How to Build an AI Financial Close Assistant (Playbook)

The month-end close compresses reconciliations, variance explanations, evidence collection, and sign-offs into a few pressured days, much of it repetitive and all of it controlled. An AI financial close assistant removes the repetitive work while leaving every journal, approval, and judgment with accountants. This playbook covers the build, following FISTA's AI agents practice and the controls in the AI controls for financial services whitepaper. Context is in ai for finance teams.

What does the assistant do?

Capability	Function	Accountant action
Close orchestration	Task list, dependencies, status, reminders	Owns tasks
Reconciliation support	Matching, unmatched grouping, proposed explanations with evidence	Decides; clears
Journal preparation	Proposed entries with support	Reviews; posts
Variance analysis	Driver decomposition, drafted narrative with citations	Verifies; finalizes
Evidence gathering	Collects and organizes support per task	Reviews
Status and questions	Answers from close data	Reads

Step 1: Model the close with the controller

Document the close as tasks with owners, dependencies, deadlines, systems, and evidence requirements; define which tasks the assistant supports and how; define prohibited actions (no posting, no clearing, no sign-off); and set quality thresholds. The controller owns the specification. See how to write an ai spec.

Step 2: Build orchestration

Implement the task model with dependencies, status tracking, reminders, and dashboards; integrate with the ERP and close tools for task status where available. Orchestration is deterministic and makes bottlenecks visible. See what is an agentic workflow.

Step 3: Build reconciliation support

For each reconciliation type, match transactions across systems using rules and, where patterns are complex, models; group unmatched items; propose explanations such as timing differences, fees, or duplicates with the evidence; and present them to the preparer for decision. Matching performance is measured per reconciliation type. Patterns overlap with how to build an invoice processing agent.

Step 4: Prepare journal proposals

For recurring and reconciliation-driven entries, prepare journal proposals with accounts, amounts, descriptions, and supporting evidence for accountant review and posting under segregation of duties. Journal posting remains a human action through the ERP's controls. Integration design is in how to build an erp ai integration.

Step 5: Draft variance analysis

Compare actuals to prior periods and budget; decompose variances by driver using ledger detail and operational data (volumes, prices, headcount); retrieve known events and prior explanations; and draft narratives with citations to the data used. Analysts verify and finalize. Grounding controls are in how to ground an llm.

Step 6: Gather evidence

Collect statements, confirmations, reports, and approvals per task from connected systems and shared locations, organize them against evidence requirements, and flag gaps. Evidence organization is a significant time saver and an audit asset. See how to build an ai audit trail.

Step 7: Enforce controls

Segregation of duties in tool scopes; no posting, clearing, or sign-off capability; read access scoped by role; audit logging of every proposal and its disposition; and change control on rules and models with re-evaluation. Governance is in the agentic AI governance whitepaper.

Step 8: Evaluate on prior closes

Replay prior closes: measure matching precision and recall per reconciliation type, accuracy of proposed explanations against what preparers found, journal proposal correctness, variance narrative quality rated by the controller's team, evidence completeness, and control compliance. Pilot on one entity or a subset of reconciliations and measure cycle time. Method is in the AI evaluation and testing whitepaper.

Worked example: a multi-entity company

A company closing several entities pilots the assistant on bank and intercompany reconciliations and operating expense variance analysis for one entity. Orchestration surfaces that intercompany confirmations are the recurring bottleneck. Bank matching handles most transactions by rule; the model groups the remainder and proposes timing and fee explanations with statement evidence, which preparers accept or correct. Intercompany differences are matched across entities with proposed explanations. Journal proposals for recurring accruals arrive with support for the accountant to post. Variance narratives for operating expenses cite ledger detail and headcount data, and analysts edit rather than write from scratch. Replay of prior closes confirms matching accuracy; the pilot shortens the entity's close and the assistant expands to other entities and reconciliation types with the same controls.

What does it cost to run?

Cost is modest relative to accountant time; build cost is driven by system integration and reconciliation rule development. Value is measured in close cycle time, preparer hours, and audit readiness. Drivers are in the AI total cost of ownership whitepaper.

What are the common mistakes?

• Giving the assistant posting or clearing capability.
• Matching without evidence, so preparers cannot trust proposals.
• Variance narratives that speculate beyond the data.
• Ignoring segregation of duties in tool scopes.
• Piloting on the hardest reconciliations first.
• No replay of prior closes before live use.

How do you phase the assistant?

Begin with close orchestration alone, which needs no model risk and immediately shows where time goes. Add reconciliation support for the highest-volume, lowest-judgment reconciliations such as bank and clearing accounts, replayed against prior closes before live use. Add journal proposals for recurring entries once preparers trust the matching, and add variance narratives last, because they require the most context and the most careful review. Each stage is measured on preparer hours and cycle time for one entity before extending to others.

Who owns the assistant?

The controller owns the specification, controls, and quality thresholds; reconciliation and reporting leads own their task definitions and evaluation cases; engineering owns integration, tools, and the evaluation harness; and internal audit reviews control design and the audit trail.

How do you satisfy auditors?

Keep every matching decision, variance explanation, and piece of evidence attached to the journal or reconciliation it supports, show which suggestions were accepted, edited, or rejected and by whom, and preserve the ability to reproduce any explanation from the data as of the close. Auditors accept assistance that leaves a better trail than the manual process did.

This guide is general guidance, not legal or accounting advice; confirm control and audit requirements with your auditors before deployment.

How FISTA Solutions builds financial close assistants

FISTA Solutions builds financial close assistants to this playbook: controller-owned specifications, deterministic orchestration, evidence-backed reconciliation support, journal proposals under segregation of duties, cited variance narratives, evidence gathering, enforced controls, and evaluation by replaying prior closes. The AI agents practice delivers the assistant, AI enablement the integration and retrieval platform, and forward deployed engineers embed with your accounting team. The record behind the work is 150+ projects with 99.9% uptime.

To scope a financial close assistant, message FISTA on WhatsApp, or read ai financial forecasting for the planning counterpart.