AI Auto-Categorization
Wired Church uses Claude AI to analyze incoming bank feed transactions and suggest the correct fund, account, and category for each one. Instead of manually reviewing every utility payment, payroll debit, and vendor charge, the AI reads the transaction description and matches it against your chart of accounts and historical categorization patterns. This guide explains how it works, how to review suggestions, and how to train the system for your church's specific spending patterns.
How It Works
When new transactions sync from your bank, the AI categorization engine processes them in the background:
- Transaction arrives — A new transaction is pulled from your bank via Plaid with a description like "ALABAMA POWER AUTOPAY 03/10" and an amount of -$847.32.
- AI analysis — Claude AI examines the transaction description, the amount, the transaction type (debit or credit), and the date.
- Pattern matching — The AI compares the transaction against:
- Your chart of accounts (account names, categories, and types)
- Previously categorized transactions with similar descriptions
- Common merchant recognition patterns (utilities, insurance, payroll services, etc.)
- Suggestion generated — The AI assigns a confidence level and suggests a specific account from your chart of accounts.
- Result applied — Depending on the confidence level, the suggestion is either auto-applied or queued for your review.
AI categorization runs automatically for all new transactions on connected accounts. There is no button to trigger it — it happens as part of the sync process. If you add or rename accounts in your chart of accounts, the AI immediately incorporates the changes.
Confidence Levels
Every AI suggestion comes with a confidence level that determines how it is handled:
High Confidence
- The AI is very confident in the categorization — the description closely matches a well-established pattern.
- The suggestion is auto-applied to the transaction. The transaction moves from New to Categorized status automatically.
- A small "AI" badge appears next to the category, indicating it was set by the AI rather than a human.
- Examples: recurring utility payments, known payroll providers, established vendor charges that appear every month.
Medium Confidence
- The AI has a reasonable guess but is not certain enough to auto-apply.
- The suggestion is queued for review — the transaction remains in New status with a suggested categorization attached.
- You will see the suggestion in the review screen and can approve it with one click or change it.
- Examples: a new vendor the AI has not seen before but whose name suggests a category, an amount that could belong to multiple accounts.
Low Confidence
- The AI could not determine a likely category.
- No suggestion is made. The transaction remains in New status with no category attached.
- You must categorize it manually.
- Examples: vague descriptions like "ACH DEBIT 48291030," internal transfers with no identifying information, very unusual one-time charges.
During the first month of using bank feeds, expect more Medium and Low confidence results. As you review and correct suggestions, the AI learns your patterns and the proportion of High confidence auto-categorizations increases significantly.
Reviewing AI Suggestions
All transactions with pending AI suggestions are accessible in a dedicated batch review screen:
- Navigate to Finance > Bank Feeds.
- Click the Review Suggestions button in the dashboard header. The button displays a count badge showing how many transactions need review.
- The review screen presents transactions one at a time or in a list view, depending on your preference.
Reviewing in List View
- Each row shows the transaction date, description, amount, and the AI's suggested category.
- A confidence badge (High, Medium) appears next to each suggestion.
- For each transaction, you can:
- Approve — Click the checkmark to accept the AI's suggestion. The transaction moves to Categorized status.
- Change — Click the pencil icon to select a different account from your chart of accounts. The AI records your correction for future learning.
- Skip — Leave the transaction for later review.
- Exclude — Mark the transaction as excluded (e.g., internal transfers you do not want in your books).
Batch Actions
For efficiency, the review screen supports batch operations:
- Select All High Confidence — Check all transactions where the AI has high confidence. One click approves them all.
- Approve Selected — After checking individual rows, click to approve all selected transactions at once.
- Filter by Confidence — Show only Medium confidence suggestions if you want to focus on the ones that need the most attention.
A fast weekly workflow is: open the review screen, approve all high-confidence suggestions in one batch, then work through the medium-confidence ones individually. This typically takes 5-10 minutes per week for a church with moderate transaction volume.
How the AI Learns
The AI improves over time based on your corrections and approvals:
What It Learns From
- Approved suggestions — When you approve a suggestion, the AI reinforces that pattern. Future transactions with similar descriptions will receive higher confidence.
- Corrections — When you change the AI's suggestion to a different account, the AI learns that this description pattern should map to the account you selected.
- Manual categorizations — When you categorize a transaction that had no AI suggestion (Low confidence), the AI adds this as a new pattern.
What It Recognizes Well
After a few weeks of training, the AI typically achieves high accuracy on:
- Utilities — Electric, water, gas, internet, phone. Most utility companies have distinctive transaction descriptions.
- Payroll — ADP, Gusto, Paychex, and direct payroll debits are recognized quickly.
- Insurance — Monthly premium payments from major carriers.
- Office supplies — Amazon, Staples, Office Depot, and similar retailers.
- Subscriptions — Software subscriptions (Microsoft, Google, Adobe, Zoom) with recurring amounts.
- Specific vendor patterns — Any vendor your church pays regularly. After 2-3 correct categorizations, the AI handles that vendor automatically.
What Takes Longer to Learn
- Generic ACH descriptions — Some banks use codes instead of merchant names. The AI relies on amount patterns and timing in these cases.
- One-time purchases — A vendor you pay only once does not build a pattern. The AI may suggest based on the merchant category, but confidence will be lower.
- Split categorizations — If a single bank transaction needs to be split across multiple accounts (rare), the AI currently does not suggest splits. You must handle these manually.
The first month is the training period. Review suggestions carefully and make corrections promptly — every correction improves future accuracy. Churches that actively review during the first month typically see 80-90% high-confidence auto-categorization rates by month two.
Plan Limits
AI auto-categorization is available on Pro and Enterprise plans with the following limits:
| Plan | Monthly AI Categorizations |
|---|---|
| Starter | Not available |
| Pro | 100 per month |
| Enterprise | Unlimited |
A "categorization" is counted each time the AI processes a transaction — this includes both auto-applied high-confidence results and medium-confidence suggestions, whether or not you approve them.
If you reach your monthly limit on the Pro plan, new transactions will still sync but will not receive AI suggestions. You can categorize them manually or wait until the limit resets on the first of the next month. Upgrading to Enterprise removes the limit entirely.
Manual Override
The AI is a tool, not a decision-maker. You always have the final say on every categorization:
- Before approval — Change any medium-confidence suggestion to a different account before approving.
- After auto-apply — High-confidence auto-categorized transactions can be recategorized at any time. Click the transaction, select a new account, and save. The AI will learn from the correction.
- Bulk recategorize — If the AI learned an incorrect pattern (e.g., it keeps categorizing a vendor as "Office Supplies" when it should be "Ministry Supplies"), you can:
- Filter transactions by that vendor name.
- Select all matching transactions.
- Bulk-recategorize them to the correct account.
- The AI will update its pattern based on the corrections.
If you notice the AI consistently miscategorizing a specific vendor, correct it as soon as possible. Each incorrect auto-categorization that goes unnoticed affects your financial reports. A quick weekly review of auto-categorized transactions catches these issues early.
Disabling AI Categorization
If you prefer to categorize all transactions manually, you can disable the AI:
- Navigate to Finance > Bank Feeds > Settings.
- Toggle AI Auto-Categorization to Off.
- All future transactions will arrive with New status and no suggestions.
- Previously categorized transactions are not affected.
You can re-enable the AI at any time. When re-enabled, it will process any uncategorized transactions in the New state.
Frequently Asked Questions
Does the AI see my actual bank account numbers or balances? No. The AI only processes the transaction description, amount, type (debit/credit), and date. It does not have access to account numbers, balances, or any identifying financial information.
Can I use AI categorization without connecting a bank? No. AI categorization operates on transactions that come in through the Plaid bank feed. Manually entered or imported transactions are not processed by the AI.
What happens if I delete a categorized transaction's account from my chart of accounts? The transaction will revert to New status and need to be recategorized. The AI will not suggest the deleted account for future transactions.
Is the AI categorization processed on Wired Church's servers? Yes. Transaction descriptions are sent to Claude AI (hosted by Anthropic) for analysis. No personally identifiable information is included — only the transaction description, amount, and your chart of accounts names. See our privacy policy for details.
Can two staff members review suggestions at the same time? Yes. The review screen uses optimistic locking — if two people approve the same transaction simultaneously, the first approval wins and the second person sees a "Already categorized" notice.