- Reason
- Goods delivered and signed for
Illustrative product screen with sample data. No customer, person or figure shown is real. - 03
Keep the decision with a person
A named person approves the exact effect, with a stronger sign-in step.
Approve invoice INV-2026-1187, AUD 12,480.00, to Sample Customer (fictional)?ApproveRefuseSame effectIf the amount, the recipient or the invoice version changes after review, this approval no longer applies and the action is refused.
Illustrative product screen with sample data. No customer, person or figure shown is real. - 04
Execute, and return the proof
Authority is re-checked at execution. A receipt and a signed record remain.
What the organisation keeps
Recorded- Decision
- Approved by the Finance manager, named on the record · stronger sign-in confirmed
- Execution
- Authority re-checked at execution · receipt RCT-88214
- Evidence
- Decision record DR-2026-04418 · signed and chained · checkable without relying on Teradius
Illustrative product screen with sample data. No customer, person or figure shown is real.
The board sees the chain: goal, decision, exact effect, evidence.
Approval bound to the exact effect, authority re-checked at execution, receipts, decision records and the Board Room: true now. The approval screen a person uses: in development.
Read more
AI can propose the work. The decision stays with a person. For every action that passes through Teradius, the effect, the authority and the evidence are kept together.
The board says what must be true: a directive, one accountable executive, key results, and the guardrails that must not be breached. The Board Room holds it, built from published numbers and frozen at issue.
A person, or an AI agent, proposes one typed action from the catalogue and gives its reason. An agent can propose only from that catalogue. Here, an agent proposes an invoice.
A named person reads the exact effect, the amount, the recipient and the record version, and approves with a stronger sign-in step. If any of them changes afterwards, the approval no longer applies and the action is refused.
Authority is checked again at the moment of execution. The action leaves a receipt, so a retried request returns the first result and never acts twice, and a signed decision record that can be checked without relying on Teradius. Signed records and receipts are the evidence. The Board Room links them back up the chain to the directive they served.
An agent, or a person, proposes an action from a typed catalogue: pay a bill, issue an invoice, change a payroll record. It shows exactly what it would do, and why.
A person reads it in plain words. The review is bound to the exact effect: the amount, the recipient and the record version on screen.
A named person approves with a stronger sign-in step. Change the amount, the recipient or the version afterwards and the approval no longer applies.
Authority is checked again at the moment of execution. The action leaves a receipt and a signed decision record the board, and an auditor, can read.
Teradius Strategy
Governance the board can read, not a dashboard somebody prepared.
Example Holdings (fictional)Board pack · issued 30 Sep 2026 · frozenBuilt from published numbers · verified
Revenue YTDAUD 14.2m+8.1% vs plan
Debtor days52target 45
Guardrails breached1 of 12capacity 94%
Grow recurring revenue to AUD 18.0mCEO · 62% computed from published numbersGuardrail: Churn 2.1% (limit 3.0%)Healthy
Open the second entity by Q2COO · 35% computed from published numbersGuardrail: Capex AUD 1.9m of 2.4mWatch
Reduce debtor days to 45CFO · 80% computed from published numbersGuardrail: Credit notes 0.8% of salesHealthy
Launch the service contracts lineCCO · 20% computed from published numbersGuardrail: Delivery capacity 94% utilisedAt risk
HealthyWatchAt risk
Board Room
Directives, key results and guardrails from published numbers, frozen at issue.
True nowOpen a second distribution centre next year?
HoldOpen in Q1Open in Q3
HoldRevenue AUD 21.4m to 23.1m · Margin 11.2% to 12.6% · Cash AUD 3.1m low point
Open in Q1Revenue AUD 24.8m to 28.9m · Margin 9.4% to 12.1% · Cash AUD 0.9m low point
Open in Q3Revenue AUD 23.2m to 26.0m · Margin 10.6% to 12.4% · Cash AUD 2.2m low point
Assumptions on record: 4Ranges from a calibrated model. Never a single number.
Strategy Lab (in design)
Test a decision on a snapshot, compared as ranges. Designed, not yet built.
In designRead more
Board Room
Board directives with objectives, key results and guardrails. Packs are built only from published numbers and frozen at issue. A target met while a guardrail is breached is reported as unhealthy, never as success.
Strategy Lab (in design)
A sandbox for testing a decision on a snapshot of the real figures, with scenarios compared as ranges and the assumptions written down. Designed, not yet built.
What is built today
Plain labels for every capability.
Teradius is pre-release. Every capability is labelled true now, in development or in design, and nothing is generally available. The list below is generated from the same source as the What is built today page. Teradius also carries its own optional modules, from records to agents, each labelled the same way and each running on one Core.
True now6
Who it is for
Boards, finance leaders and the people who own the agents.
CFOs and finance directors
Let agents near money: each action through Teradius needs a named approval.
Finance operations
One plain-words way to approve each job, with the reason attached.
AI owners and technology leaders
Agents propose typed actions within budgets and can be stopped.
Read more
Boards and audit committees
Answer the question who approved that for any AI action that touched money through Teradius, and see goals, decisions and evidence connected.
CFOs and finance directors
Let agents near payments, payroll and supplier records knowing that no action that passes through Teradius executes without a named person's approval of the exact effect.
Finance operations
One way to approve each job, in plain words. An agent-originated proposal is marked as such, with its reason attached.
AI owners and technology leaders
Agents can only propose typed actions, carry budgets, and can be stopped at any step. Every agent-assisted job has a person-only path, so the business keeps running without a model.
Our principle
Let AI do the work. Keep the decision, and the proof, with people.
Deterministic for money, tax and the ledger. AI only where judgement adds value. An agent may propose and explain; it never executes a binding effect without a person's approval. Plain language everywhere, especially on the approval screen.
Foundations
Exact money
Amounts are calculated with exact arithmetic. No rounding drift.
Evidence, not logs
Signed decision records, and receipts so a retried request is not counted twice.
Separate by design
Each customer's approval, evidence and board records are kept separate, and the separation is tested end to end.
Australia first
Designed for Australia first. Jurisdiction specifics are designed as packs.
Early access, by invitation.
Teradius is pre-release. Early access is limited to a small number of organisations, by invitation, and we expect to select design partners who want a say in how approval and evidence for AI agents should work. A briefing shows the approval and the evidence on our pre-release environment, with synthetic data.