KPI exception report for weekly review

Use this workflow to turn one weekly operations export into a short exception report for the review meeting. It is for operations analysts who prepare the evidence, and managers who need clear actions rather than a list of numbers.

The finished report separates missed targets from normal variation. Each exception includes the measure, the missed target, likely causes, missing evidence, an owner and a due date.

1. Set the reporting boundary

Before opening the export, write the scope at the top of your working document. Use the same boundary every week unless the review owner agrees a change.

Record:

• Reporting period, including the start and end date.
• Business unit or site, such as a warehouse, support queue or region.
• Measures in scope, for example on-time dispatch, order backlog, first response time, rework rate or supplier fill rate.
• Target source, such as the approved KPI register, service-level schedule or operating plan.
• Comparison period, normally the previous week and the equivalent prior period if you use one.
• Known events, including system outages, public holidays, stock constraints or process changes.

Do not treat a target printed in the data export as automatically correct. Compare it with the approved target source. A changed threshold can create a false exception or hide a real one.

2. Prepare the data pack

Create a folder for the review and save the original export without editing it. Name the file with the reporting period and source system. Then make a working copy for calculations and notes.

Check that the export contains, where relevant:

• KPI name and definition
• Actual value
• Target value
• Numerator and denominator, or the underlying count
• Date or week ending date
• Site, team, supplier or process segment
• Prior-period value
• Status fields, such as open, delayed, cancelled or completed

Standardise obvious issues in the working copy. Remove blank total rows, make dates consistent and label missing values as missing rather than zero. If a measure is a percentage, preserve both the percentage and its underlying volume. A 2-point miss across 20 cases needs different treatment from the same miss across 20,000 cases.

If you are uploading material to a model, use only the access route your organisation has approved. Check the current product and data-handling documentation before you do so, as capabilities and controls are version-dependent: xAI documentation overview.

3. Define what counts as an exception

Create an exception rule before looking for explanations. This prevents the loudest issue from becoming the only issue discussed.

Use three tests:

1. Target miss: the result is outside its agreed target.
2. Materiality: the gap has meaningful operational impact, such as affected orders, hours, cost exposure or customer cases.
3. Persistence or deterioration: the gap repeats, worsens against the prior period or appears concentrated in one segment.

Write the rule in your report. For example: include every target miss, then rank items by operational impact and trend. Keep borderline items in an appendix or a watch list, rather than mixing them with urgent exceptions.

4. Build an evidence table

Make a table with one row per candidate exception. Do this before asking the model to draft prose. The table is the factual base for the report.

Field	What to enter	Example use
KPI and segment	Exact measure and affected site, queue or supplier	Backlog, North warehouse
Actual and target	Value, direction of target and gap	94%, target 97%, down 3 points
Volume	Underlying cases or units	1,840 orders
Trend	Prior period and direction	Down from 96%
Evidence	Relevant breakdown or event record	Late inbound receipts rose
Evidence gap	Information not available	No shift-level staffing data
Proposed action	A specific investigation or correction	Validate receipt delays by supplier
Proposed owner and due date	Named role and review point	Inbound manager, Thursday

Separate facts from interpretations. “Late inbound receipts rose” may be a fact if the export shows it. “Supplier performance caused the backlog” is a hypothesis unless the data connects the two.

5. Draft the exception report with the model

Paste the evidence table and the reporting boundary into the model. Ask it to retain uncertainty instead of filling gaps with plausible language. Use this prompt structure:

Create a weekly KPI exception report from the evidence table below.

Reporting boundary: [period, scope, KPI definitions]
Exception rule: [rule]

For each exception, provide:
1. KPI, segment, actual, target and gap
2. Operational impact, using only supplied volume or count data
3. Likely causes, labelled as hypotheses unless directly evidenced
4. Evidence gaps and the data needed to test each hypothesis
5. One proposed action, proposed owner role and due date placeholder

Do not invent figures, causes, owners or dates. Exclude KPIs that meet target.
End with a short watch list for borderline items.

Ask for a report with these sections:

• Executive summary: the two or three exceptions needing management attention.
• Exception register: one concise entry per exception.
• Cross-cutting patterns: recurring issues across sites, teams or suppliers.
• Evidence gaps: data to obtain before the next review.
• Action register: proposed owner, action, due date and success measure.
• Watch list: items not yet severe enough for an assigned action.

Keep the action register separate from the narrative. It should be possible to copy it into the meeting notes without rewriting it.

6. Check where the output is wrong

The most common failure is not a false calculation. It is a confident causal story built from a correlation or a missing field. Read every statement and mark it as fact, calculation, hypothesis or proposal. Facts must be traceable to a row in the export or a named supporting record. Calculations must reproduce from the actual, target and volume. Hypotheses must include an evidence gap or a test.

Also check direction. For some measures, lower is better, such as backlog age or rework rate. For others, higher is better, such as service level. A report can state a numerically larger value as a miss if it has misunderstood the KPI direction.

If you see this	Treat it as	What to do
A cause with no source field	Hypothesis presented as fact	Add an evidence gap and request supporting data
A percentage with no volume	Incomplete impact statement	Add the denominator or remove the impact claim
A named owner not in the source	Unapproved assignment	Replace with a proposed owner role for the manager to confirm
An action with no success measure	Activity, not control	State the KPI or evidence that will show completion

7. Hand over the report

Send the report before the weekly review with the original export, the evidence table and any supporting records named in the report. In the meeting, confirm each proposed owner and due date. Record decisions in the action register, including actions rejected because the evidence was insufficient.

After the meeting, retain the final report and action register together. At the next review, start by checking whether previous actions closed, whether the KPI moved and whether the claimed cause was supported. This turns the weekly report into an operating record rather than a fresh narrative each week.

When the workflow does not work

If the output is too broad, tighten the exception rule and provide segment-level evidence. If it invents causes, remove narrative source material and require each claim to cite a table field. If the export lacks denominators, target definitions or prior-period values, stop drafting conclusions and request those fields from the data owner. If the review repeatedly produces unowned actions, have the meeting chair confirm the owner and due date live, then update the register before circulation.