We Ran a Full Month-End Close with an AI Agent. Here is What Happened.
We talk a lot about AI closing books faster. Time to show the receipts.
We ran Heshi's AI agent through a complete month-end close for a real client entity: a DeFi protocol with 3 legal entities (Singapore OpCo, Cayman Foundation, BVI IP holding), 20 wallets across Ethereum, Arbitrum, and Solana, with active positions on Aave, Curve, and Uniswap V3.
Here's exactly what happened.
The Setup
Entity complexity: 3 legal entities, 2 currencies (USD, SGD), intercompany management fee arrangement.
On-chain activity: 847 transactions across all wallets for the month. Mix of: token transfers, DEX swaps, Aave deposits/withdrawals, Curve LP operations, staking rewards, bridge transactions (ETH ↔ Arbitrum), and gas fees.
DeFi exposure: $2.3M in Aave lending (USDC + ETH), $800K in Curve 3pool LP, $450K in Uniswap V3 ETH/USDC concentrated liquidity, $1.2M in Lido stETH staking.
Accounting software: Xero, connected via OAuth.
Previous close time (manual): 14 business days with 2 accountants.
What the AI Agent Did
Transaction Classification: 4 minutes
The agent classified all 847 transactions:
- 412 correctly classified with high confidence (>95%): transfers, swaps, gas fees
- 389 classified with medium confidence (80-95%): DeFi interactions, bridge transactions
- 46 flagged for human review (<80% confidence): complex multi-step DeFi operations, unusual contract interactions
Accuracy on the 801 auto-classified: 98.5% (12 misclassifications found during review, all correctable).
The 46 flagged transactions took our human reviewer 2.5 hours to classify. Without the AI, classifying all 847 would have taken approximately 12 hours.
Wallet Reconciliation: 7 minutes
The agent reconciled all 20 wallets across 3 chains against GL balances:
- 17 wallets matched within tolerance
- 3 wallets had discrepancies:
- Wallet #8: 0.047 ETH discrepancy (unrecorded gas fees from 2 transactions near period-end)
- Wallet #14: 12.3 USDC discrepancy (Aave interest accrual timing difference)
- Wallet #19: Missing LP token balance in GL (Curve LP deposit recorded as transfer, not LP acquisition)
All 3 were genuine errors that would have been caught in manual review too — but the AI found them in 7 minutes instead of 2 days.
DeFi Accrual Calculations: 3 minutes
- Aave USDC lending: $4,167 interest income (calculated from aToken balance change)
- Aave ETH lending: 0.82 ETH interest income ($2,870 at period-end rate)
- Curve 3pool: $1,240 in trading fee share + $890 in CRV rewards
- Uniswap V3: $3,100 in fee accruals (position was in-range 78% of the month)
- Lido stETH: 0.48 ETH staking rewards ($1,680)
Total DeFi income for the month: $13,947
Manual calculation of these accruals previously took 6-8 hours. The agent did it in 3 minutes by querying protocol subgraphs directly.
Journal Entry Generation: 2 minutes
The agent generated 127 journal entries:
- 89 from classified transactions
- 12 DeFi accrual entries
- 8 intercompany entries
- 6 FX revaluation entries
- 5 gas fee expense entries
- 4 cost basis lot adjustments
- 3 reconciliation adjustments
All entries were created as drafts in Xero, pending human approval.
Human Review: 4 hours
Our CPA reviewed:
- The 46 flagged transactions (2.5 hours)
- The 12 misclassified transactions (30 minutes to correct)
- All 127 journal entries (45 minutes — spot-check, not line-by-line)
- The 3 reconciliation discrepancies (15 minutes)
After review, 3 JEs were modified, 2 were deleted (duplicates from a bridge that was counted twice), and 122 were approved as-generated.
Posting and Close: 5 minutes
Approved entries posted to Xero via API. Trial balance generated. Flux analysis prepared. Financial statements compiled.
The Results
| Metric | Manual (Previous) | AI + Human (Heshi) | Improvement |
|---|---|---|---|
| Total close time | 14 business days | 1.5 business days | 9.3x faster |
| Human hours | ~80 hours (2 people × 40h) | ~4.5 hours | 94% reduction |
| Transactions classified | 847 (all manual) | 801 auto + 46 human | 94.6% automated |
| Classification accuracy | ~96% (human error rate) | 98.5% (AI) | Higher accuracy |
| Reconciliation time | 2 days | 7 minutes | ~400x faster |
| DeFi accruals | 6-8 hours | 3 minutes | ~150x faster |
| Cost | ~$12,000 (blended labor) | ~$800 (4.5h CPA time) | 93% cheaper |
What Surprised Us
1. The AI was better at gas fees than humans. It caught every single gas fee across all 847 transactions and classified them correctly. Human accountants typically miss 5-10% of gas fees because they're small and easy to overlook.
2. Bridge transactions are still hard. The agent struggled most with multi-step bridge operations where tokens go through an intermediary contract. This is an area where we're improving the classification model.
3. The human review was mostly rubber-stamping. Of 127 journal entries, only 5 needed changes. The CPA's time was better spent on the 46 flagged edge cases — exactly where human judgment adds value.
4. Continuous reconciliation changed everything. Because wallets were monitored throughout the month, the period-end reconciliation was a verification exercise, not a discovery exercise. The 3 discrepancies were already known.
What We'd Do Differently
- Pre-classify bridge transactions by adding more known bridge contract addresses
- Implement parallel close for multi-entity groups (currently sequential)
- Add automated flux analysis commentary (currently generates numbers, human writes the narrative)
The Takeaway
AI doesn't replace the accountant. It replaces the 95% of the work that doesn't need an accountant. The CPA spent 4.5 hours making decisions — not downloading data, not copying numbers between spreadsheets, not reconciling wallets by hand.
That's what AI-native accounting means.
Want to see these results for your entity? Book a demo — we'll run a trial close on your actual data.