A management report can sound convincing even when its inputs disagree. One spreadsheet shows a total of completed work, another counts accepted requests and a third includes a later correction. Asking AI to explain the difference before establishing those scopes can turn a measurement mismatch into a fictional business story. The first step is a reconciliation record, not narrative generation.
The proposed process below calculates and checks reporting inputs before AI drafts commentary. It does not establish a universal accounting or management-reporting standard. The information owner defines the real metric, source authority and acceptable scope. The practical output is a reconciliation gate that leaves disagreement visible and gives a reviewer a precise next decision.
Fix what each figure is supposed to mean
Record the entity, reporting dates, timezone, metric definition and accepted record population. A total without these fields can be impossible to compare. Completed jobs and requests received are different metrics even if both columns are labelled volume. Do not assume a shared label establishes a shared definition.
Keep source extraction time separate from the period boundary. A report generated later may include corrections that an earlier snapshot did not contain. Decide whether the comparison uses one fixed cut-off or deliberately compares versions. The latter needs an explanation of the change, not a claim that the business improved overnight.
For monetary totals, keep currency and sign conventions explicit. Do not add currencies or reverse credits to make a report balance. For event counts, define whether repeated events, reopened records or cancelled items belong in the measure. Those rules belong to the information owner, not the drafting model.
Calculate with accepted records and preserve the inputs
Use deterministic formulas or application code for sums, counts and differences. Retain record identifiers and the calculation version. An accepted source total can be compared with the row-level calculation where coverage permits, but the comparison should show what was included and excluded.
Source: Google Sheets batch updates describes grouped spreadsheet changes and distinguishes spreadsheet details from cell-value operations. It is an integration capability reference, not a guarantee that a spreadsheet's business formulas or source coverage are correct. The team must implement and inspect the actual calculations and permitted write path.
Preserve the imported snapshot before corrections. If a duplicate source record is excluded, record its identity and the approved reason. Similar amounts or text do not establish duplicate identity. If an opening value or required input is missing, keep the calculation incomplete rather than replacing it with zero.
Compare scope before treating a difference as an error
Check population, definitions, date basis and accepted versions before subtracting totals. Some differences reflect incomparable inputs rather than a calculation defect. A report counting all requests will normally differ from one counting only completed work; neither total is necessarily wrong under its own definition.
Once scope is comparable, calculate the difference and identify affected records. Keep an unexplained remainder visible. Do not ask AI to supply a plausible balancing item. Even a zero difference can hide offsetting errors or missing records, so the reviewer also checks coverage and membership.
If the information owner permits a tolerance, record its purpose, unit and approval. A universal percentage threshold would be inappropriate. A tolerated rounding difference is different from an unexplained missing transaction. The workflow should not treat all small differences as harmless.
Reusable pre-narrative reconciliation gate
Use this proposed worksheet for every metric that will appear in the management report.
| Field | Required record | Gate decision |
|---|---|---|
| Metric identity | Definition ID/version, entity and unit/currency | Meaning is accepted and comparable |
| Period | Start/end, timezone, cut-off and extraction times | Same intended reporting basis |
| Source coverage | Expected records/systems and accepted inclusion rules | Missing coverage is explicit |
| Calculation | Formula/version and member record IDs | Deterministic result can be reproduced |
| Compared figures | Source totals and computed totals | Values remain separately attributable |
| Difference | Defined subtraction and any accepted tolerance | No invented balancing item |
| Exceptions | Missing, conflicting, duplicated or excluded records | Owner and next action assigned |
| Reviewer decision | Reconciled, limited scope accepted or hold | Exact report version covered |
| Narrative input | Accepted values and limitations only | Model cannot choose between unresolved alternatives |
A useful hold note is: “Metric [ID] is not ready for narrative. Comparable source A shows [value] and source B shows [value], producing [difference] under [formula]. [Specific coverage or record question] remains unresolved. [Owner] must confirm [next decision].”
The completion check is that the reviewer can reproduce the accepted figure, inspect its source coverage and identify any permitted limitation. Successful calculation alone is not a completed gate.
Give AI only the accepted narrative inputs
Source: OpenAI Structured Outputs describes schema-constrained responses. A proposed reporting schema can require metric IDs, accepted values, limitations and evidence references. It does not establish that a value is correct or that an explanation is causal. Validate output numbers against the accepted reconciliation record.
Exclude unresolved alternatives from the ordinary narrative input or label them explicitly as a held metric. The model should not average two conflicting totals, choose the latest-looking one or describe the gap as improved efficiency. If the reviewer accepts a limited report, the wording must retain that limit rather than conceal it in an internal exception list.
Source: OpenAI function calling distinguishes model tool requests from application execution. A reporting tool should enforce the accepted gate state before returning narrative-ready values. A request to generate the report cannot bypass a held metric merely because other sections are available.
Walk through agreeing totals and an unresolved difference
In a hypothetical comparable set, accepted row amounts are 100, 150 and 250 units. Their calculated sum is 500. Both accepted source snapshots report the same 500 units and the membership check accounts for all three records. The reviewer accepts that metric for the narrative. AI may describe the recorded total without inventing why it occurred.
In a second hypothetical set, the row calculation remains 500 while another supposedly comparable source shows 490. The defined difference is 10 units. The gate holds the metric and asks the owner to inspect the record-level gap. It does not subtract 10 from a convenient row or call the difference negligible without an accepted rule.
A missing-source case is different: if source B is unavailable, its value is unknown rather than zero. A duplicate imported row might explain an overstatement, but the reviewer must establish identity and rerun the accepted calculation. If a later correction changes the result, retain the earlier report input and create a revised gate decision.
Test the gate rather than the persuasiveness of prose
Evaluate the proposed procedure with fixed fixtures containing agreeing totals, mismatched definitions, omitted records, repeated imports and offsetting errors. Check that the narrative stage receives only accepted figures and approved limitations. An attractive explanation of an unresolved number is a failure, not evidence that the report is useful.
Record who resolves each exception and how the report resumes. Compare actual reviewer effort and corrections with the current process before claiming fewer incorrect reports. The system should finish with accepted evidence or a visible hold, rather than hide uncertainty to meet a reporting deadline.
Questions about reporting reconciliation
Can AI choose the total from the most recent file?
Not without the accepted authority and version rule. A recent file may use a different scope or contain an unreviewed correction. Preserve both references and let the information owner decide which value governs the report. File age alone does not establish comparability or correctness.
Does a zero difference prove the inputs are complete?
No. Missing records or offsetting errors can leave totals unchanged. Check the expected population, exclusions and membership as well as the sum. Reconciliation should explain what the figure covers, not merely demonstrate that two numbers happen to match.
Can the report proceed while one metric remains unresolved?
Only if the responsible reviewer approves a limited scope and the report clearly identifies the held section. Do not imply full completeness or fill the gap with a forecast. Keep the unresolved metric's owner and next action visible until the evidence is accepted.
If your business needs help defining this process, explore Workflow automation, the wider AI automation services, and our custom-agent workflow guide. The agents and automation comparison and custom AI agent glossary explain the terms. To discuss your records and approval rules, get in touch.