Reconciliations & Consistency Checker

What to do

1. 1Set the objectiveDescribe the purpose, intended audience, and scope of the reconciliation, including the documents and figures to be checked.
2. 2Map current processMap how figures are collected, differences are investigated, and conclusions are reviewed and approved. Identify who handles what and how the steps are connected.
3. 3Define the scopeThe agentic workflow extracts and compares figures, flags differences, and drafts reconciliation findings. Resolving exceptions and sign-off remain with finance reviewers. Publishing the reporting pack is outside the scope.
4. 4Set acceptance criteriaDefine what a good result looks like at each step. Require accurate data extraction, traceable evidence, and raise warnings when discrepancies occur.
5. 5Choose a test caseUse a reporting document and its source-of-truth data that a finance reviewer can check against expected matches, differences, and missing figures.

Example

An analyst starts a Q1 2026 check for an annual-report PDF and a management-pack PowerPoint against the financial database. The checker takes over the manual extraction, comparison, and write-up, but does not sign off or publish. Both routes are measured on correctness, evidence, review effort, and time.

What to do

1. 1Define sources and ownersName the owner of the source-of-truth data, reporting documents, and KPI definitions.
2. 2Prepare the dataDefine consistent KPI labels and units, including accepted aliases and sign conventions. Set tolerances for differences between reported figures and source data.
3. 3Assign reference IDsGive each source-data row and reported figure a stable ID and reporting period.
4. 4Add input checksCheck that all relevant KPIs and reporting periods are covered consistently across the input files, including source data and KPI definitions.

Example

The source-of-truth income statement, the annual-report PDF, the management-pack PowerPoint, and the KPI definitions are the identified inputs in the reconciliation process. The KPI definitions define the relevant metrics, each with a canonical ID, accepted aliases, a unit, and a sign convention.

What to do

1. 1Break down the taskWork backwards from the desired output and identify the processing steps required to get there. Group activities that together produce one meaningful, intermediate result. For example, extracted figures with source locations provide the basis for comparison against the source-of-truth data.
2. 2Specify each stepDescribe what each step receives, what it should do, and what it passes on. For instance, specify how extracted figures are matched to source data and how comparison results become reconciliation findings.
3. 3Define exception handlingSpecify how to handle missing or uncertain figures and comparisons that exceed the agreed tolerance. Define when to leave a comparison unresolved and request human review.
4. 4Define the outputDefine the output schema, including field names, data types, and required values. Specify how missing information is represented and provide a populated example. For each reconciliation finding, include fields for the compared figures, calculated difference, and source references.

Example

The workflow separates extracting figures from comparing values and drafting findings. This lets you check extraction accuracy and the calculated differences independently. The drafting step then brings those results together in reconciliation findings with source references.

The final output could follow a schema like this:

{
  "status": "exception",
  "canonical_id": "profit_before_tax",
  "period": "2026 Q1",
  "document": "PDF",
  "source_value": 107.9,
  "document_value": 111.9,
  "difference_usdm": 4.0,
  "unit_source": "USDm",
  "unit_document": "USDm",
  "source_location": "CSV row Profit Before Tax",
  "document_location": "page 2 management sign-off summary",
  "comparison_rule": "exact_id+period;abs_diff;tol=0.1"
}

What to do

1. 1Assign each stepDecide whether code, an AI agent, or a person should carry out each step. Use code for comparisons, agents to interpret extracted content and draft findings, and a finance reviewer for sign-off.
2. 2Implement each stepWrite the code and agent instructions required to process the step. You can use AI to assist you with this work. For the comparison step, implement the KPI matching rules and difference calculations from the blueprint.
3. 3Connect the stepsPass each step's output to the steps that depend on it. Implement the exception handling and requests for human input. Carry the stable IDs throughout each step all the way into the reconciliation findings.
4. 4Run the workflowProcess the sample inputs from start to finish. Inspect intermediate results and fix failures until the workflow produces an output that follows the schema.

Example

Implement comparison rules in code so the drafting agent receives calculated differences and assigned statuses. Carry the extracted figures and source locations into each finding. Run the sample reporting pack through the connected steps to produce a first reconciliation report.

What to do

1. 1Identify possible failuresWalk through the existing workflow steps and identify errors that could compromise the final output. For reconciliation, consider extraction errors, mismatches in units, and incomplete KPI definitions.
2. 2Implement validationsImplement checks for those errors. The checks can be performed by code or agents. For instance, verify totals and recalculate differences in code.
3. 3Raise issues for reviewFlag failed checks alongside the results and summarize them for the reviewer. Explain what failed, include the relevant evidence, and state what needs checking or correcting. For example, flag a comparison between different reporting periods and show both periods with their source documents.

Example

A figure reported in thousands is treated as millions. A code check flags that the normalized value does not follow the recorded conversion rule. The reviewer receives the original figure, its unit, and the converted amount.

What to do

1. 1Collect representative casesGather realistic inputs covering common tasks, difficult cases, and known failures. Include cases where the workflow should identify missing information or raise an issue for review. Include matching figures, genuine discrepancies, and extraction or unit-conversion errors.
2. 2Define expected resultsSpecify the expected results for each case before running the evaluation.
3. 3Set up gradingDefine how each result will be assessed against those expectations. Use code for exact comparisons and an agent or person with clear scoring criteria for judgments about content quality.
4. 4Run and compareRun the evaluation set and inspect failures to identify which steps need improvement. Record a baseline and repeat the evaluation after changes to compare performance.

Example

An evaluation case contains the same amount expressed in thousands in one document and millions in another. The expected result is a match after unit conversion.

What to do

1. 1Assign responsibilitiesDefine who uses the workflow, who reviews and approves results, and who maintains it.
2. 2Design the user interfaceGive users a clear way to provide inputs and inspect results.
3. 3Define the review processSpecify when review is required, what the reviewer must check, and how corrections and approvals are recorded.
4. 4Create audit logRecord the inputs, workflow version, outputs, and review decisions so each result can be traced back to how it was produced and approved.
5. 5Prepare for ongoing operationsDefine how the workflow will be monitored, maintained, and improved over time.

Example

A reviewer opens the reconciliation report and checks the flagged differences against their sources. They confirm the findings, make corrections, or request further evidence through the same interface. The audit log records their actions and final decision.