AI Presentation Governance: A Practical Framework for Enterprise Teams

AI presentation governance is the set of rules, roles and technical controls that decide what an AI system may read, retrieve, generate, change and release inside your slide production. It answers five questions: what data may enter, what knowledge the system may reach, what it may write, what artifact comes out, and who signs.

That definition sounds obvious. Almost no company I speak to has written it down.

Most enterprises now have an AI usage policy. Very few have one that survives contact with a board deck. I want to explain why presentations need their own governance layer, and then give you the framework we use when we walk an enterprise team through this.

Why presentations break a general AI policy

A generic policy says something like "do not enter confidential information into AI tools." Reasonable on paper. Useless in practice for presentations, because building confidential material is the entire job.

Here is the structural problem. Presentations concentrate information that every other system keeps apart. One quarterly board deck can carry the financial forecast from the reporting system, headcount detail from HR, a customer escalation from the CRM, an open incident from security, and a management judgement that exists nowhere else. Each of those sources has its own access model. The deck flattens all of them into a single file that then travels by email, gets exported to PDF, and lands on three laptops that were never in scope for any of it.

A general policy governs tools. Presentation governance has to govern a production process. Those are different objects.

The five control points

Every AI presentation workflow, from a simple template filler to a full agent, has the same five control points. If you can describe yours at each layer, you have a governable process. If you cannot, you have a demo.

Notice that only two of the five are technical. Governance fails far more often at input and release than at generation. Teams spend months evaluating model quality and then hand the finished deck to whoever is still online at seven in the evening.

Step one: classify before you configure

Start with a classification model small enough that people actually use it. Four levels is almost always enough.

1. Public. Approved for external distribution without further review.
2. Internal. Ordinary internal business information.
3. Confidential. Client, financial, strategic or employee information with limited access.
4. Restricted. Subject to special legal, contractual, transaction, health, security or market abuse controls.

Then write the one line that most policies are missing: which levels each tool may process, and under what conditions. Not "be careful." A specific mapping.

If your answer for a given tool is "levels 1 and 2 only," then you have bought a tool your teams cannot use for the work they actually do, and someone will paste a level 3 document into it within the month. Better to know that now.

Step two: define the source of truth for every material claim

A governed workflow knows where a number came from. Approved workbook, reporting system, controlled document, or somebody typing it into a text box. Where it matters, carry the period, the units, the currency, the scenario, the owner and the refresh date alongside the value itself.

For recurring decks, maintain a source map. For generative tasks, require a citation or an evidence entry for anything that is not common knowledge.

And set the rule that matters most: when the source cannot be established, the system flags the gap. A model that quietly fills a gap with a plausible number is doing the single most damaging thing it can do inside a presentation, because the format makes it look authoritative.

Step three: assign owners, not committees

Governance dies in shared ownership. Write the table.

The last row is the one that gets skipped. Every deck that leaves the building needs a person, not a function, who accepted it. If your process cannot name that person after the fact, you do not have an approval step. You have a habit.

Step four: tier the review by consequence

Not every deck deserves the same scrutiny. Tiering is what keeps governance from becoming theatre.

Tier 1, low consequence. Internal status updates, team meetings, working drafts. Author reviews. No formal record.

Tier 2, standard. Client presentations, internal management reporting, proposals. Subject matter owner reviews content, author checks format, one named approver releases.

Tier 3, high consequence. Board and supervisory material, regulated reporting, transaction documents, anything with personal data about identifiable employees or customers, anything a regulator may later read. Independent content review, privacy or legal review where personal or regulated data is involved, documented approval, archived source set.

Write the tier definitions once. Then let the workflow route the deck automatically based on classification and audience, rather than asking a stressed author to self assess at the end.

Step five: decide what the AI is not allowed to do

This is the shortest section in most policies and the most valuable. Be concrete and be specific to slide production.

An AI presentation workflow should not, without explicit human input and verification, create or alter:

• pricing, commercial terms or delivery commitments;
• client names, logos or permission to reference a client;
• staffing, availability or individual qualifications;
• outcome metrics, performance figures or financial results;
• legal wording, disclaimers, regulatory statements and required declarations;
• quoted material attributed to a named person;
• anything presented as a citation that it cannot resolve to a real source.

Everything on that list has one property in common. If it is wrong, the cost lands outside your company, on a client, an employee or a regulator. That is the line worth drawing.

Step six: log the decisions, not just the output

An audit trail is worth its storage cost only if it answers a question somebody will actually ask. Depending on the use case, capture the user and workspace, the task and the selected source set, the file and version identity, the objects generated or modified, validation results and unresolved items, the reviewer and approval status, and the export or release event.

Resist logging every prompt and every piece of content by default. Prompt logs become sensitive records in their own right, and an unbounded one is a data protection problem you built on purpose. Give logs their own purpose, access rules and retention period, exactly as you would for any other data store.

Where EU regulation actually touches this

Two frameworks matter for most European teams, and it is worth being precise about what applies today rather than repeating headlines.

GDPR applies whenever the workflow processes personal data, which in presentations is more often than people assume. A person can be identifiable through a unique role, a deal history, an office and a date, without a name appearing anywhere. The obligations that bite hardest in slide production are purpose limitation, minimisation, accuracy and the processor arrangement with your vendor.

The EU AI Act applies according to the system and your role. Since 2 February 2025 the AI literacy duty under Article 4 has applied to deployers of any AI system, which includes a firm rolling out a slide drafting assistant. Since 2 August 2026 the transparency obligations under Article 50 apply. The obligations for Annex III high risk systems were deferred to 2 December 2027 by Regulation (EU) 2026/1744, the Digital Omnibus on AI, which entered into force on 27 July 2026. Systems embedded in products under Annex I follow on 2 August 2028.

Most slide drafting workflows are not high risk systems. That is a conclusion you should reach by assessing your own use, not by assuming it. And the deferral changed a deadline, not a direction.

The one page policy that works

If your policy is longer than this and people still paste client data into a consumer chat window, the length was never the problem. The approved path was too slow.

What to measure

Governance quality is measurable, and time saved is the wrong first metric. Track:

• unsupported or corrected claims caught in review, and where they came from;
• review cycles per deck, and how many were format rather than substance;
• share of content drawn from approved sources versus pasted from elsewhere;
• access exceptions requested and granted;
• incidents reported, including near misses;
• share of external decks released by the named approver;
• time from request to review ready draft, alongside all of the above.

The pattern to watch for is a workflow that gets faster while review effort climbs. That is not efficiency. That is cost moving to someone with less time.

Where offgen fits

We built offgen so that these controls live in the workflow rather than in a document nobody reads. A Company Brain connects approved masters, templates, slide libraries, CVs, references and business context to both people and agents, under the permissions people already have. Output stays native and editable, so review happens on the real artifact. Brand rules, lockable elements, scoped skills and controlled access are the mechanisms behind the framework above.

Those are inputs to your control design, not a compliance conclusion. You still assess your own data, configuration, legal role, users and purpose. Our security overview, trust center and data processing agreement exist to make that assessment faster.

The test I would apply to any framework, including this one: can your team answer what went in, what the system did, what changed, who checked it and what was released? Five answers. If you have them, you can govern the process. If you do not, no policy document will produce them.