Create Presentations With AI: Script, Slides and Review Without Rework

Create Presentations With AI: Script, Slides and Review Without Rework

To create presentations with AI without rework, start with the message and the script, not the slides. Use AI to propose a structure, organize your arguments and draft each slide, and keep the review of numbers, names and factual claims for the author. If you reverse that order, you get good-looking slides about an idea that is still vague, and you end up redoing everything.

Key takeaways

• Write down in one sentence what the audience should decide or do after the presentation, before you ask AI for anything.
• Ask for the script first (the sequence of arguments), then the slide drafts. Two separate steps need fewer corrections than a single request.
• AI is good at structure, text variations and summarizing. It is not a source of numbers, dates, customer names or quotes.
• Numbers and factual claims are the author's responsibility, checked against the original source and not against the model's memory.
• Internal and customer data need care: use the tool your company allows and remove anything the request does not need.

Why do most AI-made presentations end up needing rework?

The rework rarely comes from the quality of the generated text. It comes from three common habits.

The first is asking for "a presentation about X" without saying who it is for, what it should achieve or what decision you expect. The AI returns something generic, correct in form and empty in substance. The author then edits slide by slide, which takes more effort than writing the script from scratch.

The second is treating the draft as the final product. AI-generated slides tend to include plausible numbers, round comparisons and punchy phrases that sound true. If nobody checks them, the error shows up in front of the audience, which is the worst moment.

The third is mixing tasks. Asking for message, structure, copy, design and data all at once lets the tool decide things that you should decide. The alternative is to split the work: you decide the message, the AI helps with structure and drafting, and you review whatever is a fact.

How do you define the message before opening any tool?

Write five short answers without AI help. They are worth more than any elaborate prompt.

• Who will be watching, and what that person already knows about the subject.
• Which decision or action you want to trigger (approve a budget, change a process, understand a proposal).
• What the central message is, in one sentence short enough to fit in a WhatsApp message.
• Which three or four arguments support that message.
• Which objection the audience is most likely to raise.

If you cannot answer these questions, the problem is not the tool. AI can help you think (ask it to question you about your goal instead of answering for you), but the decision about what to say stays with you. This is a good example of what is worth keeping with a person, as we discuss in what not to delegate to AI in your business.

How do you use AI to build the script?

The script is the sequence of arguments that takes the audience from the current situation to the decision. Each item in the script should become, at most, one slide. AI helps a lot here because structuring is one of its strongest tasks.

A plan that works in practice:

• Step one: paste the central message, the audience, the expected decision and the likely objection. Also state how much speaking time you have.
• Step two: ask for a bulleted script that names the function of each block (context, problem, proposal, evidence, next steps), with no slide text yet.
• Step three: read the script and cut whatever does not support the message. Ask for an alternative version with a different order and compare the two.
• Step four: mark which blocks need data. Make a separate list of those gaps and fill them in yourself, with sources.
• Step five: only then approve the script and move on to the slides.

The more relevant, well-organized context you provide, the better the structure that comes back. This holds for any use of language models, and Anthropic describes the idea in its material on context engineering. For a presentation, in practice: give the goal, the audience and the facts you have already confirmed, and leave out the noise. If your source material is long, it is worth understanding the context window so that parts of the document are not lost in the middle of your request.

How do you ask for slide drafts without losing control?

Once the script is approved, ask for the draft of each slide in a fixed format. Define the format once and reuse it:

• A title that states the idea (a full sentence, not a label such as "Overview").
• A few lines of supporting text at most, written for someone reading from a distance.
• A suggested visual element (chart, diagram, image), described in words.
• Speaker notes with what to say out loud.
• A "data to confirm" field whenever there is a number, date or proper name.

The last field is the most important one. Instruct the AI never to invent values and to write a visible marker, such as "[confirm number]", in place of any data you did not provide. This turns an invisible risk into a task list.

Then take the text to the tool your team already uses, whether PowerPoint, Google Slides or Canva. Many of them have their own AI features for layout and images, and the principle stays the same: the content has already been decided, and the tool only lays it out. If there are charts, build them yourself from the original spreadsheet, not from numbers repeated back by the AI.

It also helps to limit the tone. Ask for short sentences, no marketing adjectives and the terminology your audience actually uses. For text that leaves the company, the checklist for reviewing AI-generated text works as a second read.

Who should review the numbers, and how?

The author. Not because AI is always wrong, but because the cost of an error falls on the author: it is their name in the room. Language models generate probable text, not verified text, and that includes figures, dates, names of regulations and attributions of quotes.

A lean review, done once the draft is ready:

• List every number, date, proper name and factual claim on each slide in a checking list (a simple list in your notes app is enough).
• For each item, open the original source: spreadsheet, report, system or official document. Record where you checked it.
• Verify units and time periods. A correct number with the wrong period misleads as much as a false one.
• Reread the comparisons ("greater than", "double") and recalculate them yourself.
• Look for absolute words ("always", "never", "everyone") and replace them with statements you can support.
• Confirm that charts and text tell the same story.
• Rehearse out loud at least once, using the speaker notes.

If your data comes from PDFs, contracts or reports, the extraction can also use AI as long as a human checks it, as described in extracting data from PDFs with AI. The rule is the same: AI speeds up the reading, and a person confirms what will be said.

There is also the question of sensitive data. Before pasting information about customers, employees or contracts into any tool, confirm that company policy allows it and remove whatever the request does not need. Brazil's data protection authority (ANPD) publishes guidance on personal data protection, and the OWASP Top 10 for LLM applications lists sensitive information disclosure among the known risks of this kind of system.

Illustrative scenario: a proposal for the leadership team

Imagine, as an illustrative scenario, an operations manager who needs to convince the leadership team to change how inventory is controlled for an online store. She does not open the slide tool. First she writes: the audience is the leadership team, who know the problem from complaints and not from data; the expected decision is to approve a pilot; the message is "we can test a small change, with a clear stop criterion"; the likely objection is the risk of disrupting operations.

She pastes this into the AI and asks for a bulleted script. She gets a sequence of context, problem, pilot proposal, risks and controls, and next steps. She cuts one block that repeats the problem and moves the risks ahead of the proposal, because she knows the objection will come early.

Next, she asks for slide drafts in the fixed format, with the "[confirm number]" marker wherever data is missing. A list of gaps comes back: volume of affected orders, average correction time, cost of the current error. She fills in each one from the system reports and notes the source. Only then does she build the charts in Google Slides, from the spreadsheet.

The result is not a presentation created by AI. It is a presentation thought through by the manager, with AI in the role of drafter and organizer. The rework drops because the hard decisions came before the text.

How does this method apply to working with AI in your company?

The method we use in AI projects works on a smaller scale for a presentation: map, limit, pilot and hand over. Mapping means understanding how you produce presentations today and where the rework happens. Limiting means defining what AI may touch (structure and drafts) and what it may not (numbers, customer data without authorization). Piloting means testing the flow on a real presentation, with a simple success criterion, such as fewer rounds of corrections. Handing over means recording the request format and the checklist so other people on the team can repeat the process. If your company does not have usage rules yet, read about a company policy for employee ChatGPT use before you standardize.

Frequently asked questions

Can AI create the whole presentation on its own? It can generate a complete draft, but the result tends to be generic if you have not defined the audience, goal and message. The most reliable path is to decide the message, approve the script and only then ask for the slides. The final review is still yours.

Which AI tool is best for creating presentations? It depends on what your company already uses and allows. The point is not the brand but the flow: script first, draft second, data checked by you. Choose a tool that fits your company's data policy and test it on a real presentation before standardizing.

Can I trust the numbers AI puts on the slides? Not without checking. Language models can produce plausible values that exist in no source at all. Supply the data yourself and ask for markers in place of anything that is missing. Then confirm each number against the original source.

Is it safe to put customer or company data in the request? Only if company policy and the tool you use allow it, and even then send only what is necessary. Prefer to anonymize names and remove identifiers. If you are unsure about personal data, consult the person responsible for data protection in your company and the official guidance from the ANPD.

Sources

• Anthropic: Effective context engineering for AI agents
• ANPD: Autoridade Nacional de Proteção de Dados
• OWASP Top 10 for LLM Applications