Presentation & slide generation

Building a deck is slow: structuring the story, laying out slides, fitting a corporate template and drawing diagrams often takes hours for a few minutes of content. Generative AI drafts the storyline from a prompt or source document, picks layouts, creates charts, diagrams and images, and exports an editable deck in the company template. Typical benefit: a first draft in minutes instead of hours, with people spending their time on message and review rather than formatting.

Problem & value

• Today: staff copy text between documents, nudge boxes in PowerPoint, redraw diagrams and re-apply brand rules. Consultancies and sales teams produce many near-identical decks per week (proposals, status reports, QBRs, training material, lecture slides).
• AI value: (1) storyline and slide-text drafting from sources, (2) automatic layout within a template, (3) chart and diagram generation from data or text, (4) restyling, translation and shortening of existing decks, (5) speaker notes and summaries.
• Evidence: vendors report large time savings, but figures are self-reported and depend on deck type; measure against your own baseline before trusting any percentage.
• Realistic expectation: AI delivers a draft that still needs fact-checking, a point of view and visual polish. It does not replace the author's judgement about the audience and message.

Variants

• Text-to-diagram & infographic generation: flowcharts, mind maps, process visuals and infographics from text. Same stack (LLM structure extraction plus template- or layout-based rendering of editable vectors) and same office users. Tools: Napkin AI, Miro AI and Whimsical AI diagram features, and LLM-to-Mermaid / PlantUML / Graphviz / draw.io XML pipelines that render as code and stay diff-able and editable.
• Sales and proposal decks: generated from CRM data and past winning decks with brand guardrails.
• Consulting and reporting decks: data-driven slides refreshed from spreadsheets or BI (templated monthly reports).
• Education: lecture slides, quizzes and summaries from course material (see AI literacy duties).
• Public sector: briefing decks from reports; EU-hosted or on-premise only.
• Private use: school talks, club events, wedding or birthday slideshows.
• Developer variant: "slides as code" (Marp, Slidev, reveal.js) where an LLM writes Markdown.

How it works

1. brief / source documents / data
2. ingestion (document parsing, transcript, table load)
3. storyline: audience, goal, outline (LLM, human approves outline)
4. slide content: titles, bullets, speaker notes, chart specs, diagram code
5. layout & rendering: map to template layouts/placeholders (or generate HTML/Markdown)
6. visuals: charts from real data, diagrams as vectors, images (image-gen or stock library)
7. review (human): facts, numbers, tone, brand
8. export (.pptx / Slides / PDF) and publish

• Structure extraction: an LLM turns unstructured input into a typed outline (JSON schema with slide type, title, body, notes). Approving the outline first is the cheapest quality gate.
• Rendering: either native editable PowerPoint objects (python-pptx, Office add-ins, Open XML SDK) built on the company slide master, or HTML/Markdown decks exported to PDF/PPTX. Editable output and master-slide compliance matter more than fancy auto-design.
• Charts and numbers: never let the LLM retype figures. Pass data to chart code (native PowerPoint charts, Vega/matplotlib) so numbers are exact.
• Human-in-the-loop: outline approval, fact and number check, final design review before external use.

Data preparation

Slide generation needs little data preparation; the main inputs are the corporate PowerPoint template, brand kit (fonts, colours, logos), an approved image library and the source documents or notes behind a deck. In an SMB these typically live on file shares, in the DMS, in e-mail attachments and on individual laptops. Typical gaps are several competing template versions, outdated logos, scanned PDFs without text layer, unlabeled image libraries and confidential figures mixed into shared folders.

Preparation steps: (1) consolidate one master template with proper layouts and named placeholders, and one approved brand kit; (2) tag the approved image and icon library with usage rights; (3) run OCR on scanned source PDFs (PREP-006) so the tool can read them; (4) collect two or three gold-standard decks as examples (PREP-016); (5) define what may be uploaded into which tool (PREP-024). Tools: Paperless-ngx or the DMS for OCR, simple tagging in the file share. Share of total effort: roughly 10 to 20 percent, mostly template clean-up. See also DOC-014 for document drafting from sources and GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A, because mature EU-hostable deck tools exist at per-seat prices and slide creation is a standard process.
• Enterprise: Path A, because decks are a commodity function and the vendor suites already integrate with the corporate template.
• Developer: Path B, because deck generation is easiest to embed in own products through APIs and python-pptx style templating.
• Private: Path A, because free or low-cost consumer tools cover occasional presentations.
• Public sector: Path B on a sovereign EU or German cloud, because procurement rules and confidentiality of administrative content favour controlled hosting.

Path A - Buy

• Office suites: Microsoft 365 Copilot in PowerPoint (draft from a prompt or Word file, using the company template; EU Data Boundary applies for M365), Google Gemini in Slides / Workspace (EU data regions), Canva Magic Design and Canva Presentations (Australian company, EU data processing options).
• Dedicated AI deck tools: Gamma, Beautiful.ai, Pitch (Berlin, EU), SlidesAI, Presentations.ai, Prezi AI (Hungarian origin), Decktopus. Specialist: Napkin AI and Miro/Whimsical for diagrams and infographics. Check each tool for EU hosting and a DPA.
• Choose A when: the organisation already uses M365 or Workspace (Copilot/Gemini is the lowest friction) or needs quick results without engineering. Check: brand-template import, editable PPTX export, EU region, training-on-customer-data opt-out, SSO.

Path B - Build on cloud APIs

• Models: frontier LLMs with structured outputs (Claude, GPT, Gemini, Mistral as EU provider) for outline and content; image models for illustrations (see GEN-001).
• Rendering: python-pptx or PptxGenJS on the corporate template; Office.js add-ins for in-PowerPoint assistants; Google Slides API; vendor APIs such as the Gamma API for hosted generation. Claude and other vendors also offer ready-made document/PPTX-creation skills for agent workflows.
• Orchestration: n8n / Power Automate / Make to trigger decks from CRM, ticket or BI events; use EU-region endpoints (Azure OpenAI EU, Vertex AI EU, Bedrock Frankfurt, Mistral La Plateforme).
• Choose B when: you need repeatable decks (monthly reports, proposals) from internal data.

Path C - Self-host (open source)

• Slides as code: Marp, Slidev, reveal.js (Markdown/HTML), python-pptx for PPTX, LibreOffice Impress for conversion/preview, Mermaid / Graphviz / PlantUML for diagrams.
• Models: open-weight LLMs (e.g. Qwen, Llama, Mistral families, gpt-oss) served via vLLM or Ollama; open image models (FLUX, Stable Diffusion family) where licences allow; open projects such as PPTAgent and presenton for research and self-hosted deck generation.
• Hardware: HW2 is enough for a 20-30B quantised LLM; add GPU capacity if image generation is included.
• Choose C when: confidential board, M&A or public-sector content must not leave the premises.

Path D - Custom training

Rarely worthwhile. Few-shot examples from the best past decks plus a strict schema usually beat fine-tuning. Consider fine-tuning only for a large, consistent deck corpus (hundreds of decks) where style and structure must be reproduced reliably.

State of the art (2026)

• Agentic deck creation inside the office suites is now standard: assistants read the company template, pull in files and emails, build multi-slide drafts and iterate by chat (M365 Copilot, Gemini, Canva, Gamma). Quality depends heavily on template discipline (proper layouts and placeholders).
• Editable output is the main differentiator versus 2023-24 tools that exported flat images.
• Text-rendering in image models has improved, enabling infographics with legible labels, but exact numbers still belong in native charts.
• Slides as code and diagram as code make decks versionable and reviewable, and fit agent workflows.
• Limits and failure modes: hallucinated facts and statistics, invented citations, generic wording, inconsistent visuals, overfilled slides, broken layouts on unusual templates, wrong numbers when charts are retyped, and copyright/trademark problems with generated images.

Shared resources & combinations

• Same stack: GEN-004 (graphic design and layout), GEN-027 (landing pages), GEN-006 (brand assets), GEN-001 (imagery) share image-gen and brand-kit infrastructure; DOC-014 (document parsing) feeds source content.
• Upstream: KNOW-015 deep-research agents produce the research that becomes the deck.
• Downstream: GEN-003 avatar presenter videos can narrate generated decks.
• Developer: CODE-007 (vibe coding) overlaps for HTML-based decks.
• A single brand-template library and an approved-image repository serve all of these.

Effort, cost & KPIs

• Path A: days; per-seat subscription (Copilot, Gemini, Gamma and similar are priced per user per month).
• Path B: 2-6 weeks for a templated pipeline; LLM cost per deck is typically small compared with staff time.
• Path C: 1-3 months incl. template engineering; GPU server and maintenance.
• KPIs: time to first draft and to final deck, revision rounds, template compliance rate, factual errors found in review, user adoption, share of decks reused/refreshed automatically.

People & change

Involve management early for the AI policy and budget, and the department lead (sales, marketing, consulting) as process owner to name the first decks. Bring IT in for accounts, single sign-on and template deployment, and the data protection officer before customer or personnel data goes into prompts. End users should be part of the tool selection, since they judge the output daily.

Works council: depends. A deck generator alone does not monitor people, but cloud suites usually log usage per user, and the BAG treats objective suitability for monitoring as sufficient under §87(1) no. 6 BetrVG; additionally §90(1) no. 3 requires timely information when work processes are planned with AI. Check the licence's telemetry settings and agree on usage rules. Public bodies involve the Personalrat. A DPIA is rarely needed, but an AVV (Art. 28 GDPR) with the vendor is.

AI literacy: the AI Act Art. 4 requires measures that support staff competence; offer a short training on prompting, fact checking numbers and copyright of generated images (GUIDE-006). Start with a pilot group of 5 to 10 volunteers, name AI champions, and communicate that the tool removes formatting work, not the message. Typical concerns are job loss, uniform "AI look" and leaking confidential figures; address them with review duty (the author remains responsible), a policy (GUIDE-007) and visible human sign-off. See also GUIDE-008.

Risks, regulation & compliance

• EU AI Act: generally minimal risk, but synthetic images, audio or video that look authentic (deepfakes) require disclosure under Art. 50 for deployers; AI-generated text published to inform the public on matters of public interest needs labelling unless human-reviewed under editorial responsibility. Hence rated limited. Art. 4 AI literacy applies to staff using these tools. No Art. 5 issue unless decks are used for manipulative practices.
• GDPR: decks built from CRM, HR or customer documents contain personal data; require a DPA, EU data region, and no training on customer data. Do not upload personal data into consumer tools.
• Confidentiality and IP: trade secrets leaking into third-party services; check copyright and licence of generated images and fonts; avoid logos and likenesses of real persons without consent.
• Accuracy and liability: fabricated numbers or quotes in client or investor decks create legal and reputational risk. Mitigate: source-linked content, native charts from data, mandatory reviewer sign-off.
• Accessibility: public bodies and many companies must meet accessibility rules (EU Accessibility Act, EN 301 549); check alt text, contrast and reading order in generated decks.

Getting started

1. Fix the template first: clean slide master, real placeholders, brand colours and fonts.
2. Pilot Copilot, Gemini or one dedicated tool with 3-5 recurring deck types and 10 users.
3. Define a review checklist (facts, numbers, sources, tone, accessibility, image rights).
4. For recurring reports, automate the pipeline: data to native charts, LLM only for commentary.
5. Track time saved and error findings after one month, then decide on Path B or C.

Sources

1. Microsoft - EU Data Boundary for Microsoft 365 - https://learn.microsoft.com/en-us/privacy/eudb/eu-data-boundary-learn
2. Gamma - API and developer documentation - https://developers.gamma.app
3. python-pptx documentation - https://python-pptx.readthedocs.io
4. Marp - Markdown presentation ecosystem - https://marp.app
5. Slidev - presentation slides for developers - https://sli.dev
6. Mermaid - diagram as code - https://mermaid.js.org
7. PPTAgent (research and open-source slide generation) - https://github.com/icip-cas/PPTAgent
8. Regulation (EU) 2024/1689 (AI Act), Art. 4 and Art. 50 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj