Fiscent

Our own product: finance operations run by agents

Fiscent is a product we designed without a client: a finance operations platform where AI agents do the routine work of payables, receivables, spend, treasury and the month end close. The name is fiscal and agent, put together. We sell it as a Figma kit for teams building finance operations, payments or AP automation products, and it is how we show what designing for AI agents looks like when the money is real.

Finance is the category where agent design is hardest to fake. A wrong summary is an annoyance; a wrong payment is a loss, a vendor dispute and an audit question. So the design question was never what the agents can do. It was how a controller stays in charge of them without reviewing every line by hand.

Autonomy per action, not per agent

Most agent products give an agent one setting: on or off, assisted or autonomous. Fiscent sets autonomy for each action an agent takes, on a scale from watch to suggest, ask first and act. A payables agent can be allowed to match invoices to purchase orders on its own while it still asks before paying a new vendor or anything over $25,000.

Every agent also carries daily limits that reset each morning, a mission written in plain words, and a pause switch. Trust is easier to give in small steps, and easier to give at all when it is easy to take back.

Brief an agent like a new hire

Adding an agent is a wizard that reads like briefing a person. You describe the job in plain words, Fiscent drafts the rules from that description, replays last month to show what the agent would have done, and only then asks for limits. A new agent starts with a watch week, in which it proposes everything and does nothing, before it is allowed to move money.

Policies with versions, diffs and drafts

Every rule the agents follow lives in a versioned policy. A change starts as a draft, is compared with the live version rule by rule, and is activated with an approval. The history stays, so the question of which rule allowed a payment in March has an answer in March’s version, not today’s.

Proof for auditors, a lock for bad days

Every decision is logged with the policy that allowed it and the person who approved it, and the log can be shared with an auditor or exported as an audit pack for a date range.

For the day something goes wrong, there is an emergency lock. It can stop payments, stop everything, or lock one agent, and it is one screen away from anywhere in the product. There is even a lock drill, because a control nobody has practised is a control nobody trusts.

A system made for agent products

The file is built on 47 variables in two collections and 73 components in nine documented groups, with 16 text styles in Figtree. The tokens are semantic for the things this product talks about, money, status and autonomy, and the components cover what agent products need and most kits do not have: agent tiles, approval rows, limit controls, overlays and charts with hover tooltips.

The 73 screens run across nine flows, from the morning briefing to the settings change log, at 1440 wide, with 97% of frames in auto layout. Five flows are wired as clickable prototypes: hiring an agent, releasing a payment, shipping a policy change, the emergency lock and sharing proof with an auditor. The kit is on our products page.