Business presentation deck datasets for AI training

Presentation deck datasets are real slide files, such as proposals, pitches, board updates and training decks, delivered as editable files with layouts, chart data, speaker notes and revision history, plus a rendered image of each slide. SourceX sources these decks from established businesses' PowerPoint, Google Slides and Keynote archives and links them to the briefs and data behind them where they exist. Images and fonts without cleared rights are removed before delivery, and the license fixes what the decks may be used for.

Request this dataset Email the data team

Dataset manifest

Sourced to your spec

What it is: Real slide decks with layouts, chart data, speaker notes and revisions
Typical systems: PowerPoint on SharePoint or OneDrive, Google Slides, Keynote, Box
Typical history: Decks sit in company archives that often span 3–15+ years; varies by partner
Modality: Native slide files, rendered slide images, extracted text and structure
Delivery formats: Native pptx files, PNG slide renders and slide-level JSONL
Preparation: Names and client identifiers de-identified; third-party images and fonts screened
Licensing: Agreed per dataset; decks made for clients can require the client's sign-off
Availability: Depends on partners willing to license their decks; never guaranteed

Request this dataset

What a delivery contains

Fields vary by source system and are fixed per order. A typical delivery includes:

Field	Type	What it holds
deck_id	string	Pseudonymous deck identifier that ties together its saved versions, slide renders, brief and source workbooks.
deck_type	enum	Purpose, such as client proposal, sales pitch, board update, quarterly review, training or internal strategy.
template	object	Slide master, layouts and theme colors and fonts, with an ID shared by every deck built on the same template.
slides	array	One entry per slide in deck order, with the layout it uses, its section and whether it is hidden.
slides[].shapes	array	Placeholders, text boxes, tables, charts, images and groups with position, size and stacking order.
slides[].text	array	Text by shape, with bullet levels and emphasis, in stored order and in visual reading order.
slides[].chart_data	object	Chart type, series and the values from the chart's embedded workbook, not read off the image.
slides[].notes	string	Speaker notes, kept separate from on-slide text.
slides[].render	string	A PNG render of the slide at an agreed resolution, with any substituted fonts listed.
comments	array	Review comments with author roles, the slide they anchor to, replies and resolution status.
versions	array	Saved states of the deck, each listing the slides added, removed, reordered or edited since the previous one.
assets	array	Images, icons and fonts with their source (in-house, licensed stock, client-supplied, unknown) and whether each was kept.
context	object	The brief, source workbooks and recorded outcome linked to the deck, where they exist and are in scope.

Example record

{
  "deck_id": "dk_62a0f3",
  "path": "/Sales/Proposals/2023/[CLIENT_NAME] fulfillment proposal FINAL.pptx",
  "deck_type": "client_proposal",
  "template": { "id": "tpl_7c1e", "master": "Corporate 16x9", "fonts": ["[BRAND_FONT]", "Arial"] },
  "context": {
    "brief": { "id": "em_19c4e2", "text": "Need a proposal for [CLIENT_NAME] by Thursday. Two options, pricing, onboarding plan." },
    "source_files": ["wb_73b1d0"], "outcome": "shortlisted"
  },
  "slides": [
    { "n": 1, "layout": "Title Slide", "text": ["Fulfillment proposal for [CLIENT_NAME]", "[MONTH] 2023"] },
    { "n": 4, "layout": "Title and Content", "text": ["Two options for year one"],
      "shapes": [{ "type": "table", "rows": 4, "cols": 3,
                   "header": ["", "A: shared site", "B: dedicated zone"] }],
      "notes": "Lead with B. Their peak doubled last year and A would miss the SLA in Q4." },
    { "n": 7, "layout": "Title Only", "text": ["Option B stays ahead of your Q4 peak"],
      "chart_data": { "type": "bar_clustered", "source_file": "wb_73b1d0",
                      "categories": ["Q1", "Q2", "Q3", "Q4"],
                      "series": [{ "name": "forecast orders/day", "values": [1800, 1900, 2100, 4200] },
                                 { "name": "A capacity", "values": [3000, 3000, 3000, 3000] },
                                 { "name": "B capacity", "values": [5000, 5000, 5000, 5000] }] },
      "render": "renders/dk_62a0f3/07.png" },
    { "n": 9, "layout": "Title and Content", "text": ["Live in ten weeks"],
      "notes": "If they push on timing, offer a phased start. Do not commit to eight weeks." },
    { "n": 12, "layout": "Blank", "hidden": true, "text": ["Backup: cost per order by option"] }
  ],
  "comments": [
    { "slide": 4, "author_role": "sales_director", "resolved": true,
      "text": "Drop the cost-per-order row. Price lives on slide 11." }
  ],
  "versions": [
    { "v": 1, "t": "2023-08-14T10:02:00Z", "author_role": "solutions_engineer", "slides": 9 },
    { "v": 5, "t": "2023-08-17T17:45:31Z", "author_role": "account_exec", "slides": 12,
      "changes": ["added slides 7, 10 and 11", "moved slide 5 to 4", "hid slide 12"] }
  ],
  "assets": [
    { "slide": 1, "kind": "photo", "source": "licensed_stock", "kept": false },
    { "slide": 1, "kind": "logo", "source": "client_mark", "kept": false },
    { "kind": "font", "name": "[BRAND_FONT]", "embedded": false, "render_substitute": "Arial" }
  ]
}

Synthetic record for illustration. Field names, structure and format are agreed per order.

What AI teams use it for

Generate decks from a brief

Briefs paired with the decks that answered them show how professionals turn a request into a storyline, choose what to put on each slide and what to move to an appendix.

Edit decks inside an existing template

Masters, layouts and placeholders show how real content fits a company template, so a model can add or restyle slides without breaking it.

Read slides for search and question answering

Renders paired with extracted structure and chart values train models to read charts, tables and diagrams, where much company knowledge exists only as slides.

Draft speaker notes and talk tracks

Notes record what presenters planned to say beyond the slide, including caveats, emphasis and how to handle pushback.

Learn from review and revision

Version chains and review comments show how decks are cut, reordered and corrected before they are presented.

Use-case guides: Document AI, enterprise search and RAG, Enterprise and computer-use agents, Sales agents

What makes this data valuable

Editable structure

Native shapes, placeholders and layouts are worth more than PDFs or images of slides.

Linked briefs

The request behind a deck defines what the deck had to achieve, which a deck alone cannot show.

Speaker notes

Notes carry the spoken argument, caveats and handling of objections that never reach the slide.

Chart data

Embedded chart workbooks keep exact values, so chart reading can be checked against ground truth.

Review trail

Comments and versions between first draft and the presented deck record what reviewers changed.

Template variety

Decks from many teams and templates keep a model from overfitting to one house style.

Why decks need their briefs

A deck can only be judged against the request it answered, and public slide collections almost never include that request. They are mostly finished presentations made for outside audiences, such as conference talks, lecture slides and investor materials, and many survive only as PDFs or images. They show what slides look like, but rarely what a deck was for: who asked for it, what it had to argue, by when, and from which numbers.

Inside a company, a deck is the answer to a specific request, such as two options with pricing and a timeline for a client by Thursday, or a quarterly review that has to explain a missed target. That request sets the storyline, which slides lead and what goes to the appendix. Training on decks alone teaches a model to produce slide-shaped output. Training on brief, source data and final deck teaches it to produce the deck the request called for. The same pairs make useful evaluation items, because each has a reference deck that a professional built and reviewers signed off.

What a slide file holds beyond the picture

A pptx file is a package of XML parts: masters and layouts define placeholders and styles, and each slide inherits from its layout. Text lives in placeholders or in free text boxes, and the order in which shapes are stored, which is also their stacking order, need not match the order a person reads them. That mismatch is a common source of scrambled text in slide extraction, and it is why reading order is worth delivering as a separate field.

Charts are their own parts, with cached values and usually an embedded workbook holding the data, so a chart's numbers can be read exactly instead of estimated from pixels. A placeholder that was never moved stores no position of its own and inherits it from its layout and master, so an extractor has to resolve that chain to know where text sits. SmartArt keeps its text in separate diagram parts that simple extractors skip. For a model that edits slides, these details decide whether its output opens cleanly and fits the template, which is why native files matter more for decks than for most document work.

What to check before licensing

Ask how every embedded image, icon and font was classified by source, and what happened to licensed stock, client-supplied and unknown assets. Some stock image licenses expressly exclude machine learning use.
Confirm who owns client-facing decks. Consulting and agency deliverables are often the client's property under the engagement contract, and client logos and data appear throughout.
Check de-identification off the visible slide — speaker notes, comments, hidden slides, alt text, layouts and masters, file properties and embedded chart workbooks.
Measure duplication on the sample. Archives hold many near-copies of the same deck and reused slides, which should be clustered before training.
Compare renders with native files on a sample of slides. Fonts missing on the rendering machine are substituted, which changes line breaks and can push text out of its box.
Ask how briefs and outcomes were linked to decks — explicit links, shared folders or email threads, or inferred matching — and what share of decks has each link.
Ask which files were converted from Google Slides or Keynote, since conversion to pptx can alter fonts, spacing, charts and effects.

How licensing works through SourceX

1
Define
Send the domain, modality, volume, format, timeline and permitted use you need.
2
Source
SourceX identifies businesses that hold matching data and are open to licensing it.
3
Qualify
Fit, rights and quality are checked, and you review samples before committing.
4
License
Scope, permitted use, exclusivity, price and obligations are agreed in writing.
5
Deliver
Approved data is prepared, de-identified where required and transferred securely.

Questions buyers ask

Can I license real business presentations for AI training?

Yes, when the business that made them owns them and agrees to license them. Decks carry more third-party material than most documents, such as client logos, stock photos, licensed fonts and charts from purchased research, so assets are screened before delivery and anything the partner cannot license is removed or replaced with a placeholder. Whether a match exists depends on which partners hold suitable decks, so describe the deck types, industries and languages you are after.

Do I get native files or images of slides?

Both can be delivered, keyed to the same slide. Native pptx files keep shapes, layouts, text, chart data and speaker notes, which editing and generation models need. Rendered images show what the audience saw, which vision models need. A PDF export keeps the look but drops the editable structure, and usually the notes too, so it is a weaker substitute for either.

How are stock images, logos and fonts handled?

Each asset is classified by source and treated accordingly. Licensed stock images are removed or replaced unless the rights are cleared, because some stock licenses forbid use in machine learning. Client logos are removed or masked. Commercial fonts are licensed software, so embedded font files are usually stripped and the font name kept, and renders list any substitute font used.

Can decks be linked to the brief or data behind them?

Where the partner kept them, yes. Briefs turn up as emails, tickets, CRM notes or kickoff documents, and source data as workbooks that charts were pasted or linked from. Links can come from shared folders, references inside files or the partner's systems, so coverage varies by archive. If briefs matter for your use, set a minimum linked share in your request.

Which kinds of decks can be sourced?

It depends on the partner's business. Sales teams hold pitches, proposals and quarterly business reviews; consultancies and agencies hold client deliverables; corporate teams hold strategy, planning, training and project decks. Board and investor decks are often excluded or released only after extra review because they can contain non-public financial information.

Can I see how a deck changed before it was presented?

Often, from two sources. Platform version histories record saved states of the same file, and review comments record what a manager or client asked to change and whether it was resolved. Decks also evolve by copying, as a new pitch starts from the last one, so a template ID and slide-level similarity can trace each slide back to the deck it came from. Together these show what was added, cut and reworded on the way to the final version.

Enterprise document archives
A company's working files with folders, versions and sharing metadata
Real-world spreadsheets and financial models
Working Excel and Google Sheets files with formulas, links and version history
Sales CRM pipeline histories
Opportunity histories with stage changes, logged activities and won or lost outcomes
Approved workplace email and chat exports
Approved, de-identified exports of team email threads and chat channels

Evaluating this data for procurement?

Diligence packets are prepared per dataset. Rights, privacy processing and quality differ between datasets.

Request dataset diligence

Tell us what your models need

Send your spec — domain, volume, format, timeline and permitted use — and SourceX will match it against partner data and come back with what can be licensed.