A push-first transaction review loop that meets the SMB owner in the moment a charge happens, lets AI handle the obvious, and keeps the accountant out of email jail.
By the time Casey opens her review queue, the receipts have scattered, Avery is on location, and the context behind a $524 Amazon charge has evaporated. The conversation that follows — email, text, voicemail — is a tax on both of them.
Transactions accumulate quietly all month
EOM hits — Casey reviews a pile she can't categorize alonepile-up
Emails + calls Avery for every ambiguous chargecontext-loss
Sees the email two weeks later, on set, can't recall the purchaseno-memory
Half-answers, drops thread, never uploads the receiptno-response
EOM slips. Both sides frustrated. Repeat next month.EOM-miss
Amex / Ramp / PayPal / Amazon webhook fires within seconds
≥95% confidence → auto-categorize. <80% → flag for Avery.
Lock-screen push the same hour: 'Amazon $524 — production gear?'30s
One tap to confirm. Or delegate to Maya. Or attach a photo.
Quiet escalation if 3 days pass — SMS, then email card, then Casey.
EOM arrives at 98% done. Casey reviews exceptions, not the pile.
“If I have to chase Avery one more time, I'm dropping Lion Snacks.”
“I'll answer when I have a second. I always mean to.”
The current loop fails in four different places, so the fix is four connected moves. Each one stands alone, but the leverage comes from running them as a system.
Where Leafly has a partner connection (Ramp, Amex Business, Amazon Business, PayPal webhook), pull line-items, not totals. Gaffer tape $48, Sandbags $138 — fact-based categorization, not pattern matching.
Ambiguous charges fire a lock-screen push to Avery within minutes, with the top AI-suggested category pre-filled. The reply path is one tap, while the memory is still hot.
No reply in 24h → SMS magic link. 48h → email action card (Stripe-style, reply inline). EOM−5d → Casey gets a single 'needs your attention' tile. Push fatigue is real, so unrelated charges batch into a daily digest.
Casey's dashboard shows clients by % categorized, not by alphabetical order. Exception queue surfaces only what AI + Avery couldn't resolve. The week-of-EOM view is a single screen.
Lock-screen with AI suggestion. Tap to confirm.
Browser-based reply for Avery's teammates without the app.
Reply buttons inside the email body. No app jump.
A single tile on Casey's dashboard, only for the unresolved.
The four-layer strategy is the shape of the product. Underneath it is an LLM-native architecture that changes the question accounting software is even asking — from "how do we categorize transactions?" to "how do we preserve financial context before it disappears?"
Transaction
Rule / OCR
Categorization
Human review
Month-end cleanup
Transaction
Context ingestion (bank + email + receipt + history + org memory)
LLM reasoning layer
Conversational clarification
Adaptive escalation (push · SMS · delegate)
Continuous memory + audit trail
Bank feeds, business cards, PayPal metadata, email receipts, OCR uploads — fused with vendor history, prior approvals, location & timing, recurring project patterns.
Not 'Uber → Travel'. Instead: 'this resembles prior production transportation approved by Avery during late-night studio shoots.' Confidence-scored, evidence-linked.
Avery can ask: 'Why is this flagged?' · 'Split this expense.' · 'Find the receipt.' · 'Who handled this last time?' Human language in, accounting structure out.
Tone shifts with stakes — urgent at EOM, casual mid-month, delegate-aware when Maya is the right asker. Replaces reminders with judgment.
Maya → props. Jordan → lighting. Recurring Airbnb → crew lodging. Late-night Uber near studio → production transport. Each client converges on its own model.
Every confirmation, correction and delegation reweights the model. Manual review drops month over month as confidence compounds.
Casey doesn't have to trust a black box — she can read why. That's what unlocks accountant confidence, compliance, and the right to reduce human review.
Accounting automation tool.
AI-powered operational memory system for SMB finance teams.
The full mobile flow is interactive here. Below: the three moments that decide whether this whole strategy works, plus Casey's portfolio view.
Clients sort by close-completeness, not alphabet. The progress bar decomposes the month so Casey can see at a glance how much Leafly handled, how much Avery cleared, and what's still waiting.
Open full dashboard →Amazon doesn't have a public consumer Order API. Phase 1 would lean on Gmail/Outlook receipt parsing; Amazon Business is the real long-game integration.
Active production weeks could generate 10–20 charges/day. Anything sub-$500 from a known merchant batches into a daily digest. Only new payees or large amounts interrupt.
Right now a delegated charge could become a second black hole. Maya gets an accept/decline; if she's silent for 24h, the ball returns to Avery automatically.
Connecting Ramp, inviting Maya, choosing what auto-categorizes — there's a whole first-run flow I'd want a sprint for, not a screen.
Target should be 'uncategorized $ at EOM < 5% of total spend' — not count of transactions. Need a baseline measurement before we can claim a delta.
When Avery corrects a suggestion, that signal should reweight the merchant + amount + time-of-day pattern for this specific client. Each client converges on its own model.
Six screens, two minutes — Avery gets a push and resolves a charge without ever opening her email.
Walk through the prototype