7 data tools. 33 seconds. A completed investigation.
Other platforms flag a variance and leave the work to you. Arvexi Cortex's investigation agent queries your GL, compares prior periods, checks IC counterparts, and produces structured findings with specific transaction evidence. Autonomously. Each finding includes a description, amount, classification, and recommendation. Your team reviews completed analysis instead of starting the investigation from scratch.
The problem
Your accountants are doing detective work every month. The same detective work.
For every variance, the process is identical: pull the GL detail, compare to last month, check the subledger, look for timing items, search for IC differences, document the findings. It takes 30–90 minutes per account. Multiply by 200 accounts and you have a team buried in investigation work that follows the same pattern every single time.
The 7 data tools
Every tool the agent uses to investigate your accounts
get_balance_details
Retrieves the current GL balance, subledger balance, and variance for the account being investigated. This is always the agent’s first tool call. It establishes the scope of the investigation.
get_transaction_history
Pulls individual GL transactions for the account within a date range. The agent uses this to identify specific journal entries causing the variance. By amount, date, reference, and description.
compare_periods
Compares balances and transaction patterns across 2–4 prior periods. The agent uses this to detect recurring variances, seasonal patterns, and timing differences that resolve month-over-month.
get_reconciling_items
Retrieves existing reconciling items (timing, permanent, IC) for the current and prior periods. The agent checks whether known items explain the variance before investigating further.
get_account_metadata
Returns account configuration: materiality threshold, reconciliation method, assigned preparer, historical confidence scores, and any Cortex rules. Context the agent needs to interpret findings correctly.
create_reconciling_item
Creates a new reconciling item as a draft. The agent uses this when it identifies a timing difference or adjustment that explains part of the variance. All AI-created items are flagged for human review.
How it works
The agent reasons iteratively | not in a single pass
The investigation agent doesn’t generate a response from a prompt. It enters a tool-calling loop: call a tool, analyze the result, decide what to check next. A typical investigation involves 4–8 tool calls. The agent might start with balance details, notice a $12,400 variance, pull transaction history, find three journal entries totaling $11,800, compare to last month, discover the same pattern resolved itself, and classify the finding as TIMING. Each step builds on the previous one.
Structured findings
Every finding has a type, amount, description, and recommendation
The agent produces structured findings, not paragraphs. Each finding is classified as TIMING, ERROR, PERMANENT, IC, or INFO. It includes the specific amount, a description referencing actual transactions from your database, and a recommendation (e.g., ‘No action needed, timing difference will clear next period’ or ‘Requires journal entry to correct misposting’). Findings are machine-readable and feed directly into work papers and confidence scoring.
Full transparency
Your auditors can trace every conclusion to source data
Every investigation produces a complete tool call transcript. Your auditors can see exactly which tools were called, in what order, what data was returned, and how the agent reasoned about each step. No black boxes. If the agent concludes a $45,000 variance is a timing difference, you can see the specific GL transactions it found, the prior period comparison it ran, and the reconciling item it created. The evidence chain is unbroken. Read about AI investigation. Part of Cortex. See how it powers Account Reconciliation.
