AI management reporting commentary.

What the model writes, and what it must not

AI management reporting commentary is the narrative around the numbers in a board or committee pack: arrears moved from 2.1% to 2.6%, funds under management rose NZ$40 million on market movement and net inflows, complaints doubled in one product. In a financial firm that narrative usually takes a finance or risk analyst most of a day each month, and much of it is restating tables in sentences.

The design rule has two halves. First, every figure is calculated in the warehouse or reporting layer and handed to the model as a table; the model never computes. Second, every explanation of why something moved comes from a note written by the business owner responsible (credit, investments, operations, complaints), and the model quotes or closely paraphrases it with attribution. If a metric moved past its threshold and nobody explained it, the draft does not guess. It lists the metric as an open question.

How the workflow runs

Inputs

Calculated metrics with prior period, budget and thresholds; owners’ notes submitted through a short form; last month’s approved commentary; the pack template and style rules.

Systems

The data warehouse or Power BI model as the source of figures, a form or Teams workflow for owners’ notes, a model API, and a validation script.

Steps

Freeze period data; calculate metrics and flag threshold breaches; request notes from owners for flagged metrics; draft each section from figures and notes; extract every number from the draft and match it to the table; mark unexplained movements; assemble the pack.

Human review

Owners confirm their section. The CFO and chief risk officer edit and approve the full commentary. The company secretary confirms the AI-use note before the pack goes to the board.

Outputs

Draft commentary per section, a numbers-matched validation report, a list of open questions, and an approval record.

Measures

Number mismatches caught by validation (target zero after review), share of sentences edited, hours to final pack, and board questions about unexplained movements.

Risks

Invented causes, wrong comparatives, softened bad news, and stale notes copied from last month.

The NZ supervisory context

The Reserve Bank said in May 2025 that regulated entities are expected to manage AI-related risks as part of their existing obligations. On 15 October 2026 it and the FMA added that boards and senior management need to know where AI is being used, the risks it brings and who is accountable. A pack partly written by AI should say so.

Conduct reporting belongs in the pack too. For firms under CoFI, the FMA expects the governing body to consider the adequacy and effectiveness of the fair conduct programme. Complaints, remediation and product review metrics, with owners’ explanations, give the board something to consider.

Illustrative example: a fund manager’s monthly board pack

Illustrative example, synthetic numbers: a manager running eight retail funds produces a monthly board and investment committee pack with about 60 metrics in 14 commentary sections. Assumptions: writing commentary takes 16 analyst hours a month today; editing drafts takes 6 hours; owners already write short notes, now through a form instead of email.

• Analyst time: about 10 hours a month back, and the pack reaches directors a day earlier.
• Model fees: volume is tiny, so use the strongest model. Three drafts of 14 sections at 15,000 input and 800 output tokens on OpenAI’s gpt-6-astra, at US$10 and US$50 per million tokens (OpenAI, as listed on 19 October 2026), is about US$8 a month.

What stays with people

The CFO and risk lead decide what the board is told, how risks are characterised, and what is recommended. Owners are accountable for their explanations. Directors ask the questions. The model does not assess risk appetite, propose provisioning or write recommendations.

Where commentary goes wrong

• Plausible causes. Ask a model why arrears rose and it will offer reasons. Without an owner’s note, the answer is an open question.
• Softened bad news. Drafts drift toward reassuring tone. Style rules should require plain statements of adverse movements first.
• Copied notes. Last month’s explanation reused for this month’s movement. Date-stamp notes and reject stale ones.
• Tool limits. Microsoft says Copilot in Power BI needs F2 or P1 capacity and that Australian capacities are processed in the US. Check before assuming it covers this.

Grounding every claim in a cited source is the design principle behind Lucent, a Levy Street venture building agentic investment research for wealth, equity and private-markets teams. The practice-level version of this workflow is month end reporting preparation, and our lesson on using AI to prepare management reports covers editing habits. Related: internal policy retrieval, client onboarding document checks, our AI reporting service and the financial services overview.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.