Applied AI library

Quarterly Commentary Writer

Turn financial data extract, controller notes and past management actions into commentary and recommendations.

At a glance

Without an agentic workflow, the starting point is a data extract in CSV format. A finance analyst compares the current and prior-period figures, reads controller notes to understand the movements, and reviews past management actions to assess their outcomes.

The Quarterly Commentary Writer turns these inputs into concise commentary by analyzing the movements, identifying the largest changes, explaining what happened, and recommending what to do next. It automatically extracts useful information, checks for missing explanations and outdated inputs, and creates a draft that follows the team’s required style guide.

  1. Person: Provide approved figures for the current and comparison periods.
  2. Code: Calculate movements and check the regional breakdown against the total.
  3. Person: Provide explanations for material movements in the reporting period.
  4. Agent: Tie explanation claims to their source passages and regions.
  5. Person: Provide dated management actions and recorded outcomes.
  6. Agent: Extract action IDs, decision periods, and supported outcomes.
  7. Person: Provide approved format, materiality, and evidence rules.
  8. Agent: Draft a sourced paragraph and proposed next actions.
  9. Code: Check figures, coverage, format, and source evidence.
  10. Person: Approve, edit, request evidence, or reject the draft.
Primary user
Finance reporting analyst
When to use
Quarterly commentary is due
Inputs
Financial data, notes, past actions, style guide
Outputs
Draft commentary and proposed actions

Preview the workflow steps

Follow one Claims Incurred example from the input files through calculation, explanation, drafting, and review. Select a step to inspect its example.

Person

Supply the figures and supporting evidence

The analyst selects the CSV extract, controller notes, past management actions, and commentary guide for 2026 Q1 versus 2025 Q1.

Commentary inputs

Income statement CSV excerpt
line_itemregion2026 Q12025 Q1movement
Claims IncurredEMEA-158.2-143.1-15.1
Claims IncurredAmericas-51.6-45.6-6.0
Claims IncurredAsia-67.8-58.4-9.4
Claims IncurredTotal-277.6-247.1-30.5

Step-by-step guide

The chapters below explain how to get to the result above by building an agentic workflow. Each chapter includes the desired outcome and the steps to get there.

Back to overview

CHAPTER 1

Define task

Outcome

You have a clearly defined task, with a business purpose, explicit scope, and measurable success criteria.

Why it matters

Explanations for management need to highlight material drivers, useful next steps and evidence for every fact. Defining the task first makes clear which decisions the commentary workflow should support and what is out of scope.

What to do

  1. 1Set the objectiveDescribe the purpose, intended audience, and scope of the commentary.
  2. 2Map current processMap how inputs are collected, calculations are performed, and commentary is drafted, reviewed, and approved. Identify who handles what and how the steps are connected.
  3. 3Define the scopeThe agentic workflow processes inputs and drafts the commentary. Publication and communication remain with qualified reviewers.
  4. 4Set acceptance criteriaDefine what a good result looks like at each step. Require correct figures, supported driver explanations, current action outcomes, and compliance with the commentary guide.
  5. 5Choose a test caseUse a CSV extract, controller notes, and past management actions that a reviewer can check against an expected result.

Example

An analyst prepares the Q1 2026 Claims Incurred commentary using a financial data extract, controller notes, and past management actions. The workflow takes over the analysis and drafting, then sends the commentary and recommendations to the finance reviewer for approval. Publication and implementation of recommendations remain outside the workflow.

CHAPTER 2

Prepare inputs

Outcome

You have a sample set of real input data, ready to use when building and testing the workflow.

Why it matters

Representative inputs give you a practical basis for building and testing the workflow. Numbers alone cannot explain why results changed or whether management actions worked. Controller notes and action records provide the context needed to support those conclusions.

What to do

  1. 1Define sources and ownersName the owner of the CSV extract, controller notes, past management actions, and commentary guide.
  2. 2Prepare the dataMap regional labels to one name and use a consistent sign convention, such as expenses as negative values.
  3. 3Assign reference IDsGive each dataset row, email, and action record a stable ID and reporting period.
  4. 4Add input checksCheck that all relevant regions are covered consistently across the input files, including supporting notes and applicable action records.

Example

The input set contains an income-statement CSV, controller-notes emails, a management-actions CSV, and the abbreviated commentary guide. The CSV includes regional figures and the overall total. It is the only numerical source used for the movement calculation.

Commentary workflow inputs

List of inputs used by the quarterly commentary workflow.

commentary-inputs/4 files
income-statement.csv
Income statement CSV excerpt
line_itemregion2026 Q12025 Q1movement
Claims IncurredEMEA-158.2-143.1-15.1
Claims IncurredAmericas-51.6-45.6-6.0
Claims IncurredAsia-67.8-58.4-9.4
Additional CSV row value not shownAdditional CSV row value not shownAdditional CSV row value not shownAdditional CSV row value not shownAdditional CSV row value not shown

CHAPTER 3

Design workflow

Outcome

You have a blueprint for building the workflow, from inputs to final output.

Why it matters

Designing the workflow before building it makes the processing steps and their connections explicit. For commentary, this clarifies how figures and controller notes become a draft.

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, calculated movements linked to controller notes provide the basis for drafting the commentary.
  2. 2Specify each stepDescribe what each step receives, how it should process the information, and what it passes on to the next step. For instance, specify how calculated movements are linked to controller notes and how past action outcomes are incorporated into the draft.
  3. 3Define exception handlingDefine how to handle missing explanations or outdated management actions. Define when to mark a draft as incomplete and request additional information from the reviewer.
  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 the commentary, include fields for movement explanations, recommendations, and source references.

Example

The workflow separates calculations from interpretation so figures and explanations can be checked independently before drafting. Separating calculations from interpretation lets you check numerical accuracy and the quality of explanations independently. The drafting step then brings those results together.

Commentary workflow

  1. Person: Provide approved figures for the current and comparison periods.
  2. Code: Calculate movements and check the regional breakdown against the total.
  3. Person: Provide explanations for material movements in the reporting period.
  4. Agent: Tie explanation claims to their source passages and regions.
  5. Person: Provide dated management actions and recorded outcomes.
  6. Agent: Extract action IDs, decision periods, and supported outcomes.
  7. Person: Provide approved format, materiality, and evidence rules.
  8. Agent: Draft a sourced paragraph and proposed next actions.
  9. Code: Check figures, coverage, format, and source evidence.
  10. Person: Approve, edit, request evidence, or reject the draft.

Selected stage

Supply CSV data

Input

Approved financial extract and selected reporting periods

Output

Current and comparator figures for the selected line item

The final output could follow a schema like this:

Commentary output schema

{
  "status": "drafted",
  "line_item": "Claims Incurred",
  "current_period": "2026 Q1",
  "previous_period": "2025 Q1",
  "total_movement_usdm": 30.5,
  "movement_percent": 12.3,
  "regional_coverage": [{
    "region": "EMEA",
    "movement_usdm": 15.1,
    "evidence_state": "covered",
    "source_ids": ["income-statement.csv", "controller-notes.eml"]
  }],
  "action_assessments": [{
    "action_id": "ACT-2025Q3-01",
    "outcome_period": "2026 Q1",
    "evidence_state": "covered",
    "proposed_next_action": "Maintain cover at renewal"
  }],
  "commentary": "...",
  "recommendations": ["..."],
  "review_status": "pending"
}

CHAPTER 4

Build AI workflow

Outcome

You have a working first version that follows the blueprint and produces the expected output.

Why it matters

A working first version gives you something concrete to improve. Building a first version reveals whether the proposed steps can work together on actual commentary inputs.

What to do

  1. 1Assign each stepDecide whether code, an AI agent, or a person should carry out each step. Use code for movement calculations, agents for interpretation and drafting, and a finance reviewer for approval.
  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 drafting step, supply calculated movements, controller explanations, and past management action outcomes alongside the commentary guide.
  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 commentary draft.
  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 movement calculations in code so the drafting agent uses fixed figures. Connect those results to the extracted controller explanations and action outcomes, preserving their reference IDs. Run the sample inputs to produce a first commentary draft in the agreed schema.

Commentary workflow

  1. Person: Provide approved figures for the current and comparison periods.
  2. Code: Calculate movements and check the regional breakdown against the total.
  3. Person: Provide explanations for material movements in the reporting period.
  4. Agent: Tie explanation claims to their source passages and regions.
  5. Person: Provide dated management actions and recorded outcomes.
  6. Agent: Extract action IDs, decision periods, and supported outcomes.
  7. Person: Provide approved format, materiality, and evidence rules.
  8. Agent: Draft a sourced paragraph and proposed next actions.
  9. Code: Check figures, coverage, format, and source evidence.
  10. Person: Approve, edit, request evidence, or reject the draft.

Selected stage

Supply CSV data

Person

Output

Current and comparator figures for the selected line item

Guardrail

Require the period, comparator, unit, sign convention, and approved version.

CHAPTER 5

Add controls

Outcome

You have checks in place that catch invalid results and raise them for review.

Why it matters

Controls reduce the risk of users acting on incorrect results. They also give reviewers the evidence needed to investigate an issue without retracing the entire workflow.

What to do

  1. 1Identify possible failuresWalk through the existing workflow steps and identify errors that could compromise the final output. For commentary, consider incorrect amounts, unsupported explanations, and outdated management action outcomes.
  2. 2Implement validationsImplement checks for those errors. The checks can be performed by code or agents. For instance, compare draft amounts with calculated results in code and use an agent to assess explanations against the controller notes.
  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, highlight an amount that differs from the calculated result and show both amounts with their sources.

Example

A commentary draft contains a movement amount that differs from the calculated result. A code check highlights the discrepancy and adds it to the review summary. The reviewer receives the draft with both values and their source references so the amount can be corrected.

Control ownership

Assign each check to code, a model, or a person.

Filter controls by owner

Every selected period, region, and line item is present

Code

Movement percentage matches code calculation

Code

A controller note supports each material regional movement

Model

The action ID exists and its outcome period is current

Code

The proposed driver explanation and action assessment cite supporting sources

Model

The commentary follows heading, length, amount, and tense rules

Code

The sourced explanations, action assessments, and proposed actions are correct and can be approved

Person

CHAPTER 6

Build evaluation set

Outcome

You have a set of test cases and scoring criteria to systematically measure workflow quality and reveal where improvements are needed.

Why it matters

A successful run does not show how the workflow handles difficult cases. Evaluations give you a consistent basis for comparing versions and detecting when a change makes results better or worse.

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 complete inputs, missing controller explanations, and outdated management action outcomes.
  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 includes a material movement with no explanation in the controller notes. The expected draft reports the correct amount and identifies the missing explanation.

Evaluation results

Test cases

3

Pass rate

67%

Latency

4.2s

Cost per task

$0.26

Test caseExpectedPassed
Score ≥ 90Failed
Input
Draft claims commentary from the regional figures, controller notes, and current management actions. The total movement is USD 30.5 million adverse, made up of EMEA USD 15.1 million, Americas USD 6.0 million, and Asia USD 9.4 million.

Grading

84 / 100

Driver accuracy20/24
Management actions12/20
Total variance16/16
Format12/12
Style & language10/12
Materiality6/8
Regional amounts4/4
Commentary quality4/4
Score ≥ 90Passed
Score ≥ 90Passed

CHAPTER 7

Design for everyday use

Outcome

You have an operating model that defines who reviews the results, supported by a user interface for everyday operations.

Why it matters

A user interface brings the results, supporting evidence, and review actions together for everyday use. The reviewer must be able to challenge the workflow outputs without having to redo all the work.

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 commentary draft and checks the figures and supporting explanations. They approve the draft, make corrections, or request further evidence through the same interface. The audit log records their actions and final decision.

Commentary review

Claims incurred - 2026 Q1

Draft commentary

Missing

Incomplete draft: Explanation omitted in this example.

Proposed next actions

Regional coverage

Regional coverage status
RegionUSDmExplanationSources
EMEACovered
AmericasCovered
AsiaMissing
TotalCovered
Sources

Select a file to inspect the evidence used by the draft.

Review source files
Income statement CSV excerpt
line_itemregion2026 Q12025 Q1movement

Production-readiness checklist

Before releasing the Quarterly Commentary Writer to production, the following additional steps are required:

Performance Efficiency

3 items

Reliability

3 items

Security

2 items

Cost Control

3 items

Operational Excellence

4 items

End of this guide

Return to the overview, or explore the other guides in the Applied AI library.