AI earnings call preparation: a controlled workflow for CFOs and IR leaders

AI can help a CFO and investor-relations team prepare an earnings call by assembling approved evidence, locating prior commitments, organizing likely questions, and drafting internal talking points. It should not decide what is material, determine whether a statement is compliant, set guidance, or communicate with the market. Those remain accountable human decisions.

That boundary matters because earnings preparation is not simply a writing task. The CEO and CFO need a coherent account of the quarter that can survive questions about results, drivers, outlook, and prior statements. Investor relations needs consistency across the release, prepared remarks, slides, and Q&A. Finance, legal, and disclosure owners need to know exactly which figures and claims the team reviewed.

For a public company, external communications also have legal and regulatory implications. The SEC says Regulation FD addresses selective disclosure of material nonpublic information; intentional disclosure to specified market professionals or security holders must be accompanied by simultaneous public disclosure. The SEC’s Regulation FD guidance is a useful starting point, but it is not a substitute for a company’s counsel and disclosure process.

The practical first use case is not “have AI write the earnings call.” It is “prepare a traceable internal packet for one recurring earnings-preparation job.” This guide describes how to create that packet.

Start with the external decision, not the transcript

An earnings call may look like one event, but its preparation spans several separate decisions:

• What results are final enough to discuss?
• Which movements require an explanation rather than a number alone?
• Which prior commitments, guidance statements, and operating events are relevant?
• Which questions need a factual answer, a qualified response, or an explicit “we cannot address that” boundary?
• Who can approve remarks, Q&A positions, and changes before the call?

Define one job that produces a specific internal output. For example:

After finance confirms the approved earnings snapshot, prepare a Q&A evidence packet for the CFO and head of investor relations. Compare approved results with the prior disclosed period and the company’s own previously published statements; identify material movements and source-backed context; group likely questions; and draft internal talking points with visible gaps. Do not create external communications, change guidance, or make a disclosure judgment.

This statement establishes scope, timing, evidence, audience, and stop conditions. It also prevents a model from filling a missing explanation with an invented one simply because a generic prompt asks for an “earnings narrative.”

The scope is different from AI financial close. The close establishes the controlled financial state. Earnings preparation starts only after the relevant finance and disclosure process says which inputs may be used. It is also different from AI board reporting: a board pack helps directors govern; an earnings-preparation packet helps leaders decide what they can responsibly say outside the company.

Build an earnings-preparation contract

Before connecting an AI workflow to any documents or data, write a short contract that the CFO, investor-relations lead, finance owner, and appropriate legal or disclosure stakeholders can review.

The contract should name what changes invalidate a run. A late close adjustment, revised forecast, new legal issue, or change to the release draft should trigger a refreshed packet and a visible notice of which sections may have changed. Otherwise, teams can rehearse from one version while the final materials reflect another.

Separate the evidence layer from the talking-point layer

The biggest reliability problem in earnings preparation is usually not poor prose. It is allowing a fluent draft to blur facts, explanations, judgment, and commitments.

Use separate layers:

1. Reported fact: approved financial or operating result, with period, definition, and source.
2. Observed driver: a reconciled movement or record-level concentration, with the method that produced it.
3. Management context: an accountable owner’s explanation, clearly attributed and linked to support where available.
4. External position: what the company may say, including qualifications, disclosure implications, and approval status.

AI can help assemble the first three layers and expose disagreement or missing support. It should not promote a management explanation into an external position. For instance, “a handful of enterprise renewals moved” may be a source-backed observation; “renewal timing will normalize next quarter” is a forward-looking management position that needs the company’s normal review.

This distinction mirrors the discipline in an AI variance-analysis workflow: calculations and reconciled bridges establish the amount, while the business explanation requires evidence and named judgment. In earnings preparation, the cost of conflating those layers is higher because the audience is external.

Create a claim register before drafting Q&A

Prepare a compact register for every material result, recurring investor topic, and likely question. The register is an internal review tool, not a script.

The key test is simple: can a reviewer move from a proposed line to the supporting figures, source documents, and accountable owner without searching through chat, email, and spreadsheets? If not, the system has produced a summary, not a controlled preparation workflow.

Jovis is designed for the underlying business-data problem: a team needs to investigate approved context across systems, inspect the basis of an answer, and work from shared rather than private prompts. An earnings-preparation evaluation should stay bounded to one packet, one approved source set, and the company’s existing disclosure review process.

Use AI for question coverage, not answer authority

A useful preparation packet helps the team test what it might be asked. AI can compare prior public materials, approved performance data, analyst or investor questions that the team is permitted to use, and named operating events. It can then cluster topics such as growth, margin, retention, demand, capacity, capital allocation, or execution risk.

But a question list is not a prediction engine, and a drafted answer is not clearance to use it. Require every proposed talking point to fall into one of four lanes:

• Supported response: evidence and approved position are complete.
• Needs owner input: the underlying data is clear but management context is incomplete.
• Needs disclosure or legal review: the topic may involve materiality, forward-looking information, non-GAAP presentation, confidentiality, or another review boundary.
• Do not answer from this workflow: the evidence is unavailable, restricted, speculative, or outside the approved scope.

This makes uncertainty operational. Instead of generating a more persuasive answer, the workflow gives the CFO a short list of issues to resolve before the call.

For U.S. registrants, the treatment of non-GAAP measures also deserves an explicit lane. The SEC’s conditions for use of non-GAAP financial measures explain that Regulation G requires, when applicable, a presentation of the most directly comparable GAAP measure and a reconciliation. Do not rely on a model to decide whether a metric meets those requirements; link approved finance and legal materials instead.

Preserve a controlled snapshot

Earnings preparation changes quickly. A source may refresh, a forecast version may move, someone may revise a sentence, or an executive may learn new context in a rehearsal. Preserve the state behind every packet:

• reporting cutoff, extraction time, and entity or segment scope;
• finance-model, forecast, and metric-definition versions;
• release, slides, prior transcript, and filing versions considered;
• source documents and calculations used for each material claim;
• model instructions and workflow version, where relevant;
• reviewer edits, approvals, and unresolved questions; and
• any change after the snapshot, plus the decision to rerun or accept it.

This is an evidence discipline, not a request to duplicate every system. The appropriate design might retrieve live records through governed access while recording the exact query, version, and time used. The guide to connecting AI agents to enterprise data explains why source authority, data grain, freshness, and access boundaries need to be designed for the job rather than added after the fact.

Do not let the workflow silently substitute a newer number or a different definition. A reviewer needs to know whether a statement was prepared from the release snapshot or from a later management view. When timing changes, mark the packet stale and rerun the affected sections.

Set a review path that matches the consequence

An earnings-preparation workflow should make ownership more obvious, not make it easier to bypass it. Map the review path before the first live cycle.

No table can prescribe a universal approval process. Public-company obligations, jurisdictions, exchange rules, and internal governance vary. The point is to prevent “the AI draft said so” from becoming an implicit approval path.

Access should be equally explicit. A preparation packet may include unreleased results, strategic plans, customer details, personnel data, or legal advice. Limit retrieval and visibility to the people and records required for the defined job. The AI-agent permissions guide provides a useful model: separate source access, workflow visibility, and the ability to take consequential actions. In this case, the workflow should be read-first and draft-only; external publishing belongs outside it.

Rehearse with challenge cases, not only a clean quarter

Before relying on the workflow during a reporting cycle, test it on closed periods and known difficult cases. Include:

• a result where a reported metric differs from the forecast but has a clear reconciled bridge;
• a prior statement whose context changed materially;
• a non-GAAP measure that requires the approved supporting presentation;
• competing explanations from finance and an operating team;
• a late adjustment that should invalidate part of the packet;
• a restricted document that the workflow must not retrieve or expose;
• a question for which the correct response is to escalate rather than draft; and
• a source outage or stale report that should halt the run.

Score evidence coverage, source selection, calculation reproducibility, permission behavior, correct escalation, and reviewer correction separately. A polished output should not compensate for an access failure or an unsupported claim.

NIST’s voluntary AI Risk Management Framework organizes AI risk work around governing, mapping, measuring, and managing risk. Applied here, that means defining the job and owners, testing realistic conditions, documenting failures, and changing the workflow when its sources or consequence change. It does not replace securities, legal, accounting, privacy, or compliance advice.

Pilot one preparation packet before broadening the use

Begin with a repeatable slice of work, such as prior-public-statement comparison or a Q&A evidence packet for one business topic. In the first cycle, run it alongside the current process. Have each reviewer identify unsupported claims, missing context, incorrect source choice, stale inputs, and questions the workflow should have escalated.

In the next cycle, refine the contract, claim register, and test set. Measure whether the packet helps the team find gaps earlier, makes rehearsal more specific, and leaves a clearer record of what was reviewed. Do not measure success only by how quickly an outline appeared.

The goal is a more prepared leadership team, not an automated investor-relations function. Use AI to make the evidence trail and open questions easier to handle; keep the judgment, approvals, and external communication with the people accountable for them.

If earnings preparation repeatedly requires leaders to reconstruct approved business context across systems, evaluate Jovis for one controlled preparation workflow. Start with a narrow packet, named reviewers, and a standard that every material talking point can be traced, challenged, and approved.