Expensicat Agents

Freelancers and small teams do their own books. The software industry's answer to that has mostly been better filing cabinets: prettier dashboards, faster OCR, smarter search. All of it still waits for you to show up and do the work.

This year the industry found a better answer. Ramp put AI agents to work inside corporate finance teams, enforcing policy, coding transactions, and reviewing expenses, and it worked well enough to reprice the whole category. But read their launch posts carefully and you'll see who they're for: controllers, AP teams, companies that onboard five hundred employees at a time. The agents make an existing finance team faster.

Our customer doesn't have a finance team to make faster. So we built the team.

Why agents, and why for you

Expensicat Agents rest on one observation: almost everything in small-business bookkeeping is routine, and the part that isn't routine doesn't need software. It needs you, for about four seconds, with the right context in front of you.

A €17.44 charge from Resend on a developer's account is infrastructure. It was infrastructure last month and it will be infrastructure next month. No human should look at that three times a year. But a €530 charge from a new vendor the same week you took a personal trip? That's a judgment call, and no model should quietly make it for you.

So Expensicat splits the work exactly there. Deterministic logic and learned rules handle the confident bulk: free, instant, no AI theatrics. It judges only the ambiguous remainder and writes down its reasoning. And the genuinely undecidable becomes a question addressed to the only person qualified to answer it: you.

The promise, in one line: your books are always current, and anything that needed your judgment is waiting in one place, with the reasoning attached.

Closing the loop

Most automation is a one-way pipe: detect something, fire a notification, hope. We think the difference between automation you trust and automation you uninstall is what happens after the notification. We call it closing the loop, and it shaped nearly every design decision.

Expensicat asks instead of guessing. When something is genuinely ambiguous, you get a question with tap-to-answer options, in the app and on WhatsApp if you've connected it. "Is the Resend charge business infrastructure or personal?" One tap. Done. It never asks you to make its decision for it; it asks for the one fact it's missing.

Answers arrive however you answer. Tap a button. Type a reply. Send a photo of the receipt it asked about and it gets filed, matched to the right transaction, and confirmed with an "Added!", followed by the natural next question if one exists ("What was this for?"). One question at a time, never a pile.

Your own actions count. If you categorize a transaction yourself while a suggestion is open, the suggestion notices and gets out of the way. The queue never nags you about things you already handled. Obvious? Most tools don't do it.

Everything it does is on the record. Open any transaction and the activity is right there: categorized as Software, by a learned rule, on Jul 8, with an undo button. Not a log file. A working surface where undo actually undoes.

And it remembers. Correct the same vendor three times and it becomes a rule: visible, yours, deletable. Answer a question once and the answer sticks. Your finance policy isn't something you configure in settings; it's something Expensicat learns from how you already work. Month six is nothing like week one.

One more thing that fell out of this, which we haven't seen anywhere else: because every notification and every question is part of the same system, Cat, the assistant you chat with, knows about all of it. When your phone says "your receipt from Tallinna notar has been matched," you can ask Cat "what was that about?" and get an actual answer. One system, one memory, one voice.

What's live today

Bookkeeping. Expensicat categorizes synced transactions and matches receipts and invoices against bank transactions. The confident cases are handled by rules and recorded; the ambiguous band comes to you as suggestions with reasoning, including the lookalike matches that other tools silently get wrong. Suggestions surface where you already work: pills on the transactions register, a Suggested filter with approve-all, the review queue on your homepage, and cards on the documents themselves.

Receipt chasing. Expensicat notices a transaction has no receipt and asks for it: same-day for larger expenses, batched politely for the rest, with a monthly backstop so nothing slips into your accountant handoff. Reply with a photo and the loop closes itself.

Subscription audit. Duplicate charges and price creep on recurring vendors, flagged with the evidence. Unglamorous, and worth real money over a year.

Tax reserve (Italian forfettario first). As income lands, a running "set aside this much" figure, computed from your regime's actual rules and never guessed by a model, plus a heads-up as you approach the €85k ceiling while there's still time to plan. More regimes follow.

Questions, everywhere. All of the above escalates through the same one-tap question system in the app and on WhatsApp. Telegram and more channels follow on the same architecture.

What's next

The lineup grows in a deliberate order. Next is the one we think matters most and that nobody builds for people our size: getting paid. Spotting the overdue and the should-have-been-invoiced, then drafting the chase in your voice, firm where it needs to be, always with your finger on send. After that, the close sweep that hands your accountant a complete package instead of a folder of questions, and the CFO glance (runway, concentration, trends) that's only honest because the earlier work kept the books clean.

What we will not do

This part matters more than the feature list.

Everything starts as a suggestion. Acting on its own is earned per kind of task, with your explicit confirmation, and you can dial it back any time; a kill switch is always one click away. Reversible actions carry a one-tap undo; irreversible or outward-facing ones simply don't happen without you. No money moves itself. Nothing gets filed with any authority. Your accountant stays, better-fed, not replaced. Numbers come from code, never from a model improvising. And your books remain yours: exportable, always.

You'll notice we never call any of this "autonomous." That's deliberate. The industry burned that word; accounting forums are full of people cleaning up after software that promised to drive itself. What you actually want isn't a system that acts without you; it's a system that does the routine flawlessly, shows its work, and knows exactly when to ask. Agency, not autonomy.

Pricing and availability

Agents are part of every Expensicat plan (Solo at $9/month, Team at $15/month), not a separate AI add-on. Their work draws on the monthly usage allowance your plan already includes: stay within it and there's nothing extra to pay; go past it and you're only charged for what you actually used, counted to the document and the token, visible on your usage dashboard the whole time. We priced it this way on purpose: the agents are the product, not an upsell bolted onto one.

Right now they're rolling out to early-access workspaces while we tighten the loop with real books, ours included: Expensicat runs its own finances through these agents. Everything is free during the beta, early adopters lock in current rates, and there's a 14-day trial when we launch. If you're doing your own books, talk to us. We'd like them to be the ones that are always current.