Hamsa SoftwareExact limits
The agent can read and change only what was agreed: these records, this action, up to this amount.
AI INSTITUTION / APPROVALS FOR AI AGENTS
Approval granted.
Then the facts change.
AI Institution lets an AI agent prepare a refund without handing it an open cheque book. AI does the legwork. Your team keeps the authority.
Separate sign-off
The agent can propose a change but can’t approve it. The right people sign off the exact amount and currency.
An independent check
Facts are rechecked before money moves, and a separate check confirms the result before it is recorded as done.
WHO DOES WHAT
Every task has people, limits and evidence.
Prepares the change
Turns the task into proposed work. Every proposed action still has to fit the agreed limits, policy and budget.
Signs off or stops it
Approves the exact action, denies it or stops the task. A proposal cannot give itself permission.
Confirms the result
Judges the evidence separately from the agent. Missing evidence never becomes a signed-off success.
FOR ENGINEERS / RECORDED REFERENCE RUN
From approved work
to a durable record.
A recorded run of the AI Institution reference kernel with SQLite: a bounded action, independent evidence checks and a signed-off result. Reopening the database reuses the same receipts.
—Recorded audit events
—Runtime verdict
£0Model cost: no model called
RECORDED REFERENCE EVENTS / NO LIVE AGENT
- Ready to replay the captured product execution.
INSPECT THE RUN
Receipts, not
just status badges.
The file contains the runtime’s actual audit entries, policy decision, evidence, certification and receipts from the first and reopened executions.
Download runtime evidence ↓Recording & reproduction notes ↗Try the interactive business scenario ↓Inspect the actual certification record
Loading recorded evidence…
What the wider product supports.
The repository implements the capabilities below; not all of them run on this page.
Reviewed authority
Roles, delegation, precise permissions, budgets and governed amendments. Plain-language proposals need human adoption.
Teams & recurring work
Bounded planning, role-based handoffs, schedules, cancellation and recovery from durable History.
Cost & accountability
Per-Mission model usage and cost, owned exceptions, an operator inbox and evidence-linked results.
Where it is today
- Early development, private and self-hosted. There is no hosted AI Institution service, and a customer finance workflow still needs its own integrations, checks and sign-off before use.
- The runtime recording replays an actual local run of the reference kernel with SQLite. Its action is deterministic: no live model, GitHub write or payment. Reopening the history shows that records are not duplicated; it does not show recovery of a real refund.
- The refund lab is a smaller teaching model built separately for this page, with simulated Stripe. It does not run the service itself, its durable storage, credential handling, real independent checks or live payments. Records reset on refresh.
- The repository’s refund workflow has been exercised with a scripted model and simulated Stripe.
- No model calls or billable services run on this page. The $50 limit is a fictional refund policy, not a price.
aiInstitutionGOVERNED WORK / REFERENCE DEMO
Where it is today TRY IT / REFUND DESK
One refund.
Two approvals. One last check.
Set the task, inspect the change the agent proposes and follow the evidence through to the final decision.
Interactive refund lab · prepared producer · simulated StripePlay the prepared currency-change scenario. Both roles approve USD; the final check finds EUR. No provider write is allowed.
MISSION-001Draft
Refund $42.00 USD
Charge ch_sample_201 · fictional data
You can act out both approval roles in this walkthrough. The reference product requires distinct authorised principals; this page has no identity or authentication system.
ECONOMICS / PROVIDER-EVIDENCED EFFECTS
A refund is an effect.
Certification is a separate verdict.
If verification fails after a succeeded refund, the ledger must still tell the truth about the money. If the provider is pending, it must not invent a success.
Mission evidence
WHERE THIS COMES FROM
A reference workflow,
made explorable.
This walkthrough adapts AI Institution’s refund-desk reference mission: read policy and charge, request approval of an exact amount and currency, obtain separate compliance sign-off, re-check facts before the effect, record the provider receipt and verify the result.
It is a teaching model with simulated Stripe; see where the project is today.
Explore the project