Tax Deduction Finder: How SpendifiAI Scans Your Bank Data

A tax deduction finder reads your spending and your situation and tells you which deductions and credits you probably qualify for, before tax season instead of during it. SpendifiAI’s finder checks a catalog of 69 deductions and credits three ways: it scans your bank transactions, matches your financial profile against eligibility rules, and asks short questions about the rest. It shipped on 20 March 2026 together with Plaid bank statement imports and a six-step onboarding pipeline that runs the whole analysis the moment you connect a bank.

All of this is one commit, 1f08a56, in the public SpendifiAI repository: 48 files and 4,955 lines added. The numbers in this post come from that code. For what SpendifiAI is as a whole, start with the SpendifiAI overview.

What Is a Tax Deduction Finder?

Most people find deductions by memory in April: a home office, a few subscriptions, maybe the mileage. A finder turns that around. It starts from a fixed list of what the tax code allows and works backwards through your records to see which items apply. The value is in the list being complete and the search being boring and thorough, which is exactly the work people skip.

The catalog: 69 deductions and credits

The finder’s knowledge lives in a seeded table, not in a prompt. Each entry has a name, a category, keywords to look for in transactions, eligibility rules and, where it applies, a maximum amount. The categories break down like this:

How Does SpendifiAI Find Tax Deductions?

TaxDeductionFinderService runs three detection modes, and each one is designed not to overwrite the others.

Mode 1: scanning bank transactions

For each deduction that has keywords, the finder searches the merchant names and descriptions of your transactions for the tax year and totals what matches. A confidence score starts at 0.5 and rises 0.05 per matching transaction, capped at 0.95, so one coffee-shop charge never looks as certain as fifty. Rates the tax code applies are applied here too: business meals are estimated at 50% of the total, following IRS Publication 463.

Mode 2: matching your financial profile

Some deductions never appear as a transaction. Eligibility for them depends on who you are: employment type, whether you are self-employed, whether you work from home, and your filing status. The profile matcher checks those rules, and it deliberately skips anything the transaction scan already found or that you have already decided on.

Mode 3: a short questionnaire

What neither mode can settle becomes a question, served five at a time. Answer yes and the deduction becomes eligible; then you can claim it with the actual amount. Nothing is claimed on your behalf.

Is this an AI tax deduction finder?

Not in the part that picks deductions, and that is deliberate. The transaction scan matches each deduction’s keywords against merchant names and transaction descriptions; the profile check and the questionnaire are plain rules. No model decides what you can deduct. AI does the messy work elsewhere in SpendifiAI, turning bank transactions into clean spending categories, and the onboarding pipeline runs it right before the finder. Deduction logic should be something you can read and audit, not a guess.

Where Found Deductions Go Next

A deduction you found but cannot hand to your preparer is not worth much. The finder sits inside SpendifiAI’s Tax Center, which was built in February and already did the export side of the job.

The Tax Center and Schedule C mapping

The Tax Center groups deductible spending into expandable categories, each mapped to its Schedule C line, and lets you drill down to the individual transactions and receipt line items behind every total. When the drill-down was built, a subtle bug surfaced: PostgreSQL returned sums as strings, and the browser concatenated them instead of adding them. The fix casts every total to a number on the server.

Exports for you and your accountant

Exports come as a spreadsheet and a PDF tax report. The PDF opens with an IRS form quick reference listing every Schedule C field from line 8 to line 30 and the Schedule A fields, with amounts, including the $0 lines, so you know exactly what goes in every box. Accountants linked to your account can download the same export, named with your name, from their client list.

The Tax Center’s category totals depend on SpendifiAI’s AI categorization, which decides how sure it is before it acts:

So the export totals rest on a model with a person in the loop whenever it is unsure, and the deduction finder rests on rules you can read. The two meet in the Tax Center.

Plaid Bank Statements: Importing Further Back

SpendifiAI already synced transactions through Plaid. The catch, which the Connect page spells out, is that the initial transaction sync covers about 90 days. A tax year is 365. That gap is the reason this release adds Plaid Statements: the official PDF statements from the bank, fetched through the same connection.

The DownloadPlaidStatements job lists the statements available for your checking and savings accounts, downloads each PDF, parses it, checks every transaction against what is already in your account, imports only the ones that are new, then hands them to categorization. A panel on the Connect page lets you pick a date range and “import further back”.

How the Onboarding Pipeline Works

An empty dashboard is where finance apps lose people. The new onboarding page has four stages (connect bank, connect email, a quick profile, then processing) and the processing stage runs ProcessOnboardingPipeline, a queued job with a fixed order:

1. Sync bank transactions Pull every connected account through Plaid.
2. Request statements Ask Plaid to refresh statements for the longer history; the download happens when the webhook fires.
3. Scan email for receipts Process order emails from any connected inbox.
4. Categorize with AI Run every pending transaction through categorization.
5. Detect subscriptions Find recurring charges.
6. Reconcile orders Match receipts to bank charges, then fire an onboarding-complete event that sends a notification with the results.

The profile you fill in during onboarding is the same profile the deduction finder’s second mode reads, so by the time processing finishes there is data for all three modes to work on.

Fixing 19 Failing Tests

The same commit fixed every test that had been failing before it, 19 in all, leaving 169 tests and 608 assertions passing. Several fixes were real bugs the tests had been reporting:

• Deductions showed as “Deduction #N” instead of their names and descriptions.
• Profile-matched deductions came back as $0; they now take their estimate from the catalog’s maximum amount.
• An enum comparison bug in the dashboard’s subscription status check.
• A route model binding clash, fixed by renaming {provider} to {emailProvider}.
• Gmail token refresh failures now mark the connection as errored, and scheduled email sync skips inactive connections.

New tests came with the new code: feature tests for the deduction finder, Plaid statements and onboarding. By the v2.0 milestone eleven days later the suite had grown to 225 tests.

Frequently Asked Questions