AI Categorization

Reading time: 8 minutes

Learn how Atlas's AI engine auto-categorizes bank transactions, how to review its suggestions, create custom rules, and train it from your corrections. By the end of this tutorial your AI will be categorizing routine transactions with high confidence.

---

Prerequisites

• An Atlas account with Bookkeeper role or above
• At least one bank account with imported transactions
• Some transactions already categorized (the AI learns from history)

---

Step 1: How AI Matching Works

When you import a bank statement, Atlas's AI engine processes each transaction through several layers:

Layer	What It Does
Rule Matching	Checks your custom categorization rules first
Historical Matching	Compares against previously categorized transactions
Description Analysis	Parses merchant names, reference numbers, and descriptions
Amount Patterns	Recognizes recurring amounts (rent, subscriptions, payroll)

Each transaction gets a confidence score from 0 to 100:

• 90-100 (High) — auto-categorized, appears in the Categorized tab
• 60-89 (Medium) — suggested categorization, appears in For Review
• 0-59 (Low) — no suggestion, appears in Uncategorized

---

Step 2: Review AI Suggestions

1. Navigate to Banking > Bank Accounts
2. Select your account
3. Go to the For Review tab

Each transaction card shows:

• The bank description and amount
• The AI's suggested account and payee
• The confidence percentage
• A match reason (e.g., "Similar to 12 past transactions")

Taking Action

For each suggestion:

• Accept (checkmark) — confirms the AI's suggestion
• Edit (pencil) — modify the account, payee, or memo before accepting
• Reject (X) — dismiss the suggestion and categorize manually
• Split — divide into multiple account categories

Tip: Batch-accept high-confidence suggestions by checking multiple items and clicking Accept Selected.

---

Step 3: Create Custom Rules

Rules give you deterministic control over categorization:

1. Go to Banking > Rules
2. Click New Rule
3. Define your conditions:

Condition Options

Condition	Example
Description contains	"AMZN" or "AMAZON"
Description starts with	"SQ *" (Square payments)
Amount equals	99.99 (exact match)
Amount between	50.00 – 200.00
Transaction type	Debit or Credit

4. Set the action:

   • Account — which GL account to assign
   • Payee — standardized vendor/customer name
   • Tax rate — apply a specific tax treatment
   • Auto-approve — skip the review queue for high-confidence matches

5. Set priority — rules with lower numbers are evaluated first

6. Click Save Rule

Example Rules

Rule Name	Condition	Action
Office Rent	Amount = 2,500, Monthly	Office Rent Expense
AWS Hosting	Description contains "AWS"	Cloud Hosting Expense
Stripe Deposits	Description starts with "STRIPE"	Merchant Revenue
Payroll ACH	Description contains "GUSTO"	Payroll Expense

---

Step 4: Train from Corrections

Every time you correct an AI suggestion, the system learns:

1. Correct a categorization — the AI remembers this vendor + amount pattern
2. Reject a suggestion — the AI reduces confidence for similar future matches
3. Accept a suggestion — reinforces the pattern

Feedback Loop

Import → AI Suggests → You Review → AI Learns → Better Suggestions

After approximately 50-100 corrections for a given vendor or pattern, the AI typically reaches 90%+ confidence and auto-categorizes without review.

Viewing AI Learning History

1. Go to Banking > Rules > AI Insights
2. See which patterns the AI has learned
3. Review the confidence trends over time
4. Identify patterns that need manual rules (low-confidence vendors)

---

Step 5: Configure Confidence Thresholds

Customize how aggressive the AI is:

1. Go to Settings > Banking > AI Settings
2. Adjust thresholds:

Setting	Default	Description
Auto-categorize threshold	90%	Transactions above this skip review
Suggest threshold	60%	Transactions above this get a suggestion
Learning rate	Medium	How quickly AI adapts to corrections

3. Click Save

Conservative approach: Raise the auto-categorize threshold to 95% if you prefer to review more transactions manually.

Aggressive approach: Lower it to 80% if you trust the AI and want less manual review.

---

Troubleshooting

Issue	Solution
AI not suggesting anything	Need more historical data — categorize 20+ transactions manually first
Wrong suggestions repeated	Create a rule to override the AI for that pattern
Confidence not improving	Check for inconsistent categorization of the same vendor
Rules not applying	Verify rule conditions match the bank description exactly

---

Related Guides

• Banking & Reconciliation — statement import and reconciliation workflow
• Bank Reconciliation Tutorial — complete reconciliation walkthrough
• Chart of Accounts — organizing accounts for clean categorization

What's Next

Now that your AI is training:

• Create rules for your top 10 vendors to ensure consistent categorization
• Review AI Insights weekly to spot new patterns
• Run a bank reconciliation to close out the month
• Set up auto-approve for high-confidence recurring transactions