[ Finance ]
AI Board Reporting: Build a Board Pack That Can Answer Why
A governed board-question chain lets AI assemble evidence, decompose drivers, draft commentary, and build scenarios without handing the model authority over the numbers.
On this page
AI board reporting uses AI to assemble governed evidence, reconcile driver analysis, draft commentary, and prepare scenario material. Approved systems still compute every figure, and accountable finance leaders retain the narrative, board ask, and approval. The result is a board pack that can answer why without turning the meeting into a promise to follow up.
Picture the familiar 9:14 p.m. email. The gross-margin slide is finished. Then a director asks why gross margin dipped. Instead of rebuilding the explanation, the analyst opens the existing bridge. Price, mix, freight, foreign exchange, source links, and the finance owner are already attached.
The board pack's hardest problem is rarely slide production. It is what happens after the follow-up. The analyst reopens the bridge, CRM export, and operating workbook to separate price, mix, freight, and foreign exchange. A polished chart shows what happened. A decision-ready pack keeps the route from headline to cause intact.
AI should shorten the route from a reconciled number to its evidence. It should never be the place where the number becomes true. The board-question chain records that route; finance still decides what matters and what authority to request.
What should AI automate in board reporting?
AI can retrieve source records, check tie-out status, assemble the appendix, and format a reviewable first draft. Metric policy stays in controlled finance logic. Judgment and approval stay with people.
| Board-pack stage | AI can assist with | Required evidence | Human owner |
|---|---|---|---|
| Metric assembly | Retrieve approved actual, plan, and prior-period values | Definition, period, entity, currency, source, tie-out | Finance data owner |
| Driver analysis | Draft a bridge from calculated components | Driver formulas, residual, thresholds, transaction drill | FP&A owner |
| Commentary | Turn approved facts into a first draft | Fact packet, exceptions, contrary evidence | CFO or FP&A lead |
| Scenario appendix | Run approved assumption sets and format comparisons | Base-model version, assumptions, sensitivities, bounds | Planning owner |
| Final pack | Check consistency, missing citations, and stale versions | Review log, approvals, distribution controls | CFO and company secretary |
This AI-assisted board-reporting boundary matches the NIST AI Risk Management Framework, which calls for documented knowledge limits and human oversight, plus clear roles and responsibilities. NIST's framework is voluntary and cross-sectoral; it does not prescribe board-pack workflow. Here, the practical controls are named review roles and a record of what the model cannot be trusted to do.
Why does a finished deck fail the first follow-up question?
A finished deck fails when the visual and the analysis are separate products. In an illustrative pack, the slide carries a $2.1 million gross-profit miss. The bridge lives in a workbook. Pricing sits in the CRM, unit volume in an operating system, product mix in a model, and foreign exchange in another tab. The board sees one bar. Finance sees five reconstruction jobs.
Reconcile the driver bridge before anyone writes the slide. The bridge explains the variance with calculated components, exposes any residual, and links each material driver to source evidence. If the components do not tie to the headline, AI stops and surfaces the gap. A fluent paragraph cannot repair an unreconciled bridge.
The illustrative 240-basis-point gross-margin bridge is arithmetic, not customer evidence. The control is real: a board pack should distinguish a calculated driver from management's interpretation of why the driver moved. Freight expense can be calculated. Whether the increase reflects poor contracting, a temporary lane disruption, or a deliberate service decision requires operating context.
How do you build a board-question chain?
For each board question, Pluvo's editorial framework records the decision requested, the metric and baseline, the calculated bridge, the source record, and the accountable person. It is not an accounting or governance standard. A missing link becomes a visible stop condition.
1. Start with the decision
Name whether the item is for information, discussion, recommendation, or decision. The Governance Institute of Australia's board-papers guidance recommends putting the purpose, recommendation, and action required up front. AI cannot prioritize a pack intelligently when it does not know what the board is being asked to do.
2. Lock the metric contract
A gross-margin KPI needs a locked metric definition, calculation, reporting period, entity scope, currency, owner, and version. SEC Release 33-10751 governs public-company KPI and metric disclosures in MD&A, not internal board packs. Its requirements around definitions, calculation methods, usefulness, and changes are worth applying upstream when a board metric may later reach investors.
3. State the comparison
A reforecast can make a miss vanish from the slide if it silently replaces the original budget. Name the comparison: actual to plan, forecast, prior period, prior year, or a board-approved target. Then preserve that baseline version.
4. Reconcile the decomposition
Calculate the bridge before drafting prose. Name the residual and set a stop rule. If identified drivers explain $1.82 million of a $2.10 million miss, the unexplained $280,000 belongs on the page. Hiding the residual converts uncertainty into false confidence.
5. Attach evidence and an owner
The review record links the claim to its source population, transformation, and calculation. It also names two people when needed: the metric owner who can defend the number and the action owner who must do something about it.
How should AI draft board commentary?
Draft commentary from a controlled fact packet, not a chart image and a hopeful prompt. The packet needs the approved result and baseline, reconciled bridge, residual, threshold, source links, contrary evidence, owner, and intended decision.
A bare LLM predicts text; it is not a controlled calculation engine. Tool-enabled systems can execute code, but every material figure stays in deterministic calculation built on approved data and deterministic logic. If the model encounters two definitions of gross margin or a bridge that does not tie, the correct output is an exception, not an eloquent compromise.
The first draft should be plain enough to challenge. Phrases such as resilient performance, strategic momentum, or temporary headwinds must earn their place with a measured definition. If finance cannot specify the period, amount, driver, and evidence, the adjective is doing accounting work.
What belongs in an AI-generated scenario appendix?
A scenario appendix starts from an approved base model. Each branch records its base version, changed assumptions, sensitivity range, owner, run time, and comparison to the current forecast. The uncertainty belongs on the page, not under a prettier fan chart.
“Give me an upside case” is a writing prompt. “Delay 12 planned hires by 30 days and hold price and churn constant” is a scenario instruction. The model can format the result, but finance approves the causal links and bounds.
Keep rejected scenarios. When a director asks what changed since the last meeting, show the moved assumptions, editor, and split between actual performance and management's revised view.
What must stay human in the boardroom?
People decide which variance matters, what tradeoff management accepts, and what authority it wants. AI can rank a variance by size. Without complete context and accountable judgment, AI should not be trusted to infer that the smallest line exposes the biggest strategic problem or that a director has heard a different account from a customer.
The human role is not ceremonial. COSO's 2026 GenAI guidance applies internal-control principles to generative AI and highlights opaque reasoning, model drift, and frequent configuration changes. A reviewer should reopen one cited source and challenge one changed assumption. Clicking approve after reading the summary is review theater.
The board also needs management's judgment in its own voice. A recommendation should state the risk management accepts and the alternative it rejected. No model can accept responsibility for that choice.
How does Pluvo make the board-question chain reviewable?
In a Pluvo board-reporting workflow, the gross-margin definition lives in the ontology. The rendered claim opens through lineage to the bridge, transformations, and source records. Reports keep each figure cited and traceable to source records.
AI investigates and explains; approved systems compute. Pluvo's quality bar is blunt: for cash, covenant, or board-reported figures, a number that's 95% right is 100% useless.
The 15-point board pack QA checklist
Run the 15-point board pack QA checklist before distribution. It checks the pack itself; the broader AI governance checklist for finance examines the system producing it.
| # | QA question | Pass evidence |
|---|---|---|
| 1 | Is the item for information, discussion, recommendation, or decision? | Purpose stated on the first page |
| 2 | Is the board ask explicit? | Resolution, challenge, or support requested |
| 3 | Does every material KPI have an approved definition? | Definition, formula, owner, and version |
| 4 | Are period, entity scope, and currency clear? | Reporting basis shown beside the metric |
| 5 | Can every material figure reach an approved source? | Source-to-output lineage |
| 6 | Do totals and subtotals tie? | Control totals and reconciliation status |
| 7 | Does each material variance have a calculated bridge? | Driver components equal the headline |
| 8 | Is any residual visible? | Amount, threshold, owner, and next step |
| 9 | Does commentary use only approved facts? | Fact packet and source links |
| 10 | Is contrary evidence visible? | Exceptions and unresolved findings |
| 11 | Are scenario assumptions versioned? | Base version, changed drivers, bounds, owner |
| 12 | Are risks and alternatives stated? | Tradeoffs and rejected options |
| 13 | Is every action assigned? | Named owner and due date |
| 14 | Are access and distribution appropriate? | Approved recipients and sensitive-data check |
| 15 | Did an accountable person challenge and approve the pack? | Review record with exceptions, changes, and sign-off |
For field notes on reconciliations, scenario controls, and finance-system design, subscribe to the Finance Engineering Newsletter.
When the director's email arrives at 9:14 p.m., the useful pack already links the margin miss to the reconciled bridge.
Frequently asked questions
What is AI board reporting?
AI board reporting uses AI to assemble source evidence, investigate calculated drivers, draft commentary, and prepare governed scenario material for a board pack. Material figures should still come from approved calculations, and accountable people should own the narrative, decision request, and review.
Can AI create a board deck automatically?
AI can assemble and format a first draft, but autonomous deck generation is unsafe when the system can invent figures, omit contrary evidence, or detach commentary from governed calculations. Finance should require source lineage, tie-outs, versioned assumptions, and human approval.
How should AI write board commentary?
AI should draft commentary only from a controlled fact packet containing the approved metric, comparison, calculated driver bridge, materiality threshold, source links, scenario assumptions, and exceptions. A finance owner should then decide what matters and what action to request.
What should stay human in AI board reporting?
People should own materiality, the narrative arc, uncomfortable tradeoffs, the board ask, disclosure judgment, risk acceptance, and the live discussion. AI cannot know which fact will change a director's decision or accept responsibility for the answer.
What should a board pack QA checklist cover?
A board pack QA checklist should cover purpose, decision request, metric definitions, reporting period, source lineage, tie-outs, driver bridges, residuals, commentary support, scenario assumptions, risks, owners, version dates, access controls, and final human approval.
