How to Turn a Source Packet into a Reviewable Board Update with ChatGPT for PowerPoint

A reviewable board update is not merely a shorter presentation. It is an editable deck whose proposed changes can be traced back to an approved source packet, whose protected slides and content have been identified in advance, and whose claims, numbers, citations where available, and visual changes are checked by accountable people before circulation. This tutorial uses the PowerPoint add-in rather than treating every ChatGPT presentation surface as interchangeable. The first task is therefore to choose the correct surface and establish whether the packet may be processed there at all.

Evidence checkpoints

Documented point: The PowerPoint-native sidebar can create, edit, understand and polish presentations while preserving editable slide structure. [OpenAI documentation: ChatGPT for PowerPoint]

Documented point: OpenAI’s Business release notes record general availability for Business workspaces on 6 July 2026. [OpenAI documentation: ChatGPT business release notes]

Documented point: OpenAI’s guidance describes creating or revising Google Slides or PowerPoint presentations from source material, a reference deck or a reusable template. [OpenAI documentation: Generate slide decks]

Documented point: OpenAI’s guidance describes a source-material, reference-file and template workflow that asks what must remain unchanged. [OpenAI documentation: Creating and editing documents spreadsheets and presentations with ChatGPT work]

Documented point: OpenAI’s Presentations capability can create, edit, inspect, render, verify and export decks locally, including PowerPoint presentation (.pptx) file authoring. [OpenAI documentation: Presentations]

Documented point: OpenAI says content in individual ChatGPT services may be used to improve model performance depending on user settings. [OpenAI documentation: Data controls FAQ]

Define the job before opening the source packet

The reader’s job is bounded: turn access-cleared notes, key performance indicators (KPIs), documents and an existing Microsoft PowerPoint file into a proposed board update while retaining editable slide structure and preserving designated material. It is not to ask artificial intelligence (AI)Computer systems designed to perform tasks that normally require human intelligence, such as understanding language, recognising patterns, or making predictions. Open glossary entry to decide what the board should believe, certify the figures or approve the final deck.

Begin by writing a one-page job specification outside ChatGPT. Record the audience, reporting period, decision required, approved source set, slides that may change, slides that must remain unchanged and the people responsible for verification. This specification separates editorial intent from source evidence. For example, “Explain the margin variance” is an editorial objective; “April–June management accounts, approved 14 July” identifies the evidence that may support it.

A practical specification can use these fields:

• Deck: exact file name, version and reporting period.
• Audience: board, committee or executive group, including any different information entitlements.
• Purpose: inform, request a decision, record a risk or propose an action.
• Permitted sources: named files, approved notes and enabled connected sources.
• Excluded sources: drafts, privileged advice, personal data or repositories not cleared for this workflow.
• Editable range: named slide numbers or sections.
• Immutable range: slides, wording, branding elements or approved disclosures that must not change.
• Required record: proposed change log, claim-to-source mapping, unresolved questions and unsupported-claim list.
• Review owners: finance, key performance indicator (KPI)A defined measure of progress towards an intended result. The owner should specify its calculation, period and source so that an unchanged label is not mistaken for an unchanged measure. Open glossary entry owner, company secretariat, legal or compliance reviewer where required, and final deck approver.

Worked fictional example: Northstar Components is preparing a quarterly board update. Its approved packet consists of an existing deck named Northstar_Q2_Board_v7.pptx, signed-off management accounts, an operations note and a risk-register extract. Slides 1–3 and 12 are immutable; slides 4–9 may be revised; slides 10–11 may receive wording corrections but no change to risk ratings. Draft sales forecasts and an employment-law memorandum are excluded. The requested output is a proposal for edits, not an approved board pack.

The procedure is to compare every candidate file with the specification before it is uploaded, attached or made available through an app. Remove duplicates, superseded drafts and material that is not needed for the named slides. If a necessary source lacks approval or its owner cannot be identified, stop and resolve that issue rather than compensating with a broad prompt.

Decision rule: use the workflow only when the source boundary, editable slide range and human approvers are explicit. The trade-off is that a smaller packet may provide less background, but it reduces ambiguity and unnecessary disclosure. Do not enlarge the packet merely in case the system might find something useful. Keep secrets and untrusted data out of prompts, and do not paste credentials, access tokens or confidential material that has not been authorised for this processing path.

Choose the PowerPoint-native surface, not a similarly named workflow

OpenAI documents several presentation-related surfaces, but they are not procedural substitutes for one another. The dedicated ChatGPT for PowerPoint experience operates in a PowerPoint-native sidebar and is documented as able to create, edit, understand and polish presentations while retaining editable slide structure. OpenAI’s developer guidance directs work on an open presentation to that add-in.

By contrast, the generic ChatGPT Work article describes workflows using source material, reference files and templates but, when inspected on 4 October 2026, explicitly stated that PowerPoint was not included in the Work desktop flow at launch. OpenAI also documents a separate Presentations capability that can create, inspect, render, verify and export decks, including PowerPoint Presentation (.pptx) files. Its availability depends on the plugin, plan and workspace settings. These surfaces may be useful for other jobs, but instructions for them should not be presented as instructions for revising the deck currently open in PowerPoint.

Use this identification procedure before providing any board material:

1. Open the duplicate working copy in PowerPoint.
2. Confirm that the available ChatGPT experience is the dedicated PowerPoint integration, rather than a browser conversation, generic Work desktop flow or separate presentation-generation capability.
3. Ask the workspace owner or administrator to confirm how the add-in was deployed and which account and workspace it uses.
4. Confirm that the intended user is signed into the authorised organisational account, not a personal account.
5. Record the surface, workspace and deck version in the job specification.

Worked fictional example: Northstar’s analyst can access a presentation tool in a browser and can also see ChatGPT beside the open board deck in PowerPoint. Because the task is to revise named slides in that open deck, the analyst selects the latter route. The browser tool is not treated as equivalent merely because both can produce .pptx files. Before attaching any source, the analyst records “dedicated PowerPoint integration; Northstar Business workspace; duplicate deck v7-working-copy” in the review record.

Decision rule: if the required outcome is an edit to a currently open, existing deck with specified surrounding slides preserved, use the dedicated PowerPoint route documented for that job. If the available surface can only generate or export a separate deck, pause and redesign the workflow rather than assuming identical edit, preservation or permission behaviour. The trade-off is between direct contextual editing and a more detached file-generation process; neither removes the requirement to compare the finished slides with the source deck.

Before using the add-in with a board source packet, confirm eligibility in the actual workspace. The official Business release note cited here records Business general availability on 6 July 2026. For separate background on the PowerPoint add-in, ChatGPT for PowerPoint Goes GA discusses an availability announcement, but its catalogue wording does not establish this workspace’s present plan or rollout.

Verify entitlement, deployment and controls independently

“Available globally” does not mean “usable by every employee in every tenant”. As inspected on 4 October 2026, OpenAI’s core PowerPoint documentation listed availability across Free, Go, Plus, Pro, Business, Enterprise, ChatGPT Edu and K–12 plans, with limited usage for Free and Go. OpenAI’s Business release notes separately record general availability for Business workspaces on 6 July 2026. Actual use can still depend on plan limits, workspace settings, rollout, add-in access, administrator approval, supported-app controls and permissions for connected sources.

role-based access control (RBAC)A method of assigning permissions through defined roles rather than one person at a time. Open glossary entry matters because entitlement to ChatGPT does not necessarily establish entitlement to the add-in, every enabled app or every document behind a connection. A user may be able to open a deck but lack access to an approved source repository; conversely, a broad connection may expose files beyond the board-update packet. Verify each layer rather than inferring one permission from another.

Ask the workspace owner and Microsoft environment administrator to complete this preflight:

Check	Question to answer	Evidence to record
Plan	Which ChatGPT plan and organisational workspace apply to this user?	Administrator confirmation and date
Add-in deployment	Is the dedicated integration approved and available in this PowerPoint environment?	Approved deployment route or internal service record
User role	Does the user’s RBAC role permit this feature and this class of document?	Relevant role or policy reference
Apps and skills	Which apps and skills are enabled for this workspace and user?	Administrator-confirmed list
Source entitlement	May the user access and process each named source for this purpose?	Source owner or data-owner approval
Usage limits	What limits or credits apply, and who monitors them?	Current tenant guidance, dated
Review chain	Who must verify figures, narrative, risk statements and final publication?	Named owners and sign-off order

Worked fictional example: Northstar’s finance analyst has Business access, but the organisation has enabled only an approved document repository, not the team’s operational notes app. The signed accounts can therefore be made available through the approved route, while the operations owner supplies an access-cleared extract instead of asking the analyst to bypass the restriction. The limitation changes the packet; it does not justify copying content from an unapproved system into the prompt.

Decision rule: proceed only when the user, feature, source and intended processing purpose are all authorised. If one layer is unclear, use a manually minimised, approved extract only if the data owner permits it; otherwise stop. The trade-off is convenience versus controlled access. Fewer connections can mean more preparation, but broadening permissions solely to accelerate one board update may expose unrelated information and weaken the packet boundary.

Do not invent a universal model or credit recipe

The official PowerPoint guidance inspected on 4 October 2026 used a typical task “using GPT-5.5” to illustrate consumption of 10–50 credits per message. That example does not establish a mandatory model, guarantee that every tenant exposes a model selector, or provide a fixed cost for every edit. Limits can differ by plan, workspace and task, and the visible user interface (UI)The controls and visual surfaces through which a person interacts with software. Open glossary entry can change.

The safe procedure is administrative rather than speculative. Before beginning, ask the workspace owner what the tenant currently exposes, which usage policy applies and whether the team has a practical cap for the assignment. Record the answer as a dated tenant fact. Then reduce avoidable rework: duplicate the deck, define the editable range, request a plan before a large edit and approve that plan before asking for slide changes. This follows OpenAI’s guidance to ask for a plan first when making larger edits, without claiming that planning fixes credit use or output quality.

Worked fictional example: Northstar’s internal guidance says the analyst should use the model and controls exposed by the approved PowerPoint integration and monitor the workspace’s normal usage reporting. The analyst does not search for a purported “board deck model” or promise a cost. Instead, the first substantive request asks for a proposed edit plan for slides 4–9, a list of missing evidence and confirmation that slides 1–3 and 12 are out of scope. This is a suggested method, not a guarantee of any particular credit consumption or result.

Decision rule: follow the tenant’s current, documented options; do not prescribe a model that the official workflow does not require. If usage limits are too restrictive for the packet, split the assignment into approved, reviewable stages or perform the affected work manually. The trade-off is that smaller stages require more human coordination, but they make the scope and changes easier to inspect.

Clear the privacy, classification and residency boundary

A source packet should be treated as potentially sensitive even when every file is already available to the analyst. According to OpenAI’s PowerPoint documentation inspected on 4 October 2026, the integration processes the prompt, presentation content made available to it, attachments and relevant connected context. Access to a file for ordinary work does not automatically mean that the file has been approved for processing through this route.

Training treatment also varies by service and setting. OpenAI’s Data Controls frequently asked questions (FAQ)A collection of recurring questions and concise answers about a subject. Open glossary entry, inspected on 4 October 2026, says content submitted to its business offerings, including ChatGPT Business and Enterprise, is not used to improve model performance by default unless the customer opts in. The same source says content in individual services may be used depending on user settings and notes defined circumstances in which authorised personnel and trusted service providers can access content. A default for a business offering is therefore relevant, but it is not a complete privacy assessment for the organisation, plan, integration and document class involved.

Residency requires a separate check. OpenAI’s residency guidance, updated two days before it was inspected on 4 October 2026, describes data residency and inference residency as eligibility- and configuration-dependent options for specified Enterprise and education customers and regions. Inference residency requires data residency in the same supported region, does not cover every processing step, and external integrations can fall outside the selected region. Newly released features are not residency-eligible unless listed. Do not infer that the PowerPoint workflow is covered merely because the wider workspace has a residency configuration.

Run this approval procedure with the workspace owner, data owner and relevant privacy, security, procurement or legal reviewer:

1. Classify each source and the existing deck under the organisation’s own policy.
2. Identify personal data, commercially sensitive forecasts, legal advice, credentials and other restricted content.
3. Confirm that the exact plan, workspace, add-in route and connected apps are approved for those classifications.
4. Verify current training controls, retention arrangements and contractual controls applicable to the tenant.
5. If location commitments matter, confirm eligibility, configured region, inference scope, integration exceptions and whether this exact feature is in scope.
6. Minimise the packet to the files and fields required for the named slides.
7. Record approval, exclusions and unresolved conditions before processing begins.

Worked fictional example: Northstar’s packet contains customer-level revenue rows, although the board slides require only approved regional totals. The data owner supplies a minimised table containing region, quarter, revenue and approved variance commentary. Customer names and account identifiers are excluded. A legal memorandum is not attached; the company secretary provides approved wording for the risk slide instead. This example illustrates data minimisation, not a claim that aggregation by itself satisfies every privacy or confidentiality obligation.

Decision rule: if the organisation cannot verify that the exact processing path is approved for the packet’s highest classification, do not use that path. Redact or aggregate only with the data owner’s approval, or keep the work manual. The trade-off is between richer context and lower disclosure: extra detail may help drafting, but it also increases the material processed and the burden of checking whether that detail was necessary.

Create a safe working copy and a review record

Editable structure and preservation instructions are goals, not guarantees. OpenAI notes that advanced chart, shape, formatting, slide-management and template-matching work can require manual refinement. Its developer guidance says to review the final slides, not only the outline. Before any large change, duplicate the deck and establish a record that lets reviewers compare the proposal with the approved starting point.

Use the following preparation sequence:

1. Save the approved original as read-only under the organisation’s normal records process.
2. Create a clearly named working copy; do not overwrite the approved original.
3. Export or capture a reference view of every slide so reviewers can inspect visual changes.
4. List immutable slides and protected elements such as approved disclosures, risk ratings, logos or footer wording.
5. Create a source register with file name, owner, version, approval date, permitted purpose and slides supported.
6. Create an empty claim register with fields for slide, proposed claim, source location, citation where available, reviewer and status.
7. Create an empty change log with fields for slide, requested change, proposed change, reason, source and reviewer decision.

Worked fictional example: Northstar creates Northstar_Q2_Board_v7_AI-working-copy.pptx while retaining Northstar_Q2_Board_v7_approved-baseline.pptx. Its source register says that the signed accounts support slides 4–6, the operations note supports slides 7–8 and the approved risk extract supports slides 10–11. Slide 12 has no supporting source because it is immutable and outside the requested edit range. This does not prove that the working copy will remain visually identical; it gives reviewers a baseline against which to detect divergence.

Decision rule: do not begin substantive edits without a recoverable baseline and named reviewers. If preserving a complex chart, master or layout is more important than revising it automatically, leave that element unchanged and prepare a human-authored amendment. The trade-off is automation breadth versus revertability and fidelity. For consequential board decisions, human owners must verify every material claim and number, check citations where available against the underlying source, inspect the rendered slides, and approve the final deck before sharing.

For prompt-led slide-deck work in ChatGPT Work, a separate product surface from editing the open PowerPoint presentation, 12 Prompts for Document Workflows, Slide Decks is a prompt playbook that includes slide-deck and document workflows. Treat it as separate site guidance, and retain this tutorial’s approved source packet, recoverable deck baseline and rendered-slide review when using the PowerPoint add-in.

Build a board-safe input contract before asking for slide edits

The source packet is not merely a bundle of files. It is a controlled input contract: a concise statement of what the board needs, which evidence is approved, what may change, what must remain untouched and how uncertainty must be represented. This distinction matters because a plausible narrative is not necessarily a supportable one. The reader’s job is to constrain the drafting task before the PowerPoint add-in sees the material, then preserve enough provenance for finance, company secretariat, legal, compliance or other accountable reviewers to verify the result.

OpenAI’s PowerPoint guidance, inspected on 4 October 2026, recommends specifying where to edit and what to preserve, and asking for a plan before larger changes. Its developer guidance also says to identify the audience, message and material that must remain unchanged. Treat those instructions as controls rather than prompt-writing niceties. The practical decision rule is: if a fact, period, definition or permission affects how a board member could interpret a slide, encode it in the input contract rather than expecting the model to infer it.

This is a documented workflow proposal, not a hands-on test or a guarantee about every tenant or deck. Use the PowerPoint-native add-in for work inside the open presentation; the separate Work desktop surface is distinguished in the section “Choose the PowerPoint-native surface, not a similarly named workflow”.

1. Write the decision brief before collecting evidence

A board update can inform, request a decision or escalate a risk. Those purposes require different slide structures. An information update might lead with performance and variance; a decision paper should lead with the decision, alternatives and consequences; a risk escalation should distinguish exposure, controls and requested intervention. Do not start by asking AI to “improve the deck”, because that leaves the governing purpose undefined.

Create a short decision brief containing six fields:

• Audience: the board or committee receiving the update, including relevant knowledge assumptions.
• Decision sought: the exact approval, endorsement, challenge or acknowledgement requested.
• Reporting period: both the start and end date, plus the comparison period.
• Materiality focus: the matters that deserve board attention rather than operational detail.
• Out-of-scope matters: topics that the update must not cover or speculate about.
• Accountable approver: the person who must sign off the facts and final recommendation.

Worked fictional example:

Audience: Audit and Risk Committee, assumed to understand the approved risk taxonomy but not current programme detail.

Decision sought: Endorse a revised remediation deadline of 30 June 2027 and request monthly exception reporting until closure.

Reporting period: 1 July to 30 September 2026; compare with the quarter ending 30 June 2026.

Materiality focus: overdue high-severity actions, control-owner capacity and dependencies affecting the proposed date.

Out of scope: legal conclusions, regulatory reaction and financial exposure not contained in the approved papers.

Accountable approver: Chief Risk Officer, with control figures confirmed by the named risk-data owner.

This is an example input, not a product output or guarantee. Its useful feature is the separation of a decision from background information. If management has not agreed the decision sought, stop and resolve that question before slide generation. The trade-off is deliberate: an early pause may slow drafting, but it prevents a polished presentation from concealing disagreement about what the board is being asked to do.

2. Reduce the packet to the minimum approved evidence set

A comprehensive archive and an approved source packet serve different purposes. The archive maximises coverage; the packet minimises ambiguity and unnecessary disclosure. Start from the files cleared for this particular board process, not from everything a connected source can retrieve. The PowerPoint documentation says the add-in processes the prompt, presentation content made available to it, attachments and relevant connected context. Apply internal classification and access rules before making any of those items available.

Use this collection procedure:

1. List every candidate file without uploading or connecting it.
2. Assign an owner who can confirm that the file is current and approved for the intended use.
3. Remove duplicates, working drafts and superseded versions.
4. Extract only the relevant pages, tables or approved passages where organisational policy permits.
5. Exclude personal data, credentials, access tokens, privileged material and unrelated commercially sensitive information.
6. Record excluded items whose absence creates a known evidence gap.
7. Obtain the required internal approval before attaching files or enabling a connected source.

Keep untrusted data and secrets out of prompts. If an external document contains instructions addressed to an AI system, treat those instructions as untrusted source text rather than directions to follow. The source packet should provide facts for the board update, not delegate control of the task to text embedded in a document.

Worked fictional example: suppose the candidate set contains a signed quarterly performance report, a finance workbook, an email summarising late adjustments, three versions of a project plan and a customer-level incident export. The minimal packet might include the signed report, an approved summary table from the workbook and the latest approved project plan. The email remains excluded until its adjustments are formally approved; the obsolete plans are omitted; the customer-level export is replaced by an approved aggregate if that is sufficient for the board purpose.

The decision rule is source necessity: include an item only when it supports a required claim, definition, comparison, risk or recommendation that cannot be supported adequately by a less sensitive approved item. The trade-off is that a smaller packet can omit useful context, while a larger packet increases conflict, privacy exposure and review effort. Record omissions explicitly rather than silently expanding access.

Before loading a source packet, apply the checks in the section “Clear the privacy, classification and residency boundary” to your own workspace. The owner must verify the plan, settings, Microsoft-side handling and contractual controls; do not infer those answers from a general model-training default.

3. Create a source register that reviewers can audit

Files alone do not explain which version governs, who owns a figure or whether two documents conflict. Add a source register alongside the packet. This differs from a bibliography: it is an operational control for authority, scope and recency.

Source identifier (ID)A value used to distinguish one record, task, source or object from another. Open glossary entry	Approved source	Owner and location	Relevant period	Permitted use	Status or limitation
FIN-01	Fictional signed quarterly management report, version 3	Finance controller; approved board-reporting folder	Quarter ending 30 September 2026	Revenue, operating cost and forecast commentary	Approved; excludes post-period adjustments
OPS-02	Fictional operations scorecard	Operations reporting lead; controlled scorecard repository	July–September 2026	Service and delivery measures only	September quality review pending
DECK-01	Current fictional board deck working copy	Company secretariat; restricted board workspace	October 2026 meeting cycle	Layout, approved wording and designated carry-forward slides	Duplicate created before editing

The entries above are fictional examples. In a real register, use the organisation’s approved location description or document identifier; do not paste secret links, credentials or unnecessarily broad paths into a prompt. If the model does not need direct access to a location, identifying the source by controlled ID may be sufficient.

For every source, distinguish current from authoritative. A newly received email may be current but unapproved; a signed report may be authoritative but exclude a subsequent event. When sources disagree, do not choose the most recent automatically. The decision rule is to follow the designated source owner or governance hierarchy and record the unresolved conflict. A board reviewer should be able to see why one figure was used and another was not.

Source-to-draft traceability also matters in legal memoranda, but that is a different audience and review authority from a board update. Source Packets, Preference Memory, Structured Memoranda, Citation Verification describes the legal packet and attorney-review pattern; use it as a comparison, not as evidence that any board claim has been checked.

4. Freeze approved KPI definitions

A KPI label is not a definition. “Retention”, “pipeline”, “availability” and “headcount” can each have multiple valid calculations. A model may produce coherent commentary while combining incompatible definitions, periods or denominators. Prevent that by supplying a compact KPI dictionary approved by the relevant owner.

For each KPI, record:

• the exact display name and approved definition;
• the calculation or aggregation rule;
• the unit, currency and rounding convention;
• the reporting date or period;
• the approved comparator, target and status thresholds;
• the source ID and accountable owner;
• known exclusions, restatements or limitations.

Worked fictional example:

KPI	Approved definition	Period and comparison	Presentation rule	Source
Delivery on time	Milestones completed by their approved baseline date divided by milestones due in the period; cancelled milestones excluded	Quarter ending 30 September 2026; compare with prior quarter	Whole percentage; do not infer a target	OPS-02
Forecast operating cost	Approved full-year forecast on the management-reporting basis	Forecast as at 30 September 2026; compare with approved budget	Millions of pounds sterling, one decimal place; label forecast explicitly	FIN-01

These are example definitions, not recommended universal metrics. Human owners must verify the formula, units and period against the approved source. If a packet supplies a KPI value but no approved definition, classify the definition as unknown and prohibit derived commentary such as “improved efficiency”. No calculation, trend or status should enter the deck unless its definition and comparison basis are approved. The trade-off is less narrative fluency in exchange for defensible interpretation.

5. Maintain an explicit unknowns and conflicts ledger

An unknown is not a blank for the model to complete. It is a governance state. Create a ledger before drafting so gaps are visible and can be handled consistently.

Item	Type	Current treatment	Owner	Required by
September quality result	Unknown	State “pending owner validation”; do not estimate	Operations reporting lead	Before factual sign-off
Forecast cost: report differs from planning note	Conflict	Use neither for recommendation wording until finance resolves authority	Finance controller	Before slide editing
Effect of revised deadline	Unsupported inference	List as a question; do not describe as a benefit	Programme sponsor	Before board circulation

This fictional ledger distinguishes three conditions: missing evidence, contradictory evidence and an inference that no supplied source supports. Each requires different treatment. Missing evidence can be labelled pending; contradictory evidence needs an authority decision; an unsupported inference must be removed or recast as a question.

Do not use zero, “not applicable” or “no change” as interchangeable placeholders. They communicate materially different facts. If the board format cannot accommodate an unresolved field, the accountable owner—not the model—must decide whether to defer the slide, disclose the gap or seek another approved source. Consequential decisions require human review and sign-off.

6. Attach the reference deck with an edit map

Attach the duplicated current deck or an approved reference deck only after the source packet has passed classification and access review. A reference deck supplies structure, style and approved carry-forward content; it does not prove that every existing claim remains current. Mark content by edit status rather than relying on visual inference.

Create an edit map with four categories:

• Immutable: no wording, number, object, order or formatting change is authorised.
• Update in place: preserve the slide’s purpose and layout, but replace specified period data or commentary.
• Rebuild within constraints: redesign is allowed using approved branding and figures, subject to manual review.
• Review only: identify issues but make no edit until a human approves a second instruction.

Example edit map: “Slides 1–2 immutable; slides 3–5 update in place for the quarter ending 30 September 2026; slide 6 review only because legal wording is awaiting approval; slides 7–8 rebuild within the existing theme; appendix slides 9–12 immutable.” Add object-level constraints where needed: “Do not alter the approved risk heat map on slide 5; update only the commentary panel.”

The distinction between “immutable” and “preserve where possible” is important. The former withholds authority to edit; the latter expresses a preference that may still require correction. The decision rule is to mark any approved legal wording, governance statement, historical baseline, board decision or manually controlled chart as immutable unless its owner explicitly authorises a change.

Apply the safe-copy and preservation rule described above to this edit map. Before major changes, duplicate the file; after the bounded edit, compare it with the approved original, inspect charts and shapes and restore any unauthorised change.

7. Define approved figures and branding separately

A reference deck may contain old branding, stale charts or one-off exceptions. Therefore, do not tell the model simply to “match the deck”. Supply an approved visual contract that distinguishes reusable rules from historical artefacts.

Record the approved theme or template, logo asset, typefaces, colour tokens, footer convention, confidentiality marking, chart conventions and any accessibility requirements already adopted by the organisation. Identify which images and figures are approved for reuse and whether cropping, recolouring or redrawing is permitted. Do not ask the model to find substitute logos, photographs or icons on the open web when the task is restricted to supplied material.

Fictional example: “Use the theme embedded in DECK-01. Use only BRAND-01 for the corporate logo. Preserve the footer and classification marking. Tables may use the existing blue and grey styles. Do not redraw the approved regulatory diagram on slide 2. No external imagery.” A human reviewer should then inspect every rendered slide, because template adherence may be imperfect even when the constraint is explicit.

The decision rule is to treat approved assets as a whitelist. If a required asset is absent, flag the gap instead of improvising. The trade-off is a less decorative draft, but it avoids introducing an unapproved mark or visual claim.

For a separate site tutorial on building websites and presentations with ChatGPT Work rather than editing this approved open board deck in PowerPoint, see ChatGPT Work to Build Websites and Presentations. It describes a different creation workflow; verify the relevant surface before reusing any source-packet or asset-handling step.

8. Submit a planning prompt, not an editing instruction

For a substantial update, first ask for a proposed edit plan. OpenAI’s PowerPoint Help guidance explicitly recommends requesting a plan before larger edits, while its developer guidance recommends a short slide plan followed by building and review. The plan creates a checkpoint at which a human can reject unsupported changes before they affect the deck.

The following is an example prompt, not a guarantee of product behaviour:

Work only from the attached approved packet and the content made available in this duplicated presentation. Do not use external facts, general knowledge or assumptions to complete gaps.

Audience: Audit and Risk Committee. Decision sought: endorsement of the proposed remediation deadline stated in the decision brief. Reporting period: 1 July–30 September 2026, compared with the quarter ending 30 June 2026.

Apply the KPI definitions exactly as written in KPI-01. Treat the source register as the authority map. Where sources conflict, do not choose between them: identify the source IDs, values and owner required to resolve the conflict. Where information is unknown, label it as unknown or pending; do not estimate.

Follow the edit map. Slides 1–2 and 9–12 are immutable. Slide 6 is review only. Do not change surrounding slides, slide order, approved branding, footers or classification markings.

Before editing, return a slide-by-slide plan containing: proposed purpose, intended claims, supporting source IDs, figures to be used, unresolved gaps or conflicts, and the exact slides or objects proposed for change. Also list any requested content that the supplied sources do not support. Make no slide edits until a human approves the plan.

Review the plan against the packet, not merely for writing quality. Confirm that every proposed claim has a source ID, every number uses the approved period and definition, and every planned change falls within the edit map. Reject any plan that silently resolves a conflict, introduces an external fact or modifies an immutable area.

The governing rule is that plan approval authorises only the listed changes; it is not approval of the finished slides. This adds a review stage, but it separates scope approval from factual and visual sign-off.

9. Convert the approved plan into a bounded edit request

After the accountable reviewer approves or amends the plan, issue a second instruction limited to the authorised slides. Ask for traceability artefacts alongside the editable revision: a claim register, unsupported-claim list and change log. Request citations where available, but do not assume that they will be complete or correct.

Example continuation:

Proceed only with the approved plan version dated 8 October 2026. Edit slides 3–5, 7 and 8 only. Preserve all other slides and the object-level restrictions in the edit map.

Use only supplied facts. Keep unknowns visibly labelled and do not convert questions into assertions. For each edited slide, provide a review record with: claim text, source ID and page or table reference where available; number, unit, period and KPI definition; any citation available; and a concise description of the change.

Provide a separate unsupported-claim list covering statements requested by the brief but not supported by the packet. Do not hide those items in speaker notes or replace them with plausible wording.

A useful sample change-log entry might read: “Slide 4, commentary panel: replaced prior-quarter narrative with wording based on OPS-02, table 3; no change to the approved heat map.” This is an example format, not evidence that the add-in will always generate complete provenance. Human reviewers must compare the actual deck, sources and log.

OpenAI’s official guidance requires review of important claims, numbers, citations where available and slide changes. Review the rendered slides as well as the outline or text: a correct number can still be paired with the wrong label, clipped, placed in an unauthorised object or made visually misleading. The final decision rule is: do not circulate the board update until the relevant source owners have verified their figures, the accountable executive has approved the decision framing, and designated reviewers have checked citations where available, unknowns, conflicts, visual presentation and every recorded change.

Turn the approved plan into an enforceable change boundary

A planning response is not permission to edit the presentation. Treat it as a proposed change specification that a human must accept, reject or amend before the PowerPoint add-in changes any slide. This separates editorial judgement from execution: the plan identifies intended revisions and evidence, while the subsequent instruction authorises only a defined subset of those revisions.

OpenAI’s PowerPoint guidance, as inspected on 4 October 2026, recommends asking for a plan before larger edits and being specific about what to change, where to edit and what to preserve. The OpenAI Developers slide-deck guidance also says to specify the audience, message and material that must remain unchanged. These are useful controls, but requested preservation remains a goal rather than a guarantee. Template matching, advanced charts, shapes, formatting and slide-management work can require manual refinement.

Approve proposals at slide-and-claim level

Do not approve a plan with a single “looks good” response. Review each proposed slide against five fields: intended decision, proposed change, source location, protected content and unresolved question. A reviewer should be able to approve one slide without implicitly approving every other suggestion.

1. Compare the proposal with the edit map. Confirm that every slide named for revision was classified as editable or conditionally editable.
2. Check the rationale. The proposed change should serve the agreed board decision, not merely make the slide shorter or more polished.
3. Trace each factual addition. Match the proposed claim or figure to a source-register entry and a precise location such as a worksheet tab and cell range, document section or approved note.
4. Confirm preservation requirements. Record the elements that must survive the edit, including approved figures, definitions, footnotes, chart units, dates and any wording already signed off.
5. Resolve or retain uncertainty. An unanswered question must remain visible in the approval record; it must not be silently converted into a confident statement.
6. Record a disposition. Mark each proposal “approved”, “approved with amendment”, “deferred” or “rejected”, with the accountable reviewer’s name or role.

Worked fictional example: suppose the plan proposes replacing slide 7’s operational commentary with “Service performance recovered in September”. The source register points only to a quarterly average, while the unknowns ledger notes that September data has not been approved. The reviewer should reject that sentence rather than approve it as a stylistic simplification. An acceptable amended proposal might be: “Retain the approved quarterly average and add an unresolved question asking the operations owner whether September data may be used.” This example illustrates a control method; it is not a product guarantee.

Decision rule: authorise an edit only when its slide, purpose, evidence boundary and protected elements are all explicit. If a proposal introduces a new period, denominator, forecast, causal explanation or recommendation that is not supported by the approved packet, defer it. The trade-off is slower approval, but it prevents editorial fluency from obscuring an evidence gap.

Require a complete plan response before execution

A useful plan should cover more than the slides it wants to alter. It should also identify unchanged slides and explain why they remain untouched. That negative scope is important: without it, an instruction to “improve the board update” can be interpreted much more broadly than intended.

Ask the planning response to use a review table with these fields:

• slide number and current title;
• status: revise, preserve or review manually;
• proposed change, without applying it;
• board purpose or decision supported;
• source-register identifier and exact source location;
• claims and numbers that would be added, removed or reframed;
• figures, definitions, notes and layout elements to preserve;
• citations where available;
• unsupported claims or unresolved questions;
• anticipated manual work, such as chart or template repair.

The phrase “citations where available” matters. Official guidance requires human review of important claims, numbers, citations where available and slide changes; it does not promise that citations will be generated automatically, completely or correctly. A source location in the plan is therefore a review aid, not proof that the resulting slide is accurate.

Example planning instruction:

Do not edit the presentation. Return a proposed slide-by-slide plan only. For every slide, state whether it should be revised, preserved unchanged or sent for manual review. For proposed revisions, give the rationale, exact source-register identifiers and locations, affected claims and figures, protected content, citations where available, and unresolved questions. Identify any proposal that depends on unsupported interpretation. Slides 1, 2, 10 and 12 are immutable. Preserve the approved revenue, cash and headcount figures exactly as supplied. Do not infer missing values or convert ranges into point estimates.

Decision rule: reject an incomplete plan if it omits unchanged slides, source locations or unresolved questions. A compact plan can save review time, but compression must not erase the evidence needed to decide whether an edit is authorised.

Issue an execution prompt that permits only approved revisions

The execution prompt should behave like a narrowly scoped change order. It should refer to the approved plan, list the authorised slides and restate the non-negotiable constraints. Do not rely on conversational context alone: a reviewer should be able to read the execution instruction independently and understand exactly what may change.

Translate approval decisions into explicit operations

Prepare a short execution schedule before submitting the instruction:

Slide	Authorised operation	Protected elements	Evidence limit	Manual review owner
4	Tighten the headline and reduce body copy	Approved KPI values, period labels and footnote	Source SR-03, worksheet “Board KPI”, approved cells only	Finance owner
7	Add one risk row	Existing risk scores and colour meanings	Source SR-08, section 2.3; no inferred likelihood	Risk owner
9	Reorder existing next steps	Named owners and approved dates	Existing slide content only	Company secretary

This is a fictional schedule illustrating the method, not a claim about an actual company or output. Notice that each row defines an operation rather than a broad aspiration. “Reduce body copy” allows wording changes; it does not authorise changing a KPI, replacing the chart or inventing an interpretation.

A bounded execution prompt can then read:

Apply only the approved revisions listed below to the open working copy. Revise slides 4, 7 and 9 as specified. Preserve all other slides and do not change their text, order, notes, charts, shapes or formatting. On slide 4, tighten the headline and body copy without changing approved figures, period labels, units, definitions or the footnote. On slide 7, add only the approved risk from source SR-08 section 2.3; do not infer probability, financial exposure or mitigation status. On slide 9, reorder the existing approved actions without changing owners or dates. If an instruction conflicts with the deck or source packet, stop that edit and report the conflict rather than resolving it by assumption. Return a change log and an unsupported-claim list after the edits.

Decision rule: if an approved proposal cannot be expressed as a specific operation with a source boundary and preservation clause, return it for clarification. Broader discretion may produce smoother narrative flow, but it also increases the chance of unintended changes and makes human review harder.

Protect approved figures as structured objects, not just numbers

A figure is more than its displayed value. Its period, currency, unit, scale, denominator, status and definition can change its meaning. Protect the complete figure specification in the execution prompt.

For example, “12.4” is not sufficiently bounded. A safer fictional specification is: “£12.4 million recognised revenue for the quarter ended 30 September 2026, actual, using the approved finance definition in KPI-REV-01.” If the source packet contains both actual and forecast values, identify which may appear. If a chart uses thousands while commentary uses millions, preserve or explicitly authorise the conversion and have the finance owner verify it.

Ask the add-in not to:

• recalculate totals unless that calculation is explicitly approved;
• change actuals into forecasts or vice versa;
• round values differently without permission;
• alter signs, currencies, units, periods or denominators;
• derive percentages from incomplete inputs;
• replace an approved range with a single value;
• create causal language from a correlation or timing sequence.

Decision rule: preserve a figure verbatim when it has already been approved for board use. Authorise recalculation only when the formula, inputs and accountable verifier are specified. The trade-off is that a preserved number may require manual layout adjustment, but presentation fit is not a sufficient reason to alter financial meaning.

Make “unchanged” independently reviewable

An instruction to preserve surrounding slides does not prove that they remained unchanged. Record the expected unchanged set before execution and compare it with the working copy afterwards. This can be a simple slide-number list accompanied by a saved duplicate of the pre-edit deck.

For a 12-slide deck where slides 4, 7 and 9 are authorised, record slides 1–3, 5–6, 8 and 10–12 as expected unchanged. After editing, inspect those slides for text, object position, chart data, speaker notes, footers and sequence. Advanced objects or formatting may need manual checking because official guidance notes limits around complex charts, shapes, formatting, slide management and template matching.

Decision rule: if any unauthorised slide changed, do not accept the deck merely because the change appears beneficial. Revert it or route it through the same approval process. Accepting convenient scope drift weakens the change record and makes later accountability ambiguous.

A board packet may contain restricted figures or personal data, so its owner must verify the relevant account and workspace data controls before upload. ChatGPT Privacy Center Guide surveys memory, temporary chat, connected apps and model-improvement controls; do not infer a particular plan’s eligibility or privacy guarantee from that overview.

Use focused follow-ups instead of reopening the whole deck

Once the bounded revisions are complete, follow-up prompts should address one editorial problem at a time. A blanket request such as “make the whole deck more executive” can reopen approved claims, disturb preserved slides and make the resulting change set difficult to audit. Focused follow-ups retain a clear relationship between instruction, affected slides and reviewer.

Build or amend a risk slide without manufacturing certainty

A risk slide should distinguish sourced facts from management judgement. Ask for only the fields present in the approved packet. If impact, likelihood, exposure or mitigation status is absent, label it unresolved rather than inviting the system to complete a conventional risk matrix.

Example follow-up:

Revise slide 7 only. Use the approved entries in SR-08 to present risk, evidence, current mitigation, accountable owner and unresolved decision. Preserve the existing scoring scale and all approved scores. Where probability or financial exposure is not supplied, write “Not provided in approved source packet” in the review notes rather than estimating it. Do not change any other slide. Return the source location for each risk statement and list any wording that remains interpretive.

A human risk owner must verify that labels, scores and mitigations reflect the organisation’s approved assessment. For consequential decisions, the board pack owner should also obtain any required finance, legal or compliance review.

Decision rule: use a structured risk field only when its value is supported or explicitly supplied as management judgement. Leave it blank or flag it when evidence is missing. A visually complete matrix may look more polished, but completeness achieved through inference is less reviewable than an explicit gap.

Tighten executive wording without changing the claim

Executive tightening should reduce verbal load while preserving factual scope, qualification and decision relevance. It must not turn “may”, “subject to approval” or “within the reported range” into an unqualified assertion.

Use a three-part procedure:

1. identify the exact slide and text boxes that may change;
2. state the propositions, qualifiers and numbers that must survive;
3. request a before-and-proposed-after table before authorising the wording change if the statement is consequential.

Worked fictional example: the source-supported sentence is, “The programme may require up to £2 million of additional funding, subject to completion of procurement review.” An unsafe tightening would be, “Approve £2 million additional funding”, because it changes uncertainty and decision status. A bounded alternative for consideration is, “Potential additional funding: up to £2 million, subject to procurement review.” A human owner must check that the shortened form retains the intended meaning.

Decision rule: reject any shorter wording that changes modality, time period, attribution, range, dependency or approval status. Concision is valuable only when semantic fidelity is preserved.

Adapt for another audience by creating a variant, not overwriting the board version

Audience adaptation differs from factual revision. A board may need decisions, material risks and exceptions, while an operating audience may need actions, dependencies and owners. Create a separate working copy or clearly designated variant rather than overwriting the approved board deck.

Example follow-up:

Create a proposed operating-committee variant from the approved board version. Do not edit the approved board file. Use only existing approved facts. Propose changes to slides 4, 7 and 9 that emphasise actions, dependencies and owners; preserve all figures, dates, definitions and source references. Do not add implementation detail unless it appears in the approved packet. Return a plan first, including unchanged slides and unresolved audience questions.

Decision rule: create a variant when the audience’s decision rights or required detail differ. Use a wording-only follow-up in the same deck only when the audience, purpose and evidence boundary remain unchanged. Variants require more file control, but they avoid compromising an already approved board artefact.

Stop, revert or escalate when the evidence boundary breaks

The safest response to ambiguity is not always another drafting prompt. Stop the edit when the source packet conflicts with the deck, a requested claim has no approved source, an immutable object changes, or the required operation exceeds what can be checked reliably in the add-in.

Use a simple exception procedure:

1. save the current working copy without overwriting the baseline;
2. record the affected slide and attempted operation;
3. quote the conflicting sources or preservation requirement;
4. classify the issue as evidence, approval, privacy, access or rendering;
5. assign a human owner and prohibit further editing of that item;
6. resume only after the source register, approval record or instruction has been corrected.

For example, if the deck says headcount is 480 and the approved spreadsheet says 476, do not ask the system to choose the “most likely” value. Record the discrepancy, preserve the currently approved presentation value until the accountable owner decides otherwise, and route the conflict to the designated KPI owner.

Keep untrusted data and secrets out of prompts. Do not paste credentials, private access tokens, irrelevant personal data or material that has not been cleared for the workspace and add-in processing. As documented by OpenAI and inspected on 4 October 2026, the add-in processes the prompt, made-available presentation content, attachments and relevant connected context. Business-offering content is not used to improve model performance by default unless the customer opts in, but that does not settle every privacy, retention, integration or residency question. Consumer handling depends on settings, and residency is eligibility- and configuration-dependent rather than a blanket promise for this workflow.

Decision rule: escalate rather than improvise whenever resolving an issue would require a new fact, a changed approval, access to uncleared material or an unsupported privacy assumption. Manual intervention costs time, but it preserves the distinction between authorised evidence and plausible-sounding completion.

This procedure is a documented workflow proposal based on OpenAI’s official materials as of 4 October 2026, not hands-on testing. Behaviour can differ with plan limits, workspace controls, RBAC, administrator approval, add-in deployment, connected-source permissions, PowerPoint environment and template complexity. A named human remains accountable for checking the final slides, not only the plan, and for signing off claims, numbers, citations where available, changes and consequential decisions.

Run the release review against the source packet and the rendered deck

The release review is not another drafting pass. Its purpose is to decide whether the editable presentation is safe to place before the board. Compare three artefacts side by side: the approved edit plan, the original working copy and the proposed final deck. Keep the approved source packet, source register, conflicts ledger and change record open as separate evidence. OpenAI’s official slide-deck guidance, inspected on 4 October 2026, explicitly advises reviewing the final slides rather than only the outline. The ChatGPT for PowerPoint guidance also requires review of important claims, numbers, citations where available, and slide changes.

Use the PowerPoint add-in for questions about the open presentation, but do not delegate release authority to it. The add-in can help inspect, understand and revise editable slides; it does not independently certify the organisation’s facts or determine whether a disclosure is appropriate. The accountable board-content owner remains responsible for approval, with finance, legal, compliance or other specialist reviewers involved where the subject requires them.

Decision rule: release only when the final rendered slides conform to the approved plan, every material claim has been checked against an authorised source, preservation requirements have been inspected, and all unresolved items are either closed or explicitly accepted by the accountable owner. A plausible-looking deck is not sufficient evidence.

1. Preserve a revert point before any substantial correction

Duplicate the original presentation before making large changes, even if an earlier working copy already exists. Give the duplicate a controlled name that distinguishes the source version, review candidate and approved release; follow the organisation’s records policy rather than placing confidential details in a filename. Restrict corrections to the review candidate and retain the untouched original until release is complete.

1. Confirm that the original opens and contains the expected slide count, masters and embedded objects.
2. Create a duplicate in the approved storage location.
3. Record the original filename or document identifier, duplicate identifier, creation time and responsible editor in the change record.
4. Make substantial fixes only in the duplicate.
5. If a correction damages an immutable slide, master, chart or layout, stop and restore from the duplicate rather than trying to reconstruct it from memory.

Worked fictional example: suppose “Board_Update_Q3_Working” contains approved governance slides that must remain untouched. The editor creates “Board_Update_Q3_ReleaseReview” and records that slides 1, 2, 11 and 12 are preservation-controlled. If a global formatting operation changes slide 12, the reviewer copies the verified slide from the untouched original or reverts the candidate. This example illustrates a control; it is not a guarantee that every element will copy perfectly.

Trade-off: retaining versions adds storage and administration, but it provides a practical route back when template matching or advanced chart, shape, formatting and slide-management edits require manual repair. OpenAI’s official guidance warns that such edits may need refinement, so revertability should take priority over a tidier folder.

2. Reconcile the approved plan with actual slide operations

Start with scope rather than prose quality. Convert the approved plan into a release matrix with one row per slide. Record the permitted operation, the observed operation, preservation constraints, evidence status and disposition. “No visible problem” is not a disposition: use controlled terms such as unchanged and verified, changed as approved, changed outside plan, unable to verify or not applicable.

Slide	Approved operation	Observed state	Evidence and preservation checks	Disposition
1	No change	Compare with original	Title, date, logo, master and speaker notes	Reviewer records result
4	Replace KPI summary	Compare with plan and source register	Definitions, period, units, figures and chart labels	Reviewer records result
7	Add decision request	Compare wording with approved proposal	Decision, evidence, risks, owner and timing	Reviewer records result

This table is a suggested example, not an automatic product output. A reviewer must populate it from the actual files. When the planned operation says “revise slide 4 only”, inspect neighbouring slides anyway: an accidental theme, master or layout change may affect slides that were not meant to be edited.

Decision rule: any material operation that is absent from the approved plan is a release blocker until the owner approves it or the editor reverses it. Minor mechanical differences, such as a harmless line wrap, may be accepted only after checking that they do not alter emphasis, readability or preservation constraints.

3. Check every material claim at atomic level

Do not verify a slide merely because its general message appears in the packet. Break it into atomic claims: one subject, measure, period, comparison and qualification per check. Include slide titles when they make assertions, as well as chart annotations, footnotes, speaker notes and labels such as “on track”, “improved” or “largest”. These words can be material even when no number accompanies them.

1. Copy each material claim into the review record.
2. Identify the exact packet item that supports it.
3. Locate the relevant page, table, cell, paragraph or approved note.
4. Compare the slide’s subject, period, scope and qualification with the source.
5. Mark the claim supported, partly supported, conflicted, unsupported or not independently checked.
6. Route consequential interpretations to the named subject owner.

Worked fictional example: a draft title says, “Customer retention improved across all regions.” The packet contains an approved table showing improvement in North and West, no change in Central and no authorised figure for South. The reviewer must not treat the title as supported because the aggregate direction is positive. The appropriate choices are to narrow the wording to the supported regions, add the missing qualification, obtain an authorised South figure, or remove the claim. The AI system may help identify the over-broad wording, but a human must compare it with the source.

Decision rule: if deleting a qualification, changing the period or broadening the population would make a statement more favourable, treat that difference as material. Do not release it until corrected and rechecked.

4. Recalculate and reconcile every material number

Numbers need a separate pass because a sentence can remain grammatically correct while its value, unit or denominator is wrong. Check visible figures, chart data, axis labels, percentages, totals, dates, currency symbols, decimal precision and calculations embedded in narrative text. If a slide presents a derived number, preserve both its authorised inputs and its formula in the review record.

1. Trace the displayed number to an approved source-register entry.
2. Confirm the reporting period and cut-off date.
3. Confirm the unit, currency, scale and sign convention.
4. Check whether the figure is actual, forecast, target, estimate or illustrative.
5. Reperform permitted arithmetic independently.
6. Compare chart labels and visual proportions with the underlying values.
7. Obtain approval from the responsible metric owner where the figure could affect a board decision.

Worked fictional example: assume an authorised source lists revenue as £48 million for the current period and £40 million for the comparison period. A draft slide describes the movement as 20%. A human reviewer can reproduce the example calculation: (48 − 40) ÷ 40 = 0.20, or 20%. The reviewer must still verify that both amounts use the same accounting basis, currency and period. This sample demonstrates the checking method; it does not validate any real company figure.

Also inspect totals that appear only through a chart. A stacked column may display components that do not sum to the stated total because of differing definitions, rounding or omitted categories. Do not “repair” such a discrepancy by silently changing a number. Record it as a conflict and send it to the metric owner.

Decision rule: a material number without an authorised source, reproducible derivation or accountable owner confirmation is unresolved. Remove it, qualify it transparently or hold the release; do not infer a missing value from nearby figures.

5. Inspect citations where available without treating them as proof

Citations and source labels improve traceability, but their presence does not establish correctness or completeness. Official OpenAI guidance asks users to review citations where available; it does not promise that the PowerPoint add-in will generate exhaustive or verified references. Inspect each citation against the actual packet item and record uncited material claims separately.

1. Confirm that the cited item is part of the approved, access-cleared packet.
2. Open the cited file or record rather than trusting a filename shown on the slide.
3. Verify that the cited location supports the precise claim and period.
4. Check whether the citation points to a primary source or merely to another summary.
5. Confirm that the board may receive the citation or source label under internal handling rules.
6. List material claims with no usable citation under “citations unavailable”, not “verified”.

Worked fictional example: a footnote says “Source: Monthly Operations Pack”. Three monthly packs exist, but only the September version contains the displayed measure. The reviewer should identify the authorised September item and a stable internal locator if policy permits. If the locator cannot appear in the board deck, retain it in the internal review record instead. Do not manufacture a public link or expose a restricted repository address.

Trade-off: detailed slide-level citations aid reviewers but can make an executive slide crowded or disclose internal paths. Keep the visual citation concise where policy allows, while retaining full provenance in the controlled source register. Human reviewers must decide the appropriate disclosure level.

6. Compare the rendered view with the editable structure

A release candidate has two relevant forms: what the board will see and what later editors can safely amend. Review every slide in presentation or rendered view, then inspect the corresponding editable objects. This distinction matters because text may exist but be clipped, a chart may look correct while containing stale data, or an apparently editable diagram may have become a flattened image.

• Rendered check: inspect clipping, overlap, contrast, spacing, alignment, chart legibility, footer visibility and whether the main point can be read at normal presentation scale.
• Editable check: inspect text boxes, tables, charts, grouped shapes, theme use, masters, notes and object order. Confirm that expected editable structures remain editable.
• Cross-check: reopen the file in the intended Microsoft PowerPoint environment and move through the deck in sequence. Do not rely solely on a thumbnail or exported preview.

Worked fictional example: a risk table looks complete in edit view, but its final row is clipped in presentation view. The data exists, yet the rendered board slide omits it. Conversely, a chart may render cleanly but have been replaced by a picture, defeating the requirement for an editable chart. Either condition requires correction or explicit owner acceptance.

Decision rule: where editable structure and visual fidelity conflict, apply the approved requirement rather than assuming one always wins. For a board release, unreadable content is a blocker; for a reusable reporting deck, loss of required editability may also be a blocker. Record any accepted compromise.

7. Test sequence, decision logic and readability as a complete deck

Read the deck from beginning to end without consulting the planning notes. Slide-by-slide accuracy does not guarantee a coherent board update. Check whether the sequence distinguishes status, evidence, variance, risk, decision and next action. Confirm that a later slide does not contradict or silently redefine an earlier one.

1. Read only the titles in order and write down the story they imply.
2. Review the full slides and note where evidence does not support the title.
3. Check that each decision request names the decision, relevant evidence, material risks, accountable owner and timing where approved sources provide them.
4. Inspect transitions between sections for unexplained period, scope or terminology changes.
5. Check that appendices do not introduce new decision-critical facts absent from the main narrative.
6. Have a reviewer who did not draft the slides perform a comprehension pass.

Worked fictional example: slide 5 reports delivery as “on plan”, while slide 8 asks the board to approve recovery funding. Both might be individually sourced, but their relationship is unclear. The release reviewer should determine whether they concern different programmes, periods or definitions and make that distinction explicit. The solution is not necessarily to delete one; it may be to clarify scope and sequence using authorised facts.

Decision rule: if a reasonable reviewer could reach a materially different conclusion because of slide order, an undefined term or a hidden qualification, revise the narrative or escalate it. Readability is not merely cosmetic when it changes decision meaning.

8. Recheck all preservation constraints explicitly

Do not infer preservation from a slide looking similar. Compare designated immutable slides and protected elements against the original at object and content level where practical. Typical constraints include wording, slide order, master, logo placement, legal text, approved charts, page numbering, accessibility information and speaker notes.

A suggested procedure is to inspect each protected slide side by side, then compare its title, body text, objects, notes, layout assignment and position. For protected objects on otherwise editable slides, verify the object itself as well as its placement relative to revised content. Template adherence is a goal rather than a guarantee, so visual inspection remains necessary.

Worked fictional example: slide 2 is marked “unchanged”, and its visible content matches the original. The reviewer discovers that its approved speaker note has disappeared. Because the preservation instruction included notes, the slide is not unchanged. Restore the note from the original, log the correction and repeat the comparison.

Decision rule: mark a slide unchanged only after checking all elements named by the preservation instruction. If no reliable comparison is possible, use “unable to verify”, not “unchanged”.

9. Ask for a bounded inspection, then verify the response yourself

The add-in may assist with a final discrepancy search, provided the prompt contains no new secrets or unapproved data. Keep sensitive material out of prompts unless it has been explicitly approved for processing. Request observations and locations rather than a blanket assurance.

Example inspection prompt: “Review the open release candidate against the approved slide plan and the made-available source register. Do not edit. List, by slide number: operations outside the plan; material claims lacking support; numbers whose period, unit or source cannot be reconciled; citations where available that need human checking; apparent changes to slides marked unchanged; and possible clipping, overlap or readability issues. State ‘not established’ rather than guessing. Do not introduce external facts.”

Treat the response as a candidate issue list. A human must open each referenced slide and source, decide whether the issue is real, and record the disposition. Absence from the response does not prove absence from the deck.

Decision rule: use AI-assisted inspection to widen attention, not to replace the release checklist. Consequential factual, financial, legal, compliance and governance decisions always require accountable human review.

An investor pitch deck has a different audience and approval route from a board update. For founder-oriented pitch-deck, financial-model and investor-communication prompts, ChatGPT Prompts for Startup Founders is a distinct drafting resource, not a substitute for a human check of this board deck’s claims, figures and permissions.

10. Close the change log and obtain accountable approval

The final change record should show what changed, what remained unchanged, what was corrected during release review and what remains unresolved. Give every open item an owner, consequence, due point and release disposition. Avoid vague entries such as “checked” or “looks fine”.

Item	Record	Required human evidence
Changed slides	Slide number, approved operation and final operation	Editor and reviewer confirmation
Unchanged slides	Slide number and preservation elements checked	Comparison with the retained original
Claims and numbers	Source-register identifier and status	Relevant content or metric owner review
Corrections	Issue, action and recheck result	Named reviewer
Unresolved items	Uncertainty, consequence and proposed treatment	Explicit owner decision

Worked fictional example: an open item states, “Slide 6 forecast uses a range not yet approved by the finance owner.” Acceptable release dispositions might be “hold release”, “remove the range”, or “retain with an expressly approved qualification”. “ChatGPT supplied it” is not an acceptable source or approval.

Send the review candidate, source register, release matrix and unresolved-items list to the accountable board-content owner. Ask for an explicit decision: approve, approve subject to named corrections, or reject for rework. Where applicable, obtain separate confirmations from finance, legal, compliance or the relevant KPI owner before the final approval. The accountable owner should not approve by silence.

Decision rule: only the accountable board-content owner may authorise release, and only after required specialist reviews are complete. AI output, an editor’s completion note or a clean visual render cannot substitute for that approval.

11. Save and share only the approved release artefact

After approval, apply only the authorised corrections, repeat the affected claim, number, layout and preservation checks, and save a distinct release version in the approved location. Do not overwrite the original or confuse the pre-approval candidate with the approved deck. Confirm that the shared file is the same version the owner approved.

1. Check the final filename or controlled document identifier.
2. Confirm the approval record refers to that exact version.
3. Open the saved file and inspect the corrected slides once more.
4. Verify sharing permissions and recipients under internal policy.
5. Remove temporary duplicates only when retention and records rules permit.
6. Retain the source register, change record and unresolved-item decisions with the release evidence where policy requires.

Before release, confirm that the actual workspace, destination and final deck still satisfy the privacy, classification and residency gate above. A change in audience or material requires renewed approval; do not treat a plan-wide training statement as a blanket permission to distribute a confidential deck.

Confirm that the actual workspace exposes the PowerPoint-native add-in and the permitted connections. The Business general-availability date and the separate Work desktop distinction are explained above; neither establishes this user’s present-day entitlement or permission to share the deck.

Final release rule: if the exact approved file, factual evidence, preservation result, permissions or accountable sign-off cannot be established, do not share the deck. Return to the last verified duplicate, resolve the issue and repeat only the affected review steps plus the final whole-deck sequence check.

Useful Links

• ChatGPT for PowerPoint
• ChatGPT Business release notes
• Create or revise a slide deck
• Creating and editing documents, spreadsheets, and presentations with ChatGPT Work
• Presentations plugin for ChatGPT
• Data Usage for Consumer Services FAQ
• Data residency and inference residency for ChatGPT