Month-end close reconciliation AI agent: a blueprint for the financial close
How a governed AI agent workforce prepares month-end close reconciliations: it ties ledger balances to subledgers and bank feeds, links every match to evidence, and leaves certification with an accountant, so controllers certify rather than assemble.
- Problem
- Close week is spent tying out low-risk accounts by hand, and judgment items get whatever time is left.
- Agent design
- Matching, reconciliation, close-orchestration and pattern agents under controllership oversight.
- Systems
- General ledger, AP, AR and fixed asset subledgers, bank statement feeds, and the close management tool.
- Hard limits
- Agents never certify a reconciliation, never post journal entries, and never prepare reserves or estimates.
- Value levers
- Preparer hours released, shorter close, less overtime, fewer aged items written off.
- First release
- One parallel close on the low-risk, high-volume tier, then agent preparation with certification unchanged.
Where month-end reconciliation breaks today
Every balance sheet account needs a reconciliation each period, or each quarter for the lowest-risk ones. A preparer ties the GL balance to the subledger, bank statement or schedule, lists the reconciling items, attaches evidence and submits the workpaper for certification. A multi-entity group has thousands of these, and most look the same every month: a cash account with a few outstanding checks, an AP control account that ties to aged payables, a prepaid roll-forward with one new invoice. They take the same effort anyway, because nobody knows nothing changed until the work is done.
The pressure lands in the same places every close. Clearing accounts carry unmatched items that age past explanation. Intercompany balances disagree by an FX difference nobody owns. Evidence lives in five systems, so workpapers get rebuilt in spreadsheets, and reserves get reviewed last. The first design decision is a risk tiering of the accounts, because the tier decides what agents may prepare.
| Account or item type | Typical reconciling item | What the agents do | Agents may prepare alone? |
|---|---|---|---|
| Bank and cash accounts | Outstanding checks, deposits in transit, bank fees, lockbox timing | Matches statement lines to GL cash postings and clears policy-defined timing items | Policy timing items |
| Subledger control accounts | Unposted batches, cutoff timing, manual GL postings to control accounts | Ties GL to the subledger report and lists every manual posting that breaks the tie | Exact tie-out only |
| Prepaids and simple accruals | New invoices, amortization not yet booked, schedule drift | Rolls the schedule forward and attaches source invoices | Roll-forward only |
| Clearing and suspense accounts | Unmatched payments, misapplied cash, payroll and card clearing residue | Pairs offsetting entries, ages the rest, routes items past the aging threshold | Matches only, aged items routed |
| Intercompany balances | One side posted late, different FX rates, disputed charges | Pairs counterparty balances, quantifies the difference, names the entity that must act | Paired balances only |
| Reserves and estimates | Bad debt, obsolescence, warranty, legal and tax provisions | Supplies aging, usage history and claims data to the accountant; prepares nothing | Never |
| Unexplained variance or top-side entry | A difference with no matching item, or a manual consolidation adjustment | Stops, records what was tried, routes to the owner | Never |
The month-end close reconciliation agent workforce: four roles that prepare and never certify
The live design has two agents: one prepares accounts, one runs the calendar. This blueprint splits preparation into matching and workpaper assembly and adds a pattern agent, so the identity that clears items never writes the workpaper. Each agent has its own identity with preparer access only and no reviewer or approver role anywhere.
| Agent | Job | Reads | Can write | Never does |
|---|---|---|---|---|
| Matching agent | Matches bank, clearing and intercompany items; clears rule-defined timing items | Bank statement lines, GL line items, intercompany postings, auto-clear rules | Match sets with evidence links, each cleared item tagged with its rule | Clear an item no rule covers, or change the rules |
| Reconciliation agent | Prepares each account's workpaper with every item tied or flagged | Trial balance, subledger reports, match sets, prior workpaper | Workpaper and preparer status; draft adjusting entries | Certify a reconciliation, post or approve a journal entry, prepare a reserve |
| Close-orchestration agent | Runs the close calendar: task dependencies, preparer reminders and bottleneck alerts | Close task list, subledger lock status, reconciliation status by account | Task status, reminders, escalations to the close manager | Move a deadline or complete a task for a person |
| Pattern agent | Reports recurring items, aging trends and accounts whose tier looks wrong | Audit log, certified workpapers, review rework notes | Post-close pattern report and proposed tier or rule changes | Act on any open account in the current close |
Reconciliation agent architecture: beside the ledger and the close tool
The agents read the ledger, subledgers and bank feeds through governed interfaces and write in one place: the close tool or workpaper repository the team already uses. The only path toward the ledger is a draft entry in the journal approval workflow. A control plane outside the agents holds the policy and tiers, the certification workflow, the audit log and a pause switch that returns accounts to manual preparation with the calendar intact.
How one account moves through the close reconciliation design
Two questions route every account at cutoff. Is it in a tier or category agents may never prepare? If not, did every reconciling item tie out or clear under a written rule? Only a "no" then a "yes" gives an agent-prepared workpaper, and a reviewer still certifies it under their own name.
Ledger, bank and close tool integration for reconciliation agents
The pattern holds on every platform: read balances and line items through APIs or governed extracts, match where volume is high, write workpapers where reviewers already work. What changes is where matching already happens natively.
| Platform | What agents read | How agents read | Where results go | Design note |
|---|---|---|---|---|
| SAP S/4HANA | Universal Journal balances, clearing account open items | OData APIs and released CDS views | Close tool workpapers; draft entries in the journal approval process | Clearing an SAP open-item account is itself a posting, so routine clearing stays in SAP automatic clearing and agents propose rule changes |
| Oracle Fusion Cloud | GL balances, subledger detail, bank statement lines | REST resources and extracts | Account Reconciliation where licensed, else the close tool | Where Account Reconciliation and transaction matching are licensed, agents work in its formats and roles |
| Dynamics 365 Finance | Trial balance, subledger balances, bank statements | Data entities and the data lake export | Close tool workpapers; unposted draft journals | Advanced bank reconciliation has its own matching rules; agent rules must mirror them |
| Close management tool | Account list, risk ratings, prior workpapers | Vendor APIs or bulk import | Workpaper, attachments, preparer status | A distinct preparer user per agent, with reviewer roles kept for people, lets the tool enforce certification |
| Bank statement feeds | Statements in BAI2, camt.053 or MT940 | Treasury system or bank files, read only | Statement lines to the matching agent | Check completeness first: opening balance equals prior closing balance, no gaps in statement sequence |
Agents can run on ERP-native agent tooling, a data platform such as Snowflake or Databricks, or Microsoft Copilot Studio. Where BlackLine, FloQast or Trintech Cadency is in place, it stays the system of record for certification.
Human oversight of account reconciliation and certification
Reconciliation oversight is about who stands behind a balance. Agents do the matching and assembly any preparer would do. Every decision that changes the ledger, explains a judgment or signs off a balance stays with a named person in controllership.
| Decision | Who decides | Agent's role | Control |
|---|---|---|---|
| Matching statement and GL lines, clearing policy timing items | Agent | Matches, clears under the named rule, links evidence | Each rule has an owner; reviewers sample rule-cleared items and a reversal suspends the rule |
| Preparing a low-risk tier workpaper | Agent prepares | Ties the balance, lists items, attaches support | A reviewer still certifies every workpaper; rework is tracked by account |
| Close calendar reminders and escalations | Agent | Tracks dependencies, flags bottlenecks | Close manager owns the calendar and every deadline change |
| Unmatched or aged reconciling items | Account owner | Groups and ages items, suggests causes | No clearing without a rule; controllership sets aging thresholds |
| Adjusting journal entries | Preparer and JE approver | Drafts the entry with support attached | Stays a draft in the journal workflow; agent identities cannot post or approve |
| Reserves, estimates and high-risk accounts | Accountant | Supplies data only | Excluded from agent preparation by tier, at any rule setting |
| Reconciliation certification | Reviewer | Prepares the package | Certification records a human identity only; agents hold no reviewer role |
| Risk tiers and auto-clear rule changes | Controller | Proposes changes from the pattern report | Versioned policy; every workpaper records the policy version it used |
Illustrative economics of close reconciliation agents
The model uses example inputs for a group reconciling 4,000 balance sheet accounts a month; change them to your own. Only the low-risk tier is modeled. Released preparer hours become cash only if redeployed or used to replace contractors, while avoided overtime and avoided write-offs of aged items are cash either way.
Fixed assumptions: certification time sits in the extra review hours; shorter close days carry no dollar value; build cost excludes internal team time. Example inputs only. They describe no client and are not a forecast.
At these inputs the cash-only view does not pay back, the usual shape of a reconciliation case: the value is accountant time. The plan must say where released hours go before the build starts: fewer close contractors, absorbing an acquisition without new headcount, more time on reserves and flux analysis, or a day out of the close. If none is real, start with bank and intercompany accounts, where aged items carry cash.
Rollout roadmap for reconciliation agents, one risk tier at a time
Month-ends set the calendar, because every gate needs a real close to test against. The controller signs each gate. Preparation moves to agents one tier at a time, certification never moves, and reclaimed hours go to judgment items. Durations are typical for one entity cluster.
- Weeks 0 to 4
Baseline
Inventory the accounts, confirm risk tiers with controllership, measure preparation hours by tier, and write auto-clear rules from what reviewers already accept. Brief the external auditors.Gate: controller signs risk tiers and auto-clear policy v1 - Weeks 4 to 10
Parallel close
Run at least one full month-end with agents preparing the low-risk tier alongside the usual preparers, then compare balances, items and evidence account by account.Gate: agent and human workpapers agree on 98% of low-risk accounts, every gap explained - Weeks 10 to 18
Agent preparation
Agents prepare the low-risk tier for two live closes while reviewers certify every account. Test the pause switch mid-close.Gate: review rework under 2% and no rule-cleared item reversed - Week 18 onward
Scale
Add clearing and intercompany accounts, then the medium tier, then further entities and acquisitions, each with its own tiering and parallel close.Gate: per tier, per entity
How month-end close performance is measured
Every measure is baselined from your own close data in stage one. This page names measures and claims no values.
| Metric | Definition | Why it matters |
|---|---|---|
| Close duration | Business days from period end to ledger lock, by entity | The headline measure; read with on-time completion |
| Preparation hours per account | Preparer hours by risk tier, from close tool time or sampling | Shows whether the low-risk tier stopped consuming accountant time |
| Review rework rate | Share of agent-prepared reconciliations a reviewer sends back | The control metric that decides whether the next tier moves |
| Rule-cleared reversal rate | Share of items cleared under an auto-clear rule later reopened | Tests each rule; one reversal suspends it |
| Aged open items | Count and value of reconciling items older than 60 and 90 days | Aged items turn into write-offs and audit findings |
| On-time task completion | Close tasks completed by their calendar deadline | Measures the orchestration agent and locates bottlenecks |
| Judgment hours | Hours spent on reserves, accruals and flux review | Shows whether reclaimed hours reached the work that needs judgment |
Questions controllers ask about close reconciliation agents
Can the agent certify a reconciliation?
No. Agents hold preparer rights only. Certification records a human identity, and controllership keeps the tiering decision.
How will our external auditors view agent-prepared reconciliations?
The design gives them what they test: an auto-clear policy with owners, evidence on every match, a human reviewer on every account, and access controls over agent identities. Bring them in during baseline so agent identities are scoped into IT general controls before the first live close.
We already use BlackLine or FloQast. Does this replace it?
No. The close tool stays the system of record for workpapers and certification; agents work inside it as preparers.
Do the agents post journal entries?
No. The agent drafts any adjustment with support in your journal workflow, and your existing approvers decide.
How long until agents prepare accounts in a live close?
Typically about ten weeks for one entity cluster: four weeks of baseline, then at least one full parallel close. Each gate needs a real month-end, so your close calendar sets the pace.
Working with WGA on month-end close and reconciliation
The engagement starts with a readiness check on the account population: how accounts are tiered, where subledger and bank data lives, and how much preparer time each tier consumes. WGA baselines the close and agrees targets with your controller before anything ships, then writes the auto-clear rules with the reviewers who apply them today.
Senior WGA consultants do the work with fees fixed against the outcome and stay through the parallel close and first live releases. Your controllership and finance systems teams inherit the tiers, rule book, runbooks and pause procedure, so they own the workforce and can extend it to new entities without us.
Discuss month-end close in your operation
Talk to our AI practice