Tariffic

Inspiration

When the Supreme Court struck down the IEEPA tariffs in February 2026, importers started getting refunds, and finance teams got a mess. One refund covers dozens of shipments; some tariffs come back, some never will; part belongs to sold goods, part to inventory, and some may be owed to customers. That made us think about building a smoother way for the office of the CFO to handle tariff changes.

What it does

Tariffic AI fills three finance seats for an importer. It learns the company's rules from its history, catches mistakes on its own, splits refunds correctly, and credits customers according to their contracts. Every decision goes to the CFO for approval, and Tariffic learns from each approval, so it needs fewer questions over time. When tariffs rise, it also suggests alternative vendors in other countries to source from in the future. And when someone pushes it to cut corners, it holds firm: only verified evidence can change its mind.

How we built it

AI agents do the work; code sets the limits. Agents propose and learn, but only humans approve, and only a gateway can write to the books (a real Beancount ledger). Exact math, permissions and consistency checks live in code, every action is logged, and runs are scored against a hidden answer key. 216 tests pass.

Challenges we ran into

The hardest part was making everything work together: pulling data from the bank, customs records and ERP, reconciling sources that disagree, identifying exactly where each dollar belongs in the ledger, and learning from every approval, all while keeping the books correct.

Accomplishments that we're proud of

A full refund flow, from bank deposit to posted ledger entries, with every final balance matching the answer key and zero wrong postings.
Tariffic finds mistakes in the books on its own and opens a ticket for each one.
It can't be pressured: pushback without evidence changes nothing, while a verified document updates the decision correctly.
Any decision can be explained months later with a single command.

What we learned

Limits belong in code, not in prompts: use AI for language and judgment, and code for money.
Testing against a hidden answer key keeps you honest; "looks right" isn't the same as "is right."
Finance is as much about coordination as math: who does the work, who approves it, and how to prove it later.
A lot about how tariffs, refunds and inventory costs actually flow through a company's books.

What's next for Tariffic

Reusing each approval across similar contracts automatically, using retrieval over contract clauses.
Tracing any tariff dollar from customs entry to customer invoice.
Deeper sourcing advice: comparing landed costs across countries when tariffs change.