Bank Reconciliation · Start here
Who this fits
Close manager or controller reconciling a cash account across General Ledger and Treasury.
First pilot
One synthetic cash-account export pair with known discrepancies, a named reviewer and measured review time.
Required human review
Immediate review of every consequential draft: matching, exception dispositions, proposed adjustments and workpapers. The larger option also reviews feed completeness before matching.
Assumptions
Complete period exports, equal-amount matching policy, read-only agents and documented evidence. Unresolved items stay open. No agent posts an entry; estimates require pilot validation.
Bank Reconciliation
A reference design for reconciling one cash account across General Ledger and Treasury. Every agent prepares a draft that stops at an immediate human-review gate, including matching, proposed adjustments and workpapers. Unresolved items stay open; no agent has posting authority.
Partner brief · Draft for review
Bank Reconciliation · Balanced
5 agents · 5 immediate human-review gates · ~4 weeks
Estimated economics & assumptions
- Estimated hours reclaimed: 38 per month × 12 cycles/year = 456 hours/year.
- Assumed blended rate: $75/hour. Estimated value of reclaimed capacity: $34,200/year. Capacity value is not a reduction in spending.
- Estimated build investment: $40,000; ongoing run cost: $600/month.
- Estimated payback from capacity value after run cost: about 18 months.
- Hours reclaimed are illustrative, net of time at every immediate human-review gate; remeasure in the synthetic pilot.
- Build investment covers five agents, the orchestrator, connector work and back-testing against two closed months.
- Assumes a stable bank-feed connection across all reconciled accounts; one-time feed setup is excluded.
First pilot
Feed ingestion and matching with immediate review (~1.5 weeks): Bank Feed Ingestion Agent; Feed completeness review: required; Transaction Matching Agent; Matching review: required. Build the listed agents and their named human-review gates together. Prove completeness, evidence checks and downstream blocking on synthetic data before expan…
Decision requested
Decide whether to scope this pilot; assign a pilot owner and reviewers, and resolve the open questions before implementation.
Mandatory human-review gates — proposed
Any agent whose output touches a tax return, a financial statement, a client deliverable or a material number is immediately followed by a human-review gate.
- Feed completeness review: required — General Ledger: Bank Feed Ingestion Agent affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source eviden…
- Matching review: required — General Ledger: Transaction Matching Agent affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evide…
- Exception and evidence review: required — General Ledger: Bank Discrepancy Resolver affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source eviden…
- Adjustment review: required — General Ledger: Adjusting Entry Drafter affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence…
- Controller workpaper review: required — General Ledger: Reconciliation Workpaper Builder affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source…
Open questions / decisions
- Who owns each mandatory gate, and what materiality thresholds and sign-off evidence are required?
- Verify access and integration availability for NetSuite, BlackLine, Excel, Bank feeds.
- Confirm hours, rate, frequency, costs, pilot test data and success criteria before committing resources.
Draft for review. Advisory only; not professional, legal, tax, or accounting advice. A qualified professional must review the plan, prompts and controls before implementation. Proposed controls have not been audited or implemented.
Version bank-reconciliation · revision date unavailable · Prepared 2026-10-15. Long descriptions are abbreviated; full details follow.
Links include this revision, selected option and rate. Anyone with a generated plan's link can open it.
Full appendix exports · Draft for review
Detailed specifications for review with IT or a consultant before implementation.
- Hours reclaimed are illustrative, net of time at every immediate human-review gate; remeasure in the synthetic pilot.
- Build investment covers five agents, the orchestrator, connector work and back-testing against two closed months.
- Assumes a stable bank-feed connection across all reconciled accounts; one-time feed setup is excluded.
Compare the payback before you choose
The same economics for every option, at a $75/hr blended rate. The Lean option pays back fastest, but the right pick balances payback against how much of the process you want to automate. Select a row to load that option above.
| Option | Payback | Savings / yr | Build cost | Net, year one |
|---|---|---|---|---|
| LeanFastest payback | ~15 mo | $13k · 168 hrs | $11k + $300/mo | −$2,000 |
| Balanced | ~18 mo | $34k · 456 hrs | $40k + $600/mo | −$13k |
Start lean, expand as the controls prove out
These aren't just alternatives. Read left to right, they're a phased path. Most firms launch the lean option first, then widen scope once the review gates have a track record. Select a phase to see its architecture.
Build it in order. Start with Phase 1
You don't stand up every agent at once. Phase 1 is the lowest-risk, fastest-payback agent to build first. Prove the pattern there, then widen scope. (This sequences the agents inside this option; the adoption path above sequences the options themselves.)
Build the listed agents and their named human-review gates together. Prove completeness, evidence checks and downstream blocking on synthetic data before expanding the pilot.
- Bank Feed Ingestion Agentlow
- Feed completeness review: requiredGate
- Transaction Matching Agentlow
- Matching review: requiredGate
Build the listed agents and their named human-review gates together. Prove completeness, evidence checks and downstream blocking on synthetic data before expanding the pilot.
- Bank Discrepancy Resolvermedium
- Exception and evidence review: requiredGate
- Adjusting Entry Draftermedium
- Adjustment review: requiredGate
Build the listed agents and their named human-review gates together. Prove completeness, evidence checks and downstream blocking on synthetic data before expanding the pilot.
- Reconciliation Workpaper Builderlow
- Controller workpaper review: requiredGate
Every agent's failure mode, every checkpoint
The same content a reviewer would expect in an engagement memo. What each agent could get wrong, and the human sign-off that catches it.
| # | Agent | Team | Failure mode | Effort |
|---|---|---|---|---|
| 1 | Bank Feed Ingestion Agent | General Ledger | A stale or partial feed can leave the reconciliation silently incomplete. | low |
| 3 | Transaction Matching Agent | General Ledger | Can force a match on two similar transactions that are not the same item. | low |
| 5 | Bank Discrepancy Resolver | Treasury | May misclassify a genuine error as a normal timing difference. | medium |
| 7 | Adjusting Entry Drafter | General Ledger | Output adjusts a financial-statement balance. Never posted without review. | medium |
| 9 | Reconciliation Workpaper Builder | General Ledger | A presentation slip can hide an open reconciling item from the reviewer. | low |
| # | Control gate | Team | What it confirms | Why it's mandatory |
|---|---|---|---|---|
| 2 | Feed completeness review: required | General Ledger | Confirm completeness, accuracy, evidence, authority and unresolved items before downstream use. | Bank Feed Ingestion Agent affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence immediately, before another agent uses it. Unsupported items stay open; approval of a draft is not posting authority. |
| 4 | Matching review: required | General Ledger | Confirm completeness, accuracy, evidence, authority and unresolved items before downstream use. | Transaction Matching Agent affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence immediately, before another agent uses it. Unsupported items stay open; approval of a draft is not posting authority. |
| 6 | Exception and evidence review: required | General Ledger | Confirm completeness, accuracy, evidence, authority and unresolved items before downstream use. | Bank Discrepancy Resolver affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence immediately, before another agent uses it. Unsupported items stay open; approval of a draft is not posting authority. |
| 8 | Adjustment review: required | General Ledger | Confirm completeness, accuracy, evidence, authority and unresolved items before downstream use. | Adjusting Entry Drafter affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence immediately, before another agent uses it. Unsupported items stay open; approval of a draft is not posting authority. |
| 10 | Controller workpaper review: required | General Ledger | Confirm completeness, accuracy, evidence, authority and unresolved items before downstream use. | Reconciliation Workpaper Builder affects a material cash balance or its support. A named GL reviewer/controller must approve the draft against source evidence immediately, before another agent uses it. Unsupported items stay open; approval of a draft is not posting authority. |
Controllership layer
Every agent action routes through these controls before it touches the ledger, the same way every other control in your close already works. Agents aren't the risk. agents without controls are. Build the controls in, not bolt them on after.
Prebuilt Claude Finance Agents
4 of 5 agents in this option are plug-and-play. Most of the work is already an off-the-shelf Claude Finance Agent. The remaining 1 is where your firm's unique process lives.
"Build this agent" opens the Agent Builder pre-filled from this step's spec. The result lands in your Agent Library and can be pinned into future plans.
Deployment model
General guidanceThree ways to run these agents. All route through the controllership layer before touching your ERP / GL. The choice is where the agents live and who operates them.
- ●Claude for Excel, PowerPoint & Word add-ins (GA)
- ●Claude Cowork with Dispatch. Assign work by text or voice
- ●An analyst drives and reviews every run
- ●No connector build: works on the files you already have
- ●Scoped credentials
- ●Approval workflow
- ●Execution logs
- ●Versioned configs
- ●Internal orchestration (Bedrock / AgentCore)
- ●Policy engine
- ●Internal audit logs
- ●Runs inside your security boundary
What connects today, and what you'd build
Every agent has to plug into the software your firm already runs. 2 of 4 systems have an off-the-shelf connector today; the rest need a connection built. Scope this before you budget the build.
Agents read; people post. Nothing writes back to your system of record on its own — an agent drafts the entry or reconciliation, and a human approves it at a review gate before it posts to the ERP.
This is a reference design. The systems below are an example stack, not requirements. Use different software? Use Build and the same plan gets designed around what your firm actually runs.
| System | Connectivity | Today | How you'd connect it | Effort |
|---|---|---|---|---|
NetSuite System of record · ERP / general ledger | Connector ready | Buy | SuiteTalk REST API plus common iPaaS connectors. An agent reads the cash ledger and stages draft entries. Oracle NetSuite (cloud). SuiteTalk REST & SOAP plus SuiteQL — a custom connector is well-trodden. | low |
BlackLine | Custom connector | Build | Documented REST API: a custom connector pulls reconciliation templates and pushes match results and workpapers. | medium |
Excel | Connector ready | Buy | Claude for Excel (GA add-in) works in the reconciliation workbooks natively. Ties out support and flags breaks without a connector build. | low |
Bank feeds Bank / treasury | Custom connector | Build | Via an aggregator (e.g. Plaid) or direct BAI2/OFX files. A connector normalizes the feed for matching. | medium |
The first draft is never the keeper. Describe a change in plain language and we'll regenerate.