How AI investigation agents are replacing manual reconciliation

The average senior accountant spends four hours investigating a single complex reconciliation variance. They open the ERP, pull transaction histories, cross-reference posting dates, check intercompany confirmations, review prior period reconciliations, and eventually piece together an explanation. The investigation itself is systematic. The problem is that a human is doing it manually, one screen at a time.

AI investigation agents do the same work in minutes. Not by guessing. By following the same investigative logic your best accountant would use, but with the ability to query every relevant data source simultaneously.

The manual investigation problem

Consider a $47,000 variance on an intercompany receivable account. A senior accountant investigating this manually would:

1. Pull the GL detail for the current and prior period
2. Compare to the counterparty entity's payable balance
3. Identify unmatched transactions on both sides
4. Check posting dates for timing differences
5. Review cash application records for payments in transit
6. Look at the FX rates applied to each side
7. Check whether any manual journal entries were posted
8. Document the finding and recommend a resolution

Each step requires navigating to a different system or screen, waiting for data to load, and manually comparing what they find. The intellectual challenge is moderate, any experienced accountant can do it. The time cost is enormous because the work is sequential and manual.

Multiply this by 50 accounts with variances per close cycle, and your team is spending 200 hours per month on investigation alone. That is more than one full-time equivalent doing nothing but looking up data and writing explanations.

How tool-use AI agents work

An AI investigation agent is not a chatbot that summarizes data. It is an autonomous system with access to tools, the same data sources your accountants use,, and the ability to decide which tools to use, in what order, based on what it discovers.

The architecture has three layers:

The reasoning engine. A large language model that understands accounting concepts, reconciliation logic, and investigative methodology. It does not memorize your data. It reasons about it the same way an experienced accountant would: what is the variance, what are the most likely causes, and what data would confirm or rule out each cause.

The tool set. Seven specialized data tools that the agent calls during an investigation:

• GL Transaction Query: pulls detailed transaction history for any account/period/entity combination
• Sub-ledger Lookup: retrieves supporting detail from AP, AR, fixed assets, and other sub-ledgers
• Counterparty Balance Check: queries the offsetting entity's balance for intercompany reconciliation
• Cash Application Search: finds payments in transit, unapplied cash, and partial applications
• Prior Period Comparison: retrieves last period's reconciliation and identifies recurring items
• FX Rate Retrieval: pulls the exchange rates applied to each transaction for currency variance analysis
• Journal Entry Search: finds manual entries, reclassifications, and adjustments posted to the account

The output layer. Structured findings in a consistent format (cause identified, supporting evidence, recommended resolution, confidence level). That feed directly into the reconciliation work paper.

What an investigation looks like

Back to the $47,000 intercompany variance. The agent receives the assignment and begins:

Step 1: Scope the variance. The agent queries the GL balance ($347,000) and the counterparty balance ($300,000). Variance: $47,000. It immediately checks whether this variance existed last period. It did not. This is new.

Step 2: Identify candidates. The agent pulls all transactions posted to the account this period and all transactions posted to the counterparty's offsetting account. It runs matching logic and identifies three unmatched items on the receivable side totaling $47,000.

Step 3: Investigate each item. For each unmatched transaction, the agent checks:

• Was the corresponding entry posted on the payable side? (No two of three were not.)
• Is there a cash payment in transit that explains the difference? (Yes one payment of $22,000 was sent on the 29th and received on the 2nd of the following month.)
• Was the remaining $25,000 posted at different FX rates? (Yes the receivable used the March 28 rate, but the payable entity has not yet posted the entry.)

Step 4: Produce findings. The agent generates a structured finding:

• $22,000 Timing difference. Payment sent March 29, received April 2. Will clear next period. No adjustment needed.
• $25,000 Counterparty posting delay. Invoice IC-2026-0847 posted on receivable side March 15. Payable entity has not posted corresponding entry. Recommend follow-up with Entity B controller.

Total investigation time: 3 minutes. Total human time: 30 seconds to review and approve.

Structured findings, not summaries

The difference between an AI agent and a reporting tool is the quality of the output. A reporting tool shows you the data. An agent tells you what the data means.

Every investigation produces:

• Root cause classification: timing, posting error, FX variance, missing transaction, unauthorized entry, or other
• Supporting evidence: the specific transactions, dates, amounts, and rates that support the conclusion
• Confidence score: how certain the agent is in its finding, based on the completeness and consistency of the evidence
• Recommended action: no adjustment needed, post JE, follow up with counterparty, escalate to reviewer
• Audit trail: every tool call, every data point retrieved, every reasoning step, logged and reviewable

This is not a black box. Your team can see exactly how the agent reached its conclusion and override it when they disagree.

Work paper generation

Investigation findings flow directly into the reconciliation work paper. Each reconciling item gets a row with the variance amount, the root cause, the supporting reference, and the resolution status. The work paper is audit-ready the moment the investigation completes.

For accounts that reconcile cleanly, no variance, all transactions matched,. The agent certifies the reconciliation automatically with a confidence score above the threshold your team sets. No human touch required for accounts where there is genuinely nothing to investigate.

The result: your team reviews work papers instead of building them. Reviewers spend their time on judgment calls, not on checking that the preparer pulled the right data.

The economics

Manual investigation costs scale linearly. Every new account, every new entity, every increase in transaction volume adds hours to your close. An AI agent's marginal cost per investigation is $0.05 to $0.15, depending on the complexity and number of tool calls.

For a company reconciling 500 accounts per month with an average of 100 requiring investigation:

• Manual cost: 100 investigations × 2-4 hours × $75/hour loaded cost = $15,000-$30,000/month
• Agent cost: 100 investigations × $0.10 average = $10/month
• Human review: 100 investigations × 5 minutes × $75/hour = $625/month

The cost reduction is dramatic, but the real value is not cost savings. It is speed. Investigations that took days now take hours. Close cycles that ran 10 business days compress to 5. Your team goes home on time.

What agents cannot do (yet)

AI investigation agents are not a replacement for professional judgment. They excel at:

• Systematic investigation across multiple data sources
• Pattern recognition (this variance looks like the one on Account X last quarter)
• Consistent documentation and work paper generation
• High-volume processing without fatigue or error accumulation

They are not yet reliable for:

• Novel transactions with no historical pattern to reference
• Judgment-heavy estimates (litigation reserves, warranty accruals)
• Situations where the "right" answer depends on management intent or business context the agent cannot observe

Arvexi Cortex is designed to recognize its own limitations. When an investigation hits a dead end or the confidence score falls below the threshold, the agent escalates to a human investigator with everything it has found so far. The human starts at the 80% mark, not at zero.

From investigation to prevention

The most valuable output of AI investigation is not the individual finding. It is the pattern data that accumulates over hundreds of investigations. When the same root cause appears on the same account three months in a row, that is not a reconciliation problem. That is a process problem.

Arvexi's Cortex surfaces these patterns automatically: accounts with chronic timing differences, entities with recurring posting delays, vendor accounts with systematic short payments. Your team shifts from investigating variances to preventing them.

That is the real endgame. Not faster reconciliation, fewer reconciling items in the first place.