AI Personal Finance App: How SpendifiAI Actually Works

SpendifiAI is an AI personal finance app that pulls in your bank transactions through Plaid or an uploaded PDF or CSV statement, has Claude categorize every one of them, and only asks you about the ones it is unsure of. From that categorized ledger it finds recurring subscriptions, suggests where to cut, and builds a tax package for your accountant: a five-tab Excel workbook, a PDF summary and a CSV, with deductible spending mapped to IRS Schedule C lines.

We built SpendifiAI at Parkk Technology on Laravel 12, React 19 and PostgreSQL. This article covers what the app does, how each part works in the code, and where its limits are. Most AI budgeting apps give you a chatbot on top of a spreadsheet. The work here went into the plumbing underneath: getting clean data in, deciding when the AI can be trusted on its own, and producing files an accountant can use as they are.

What Is an AI Personal Finance App?

An AI personal finance app reads your transactions and does the sorting you would otherwise do by hand: what category a charge belongs to, whether it is business or personal, whether it repeats, and whether it matters at tax time. Older budgeting tools do this with fixed rules and merchant lists, so they file every Amazon order under “Shopping” and every Venmo payment under “Transfer.” A language model can read the merchant name, the amount, the account it came from and any matching order email, and reason about what the charge probably was.

The catch is that a language model is also confidently wrong some of the time. A useful AI finance app needs a policy for that. SpendifiAI’s policy is a confidence score on every decision, with fixed thresholds that decide whether the answer is saved, flagged, or turned into a question for you.

What SpendifiAI covers

How Does SpendifiAI Get Your Bank Transactions?

There are two ways in, and both feed the same categorization engine. A user can pick either one or use both at once.

How does Plaid work with a finance app?

Plaid sits between the app and your bank. You sign in to the bank inside Plaid’s own window, and SpendifiAI receives an access token rather than your bank password. SpendifiAI requests Plaid’s transactions and statements products for US accounts. It stores the token encrypted, runs a scheduled sync every night, and accepts Plaid webhooks at /api/v1/webhooks/plaid so new transactions can arrive between syncs.

Importing PDF and CSV bank statements

Some people can’t use Plaid because their credit union isn’t supported, and some just don’t want to link a bank. For them, the upload wizard turns a statement into transactions: the statement text (extracted with pdftotext, or the PDF itself) goes to Claude, which pulls out each dated transaction, and the result joins the same queue as Plaid data. The upload pipeline also spots gaps between statements and warns you when a month is missing. That pipeline gets its own write-up in how SpendifiAI imports bank statements.

How Does AI Categorize Transactions?

Transactions go to Claude in batches of 25, with a short pause between calls to stay under rate limits. Each transaction carries its merchant, amount, date, description, payment channel, Plaid’s own category guess, the name of the account, and that account’s purpose: personal, business, mixed or investment. If an order-confirmation email has been matched to the charge, the items from that order go along too.

Claude sends back a category, a normalized merchant name, business or personal, whether the charge is tax deductible and on which tax line, whether it looks like a subscription, and a confidence score from 0 to 1. What happens next depends only on that score.

Confidence scoring: when the AI asks you

At 0.85 or above, the category is saved and marked auto-categorized. You never see it unless you go looking. Below 0.85 the transaction is saved with its best guess and flagged for review. Below 0.60 the prompt requires Claude to write a question with answer options. The question can be about the category, business or personal, a split, or a plain confirm. You answer it in the app, as a multiple-choice pick or in free text, and SpendifiAI interprets the reply.

The prompt also changes confidence based on account purpose. A charge on an account marked business gets +0.15 toward a business category, and a charge on a personal account gets +0.15 toward a personal one. A mixed-use account gets 0.10 taken off, so it produces more questions, which is what you want when a card pays for both groceries and client lunches. If an order email is matched to the charge, Claude is told to score it 0.95 or higher, because it can see exactly what was bought.

The 55 categories behind the model

Claude doesn’t make up category names. It chooses from a seeded list of 55 categories in 13 groups, from Housing and Transportation to a Business Expenses group whose six categories exist mainly for tax purposes. Twelve categories are flagged as typically deductible and carry a tax line, such as Schedule C Line 24b for business meals or Line 18 for office expenses. Because the list is fixed, the tax export can map each category to a line in code instead of asking the model.

Finding Subscriptions and Money to Save

Once transactions are categorized, a nightly job scans them for recurring charges on weekly, monthly, quarterly and annual cycles. It checks against known merchants such as streaming services, utilities and insurers, and also detects repeat patterns from unknown merchants by frequency. It watches for the opposite as well: a charge that used to recur and has stopped arriving. That can mean a cancellation went through or a card was declined, and either one is worth knowing. Each subscription is marked essential or non-essential and tracked as active, unused or cancelled.

The subscription view has been rebuilt several times since launch: a Cancelled tab, an upcoming-payments carousel, and an April 2026 overhaul of detection, de-duplication and cancellation. The details are in how SpendifiAI tracks and cancels subscriptions.

Savings recommendations you can act on

The savings analyzer reads the last 90 days of categorized spending and returns specific suggestions: this subscription, this category, this cheaper alternative. It suggests up to four alternatives per item and caches them for seven days. Every recommendation has three answers, cancel, reduce or keep, and SpendifiAI records projected savings and keeps a monthly history of what you acted on. Savings goals work the same way. You set a target, and the planner builds an action plan from your actual spending, not a generic percentage of income.

The dashboard: “Where to Cut” and the budget waterfall

The dashboard answers four questions. The budget waterfall shows income flowing down through each spending category. Monthly Bills splits recurring charges into essential and non-essential. Home Affordability works out a housing budget from your real income and debts. Where to Cut shows the savings recommendations, one action per card. In late April 2026 the cost-of-living math was reworked. Loan payments are now detected as recurring fixed transfers instead of disappearing as “transfers,” and the category you or Claude chose now takes priority over Plaid’s category when bucketing essentials. That feeds a “Can you live on your paycheck?” comparison of core expenses against detected income. Charitable giving has its own dashboard block, covered in tracking charitable donations in SpendifiAI.

Can AI Find Tax Deductions?

It can find likely ones, which is where most of the value is, because deductions are usually lost by never being noticed. Each categorized transaction carries a deductible flag and a tax category. The export collects those, maps them to Schedule C (and Schedule A for items like mortgage interest and property tax), and splits business spending from personal.

What the Schedule C export contains

1. Tax Summary totals for the year, with the estimated tax bracket
2. Income & Tax Documents W-2 and 1099 income pulled from uploaded forms
3. Schedule C Mapping deductible spending grouped by Schedule C line
4. Deductions by Category the same spending grouped by SpendifiAI category
5. All Deductible Transactions the line-by-line detail an accountant will ask for

Next to the workbook, the export produces a PDF summary and a transactions CSV, and it can also write TXF, OFX and a QuickBooks-format CSV. The full package can be emailed to your accountant from inside the app. The March 2026 overhaul added a self-employment tax estimate (15.3% on 92.35% of net self-employment income) and a bracket estimate based on 2025 IRS brackets and filing status.

The Tax Vault and accountant access

Version 2.0 added a Tax Document Vault. You upload W-2s, 1099s (NEC, INT, DIV, B, R, K and others), 1098s, K-1s and receipts. Claude classifies each document and extracts its fields, but only when the classification scores at least 0.70. Below that, the document is flagged for manual review instead of being guessed at. Extracted fields get a green or amber badge at 0.85 and 0.60. Multi-document PDFs are split at the boundaries Claude detects, and duplicates are caught by file hash and by content (EIN, SSN, form and year).

Accountants get their own account type. A firm can invite clients by link, request specific documents, and leave comments on a document in an annotation thread. Households are supported too: members join by invitation, dependents are stored with tax-relevant fields, and the export combines the whole household’s data.

Is an AI Budgeting App Safe to Connect to Your Bank?

This is the right question to ask of any finance app, AI or not. The answer depends on what it stores and how it’s protected. SpendifiAI never sees your bank password when you use Plaid. The Plaid access token, OAuth credentials for connected email and business identifiers such as EINs are encrypted with AES-256-CBC before they reach the database.

• Two-factor login with a TOTP authenticator app (6 digits, 30-second window, 8 recovery codes)
• reCAPTCHA v3 on registration and login
• Rate limiting on every authentication and sensitive endpoint
• Account deletion that satisfies GDPR and CCPA requests
• Authorization policies on transactions, accounts, questions and subscriptions, updated in v2.0 for households

The AI only sees transaction data. Claude receives merchant, amount, date and description to categorize a charge. It never receives your bank credentials or tokens.

Receipts from Your Inbox

SpendifiAI can connect to Gmail through Google OAuth, to Outlook through Microsoft Graph, or to any mailbox over IMAP. It reads order-confirmation emails and matches them to bank charges. A matched receipt tells Claude what an Amazon or big-box charge actually bought instead of leaving it as “Shopping,” which is why those transactions score 0.95 or higher. The three-phase search behind it, including a fallback for Outlook when search doesn’t index amounts inside HTML tables, is covered in matching email receipts to bank transactions.

The Tech Stack Behind SpendifiAI

Laravel’s queues run all the slow work: categorization, statement parsing, email matching, subscription scans. Your dashboard never waits on an AI call. Inertia lets the React front end behave like a single-page app without a separate API just to serve its own screens. Scheduled jobs are defined in routes/console.php: bank sync daily, subscription detection at 02:00, expiry of stale AI questions (after 7 days) at 03:00, and failed-email retries at 04:00.

How it was built

The first commit landed on February 10, 2026. By May 18, the release/v2.0.0 branch had 226 commits. Most of the build happened in the first seven weeks: 138 commits in February covered the core app, and 81 in March covered the Tax Vault, accountant portal and households. The README from early in the build lists 131 tests and 459 assertions.

Frequently Asked Questions