Skip to main content
    Use-case library
    Record-to-Report

    Account Reconciliation Triage

    Auto-match the clean reconciliations and route only the genuine breaks to humans.

    LLM workflowAutonomy: Act on exceptionSpeedCash / Working CapitalMedium effortOrg maturity: AI-EnabledCFO
    What changed

    Rules clear the exact matches and a trained model scores the rest, so high-confidence reconciliations clear straight through. A language model drafts the note for each remaining break, and preparers spend their time on exceptions instead of ticking and tying lines that already agree.

    How it's built

    LLM workflow (rung 4 of 6). A coded control path: chain, router, parallel fan-out, evaluator-optimizer.

    Not an agent: the sequence of steps is fixed in advance, so the whole path can be tested end to end.

    A language model handles step 4. Rules: step 1 · Predictive ML: steps 2 and 3 · Human: step 5.

    Autonomy: T2 · Act on exception
    Acts unattended on the confident majority; routes the rest to a human. Human role: reviews only the routed cases.
    How it is governed
    The path is testable end to end. Each step is independently evaluable.
    What the certification exams call this
    • Anthropic (Claude Certified Architect – Foundations (CCAR-F))Workflow: models and tools on predefined code paths (chaining, routing, parallelization, evaluator-optimizer). source
    • OpenAI (OpenAI agent guidance (Agents SDK, Codex, OpenAI Academy))Orchestrating via code, which the Agents SDK calls more deterministic and predictable. source
    • Google Cloud (Generative AI Leader)Workflow agents (sequential, parallel, loop) that run steps in a fixed order. source
    • Microsoft (AI-901 (Azure AI Fundamentals) and AB-100 (Agentic AI Business Solutions Architect))Sequential and concurrent orchestration patterns, with steps in a set order. source

    How Koko classifies AI · the full ladder is taught in AI & Agentic Delivery Foundations.

    Try it

    Illustrative demo

    Worked example of "Account Reconciliation Triage" on synthetic data — edit the inputs to tailor it. Edit the inputs to tailor the output — it's baked synthetic data, so nothing leaves your browser until you run it live.

    Synthetic demo company (a ~$5B high-growth US AI SaaS) — swap in any name to tailor the output.

    What this run should concentrate on.

    Run live in Claude

    Workflow

    1. 1.RulesIngest sub-ledger and bank/GL balances for the accounts in scope.
    2. 2.Predictive MLClear exact matches by rule, score the remaining candidate pairs, and auto-clear those above the approved confidence threshold.
    3. 3.Predictive MLCluster the unmatched items into break types (timing, missing, duplicate, FX).
    4. 4.Generative AIDraft a proposed disposition and supporting note for each break.
    5. 5.HumanQueue exceptions for preparer action; lock auto-cleared items for reviewer sampling.

    Prompt recipe

    Match the attached sub-ledger lines to the GL/bank lines. Auto-clear pairs with match confidence ≥ [95%]. For everything unmatched, classify the break (timing / missing / duplicate / FX / other), propose a disposition, and explain your reasoning. Never auto-clear below the confidence threshold.
    Try in Ask KokoAI
    Controls required
    • Auto-match confidence threshold approved and monitored
    • Reviewer samples a percentage of auto-cleared items
    • Segregation of duties preserved (preparer ≠ reviewer)
    • Full audit trail of match decisions retained
    HomeNewsLearnIntelBuildAsk Koko