Votinova AI: the artificial intelligence that turns your deck into a conversation

Most of the AI that has arrived in presentation tools does the same thing: it writes slides for you. Votinova AI starts somewhere else. Your deck already exists and it is probably fine; what it lacks is a way for the room to answer back. So instead of generating new content, it reads what you have, looks for the moments in the script where a question earns its place, and proposes that question with its type, its options and its reason, anchored to the slide it belongs to. Later, once you have presented and there are several hundred open responses sitting in front of you, it groups them into themes, summarises the Q&A and reads the sentiment. What follows is everything it does, feature by feature.

What Votinova AI is, and what it isn't

Votinova AI is the set of artificial-intelligence features built into Votinova. There are four of them, in two families, split by the moment in a session when they work. The generative ones act before you present: they analyse your deck, build quizzes from a topic or a document, and help you word things. The analytical one acts during and after, on what your audience has actually said.

It is worth saying what it does not do, because "AI" has become a label that covers everything. Votinova AI does not design slides or write your content, it does not talk to your audience, and it does not publish anything on its own: every piece it produces is a proposal you accept, edit or discard. It does not step into the live session either. If the AI service went down on a Tuesday afternoon, importing, editing, presenting, voting and pulling reports would carry on exactly as before.

The four features, one at a time

Each one is an independent switch in your organisation's admin settings, so you can run all four, two, or none. We'll take them in the order they appear in the product.

1. Deck analysis: questions anchored to the right slide

This is the most distinctive of the four, and the one that gets the most out of Votinova having genuinely imported your deck. When you upload a PPTX, Votinova doesn't keep a picture of each slide; it keeps its structured representation — the text, the hierarchy of the blocks, the narrative order, the theme, the typography and the palette. That is what the AI reads. It doesn't look at screenshots or try to guess what a photograph says, so it knows what each slide is about and where you are in the argument.

You open the editor, press "Votinova AI" in the toolbar and choose the scope: the whole deck or a range of slides. One analysis returns at most 15 suggestions, ranked by relevance, and never more than one per slide. Each arrives with five things.

• Analyse

  One click in the editor toolbar. The panel tells you where it has got to: reading the deck, generating suggestions.

• Review

  Suggestions land in a side panel. You read the prompt and the reason, and check which slide each one points at.

• Edit if you need to

  You can change the prompt, add or remove options, move the correct answer and re-anchor the question to another slide. Editing costs nothing.

• Accept or discard

  What you accept becomes an anchored interactive question and what you discard disappears. What you leave alone simply doesn't go in: the deck stays as it was.

The analysis also knows its own limits. It first detects the deck's dominant language: if it is one of the five presenter languages (English, Spanish, French, German or Italian) it generates in that language without asking; if it is not, it asks you to pick one of the five and labels the result as indicative. When a deck has no extractable text it invents nothing: it tells you it found no opportunities for a question and suggests widening the range. Slides with no text are skipped and the panel says how many. An analysis that ends with no suggestions costs no generation either.

If you keep editing after running an analysis, the suggestions stay. The one exception is deleting the slide a suggestion pointed at: it is then marked stale, and you have to re-anchor it elsewhere before you can accept it. We prefer that to a question left dangling in the wrong place.

Inside this same feature there is a special case: converting slides that already ask something. Nearly every deck has a slide reading "What do you think?", "Any questions so far?" or "Ready for a challenge?". The analysis spots them and, rather than proposing an invented question, offers to convert that one: it reuses the wording you wrote, proposes a sensible type and leaves the slide visually untouched.

2. Quizzes from a topic, a document or your own slides

The second feature needs nothing imported. It is for the opposite case: you have to assess something and you are starting from scratch, or from material that isn't a deck at all. There are four starting points and all four end in the same place, an editable draft.

Before generating you choose three things: how many questions you want (between 3 and 20), the difficulty (easy, medium or hard) and the output language among the five presenter ones. A topic that works well looks like this: "UK GDPR basics for HR teams, intermediate level". The more specific the brief is about audience, level and angle, the less there is to correct afterwards.

What you get back is a draft, not a finished object. You can add questions, remove them, reorder them, rewrite prompts and options, change which answer is correct and edit the explanation attached to each one. When it is ready you insert it into an existing deck or create a new one from it. Until that moment nothing has entered anything.

About the documents you upload: they are processed only for the generation you asked for and deleted when you resolve the draft, whether you insert or discard it, or after 72 hours if you leave it. They are never shared between organisations or used for anything else. And if the PDF is scanned, password-protected or empty, it tells you so in those words ("no readable text found") and costs no generation.

3. The writing assistant

It is the smallest of the four. On any question — whether the AI wrote it or you did — you can ask for four things: rewrite for clarity, shorten, change the tone (professional, friendly or energetic) and translate between the five presenter languages. The translation takes the prompt and the options together, not one at a time.

They all behave the same way: they show you a preview with the current version on one side and the proposal on the other, and touch nothing until you press "Apply". If you discard, the question is exactly as it was, character for character.

4. Response analysis: reading a room of two hundred

The first three features help you prepare. This one helps you understand. You put an open question to two hundred people and two hundred sentences come back: on paper that is rich data, in the room it is a wall nobody reads while you are still talking. From ten responses onward, Votinova AI groups them into labelled themes — up to eight per question — with the count for each, and the grouping updates itself as more arrive. You can see that "hybrid working policy" and "pay review" are taking half the room without having read a single sentence in full.

The AI proposes the themes, but you are in charge, and your judgement sticks: no automatic update undoes what you have curated.

• Rename

  Swap a theme's label for the one your organisation actually uses. The new name wins everywhere it appears.

• Merge

  Two themes that are really the same thing become one, and responses that arrive later go to the survivor.

• Split

  A theme holding mixed content is broken up into the proposal you confirm.

• Project

  Once the grouping convinces you, switch the projected results from the standard view to the themed one. Participants' own devices always stay on the standard view.

Alongside the themes there are two more readings. Aggregate sentiment gives you the distribution of positive, neutral and negative responses and the tendency for that question; it is how you notice that "office move" ran tense while "new product" ran warm. And the Q&A executive summary hands you, on request during the session or automatically when you close it, the topics people asked about most, the five most-upvoted questions still unanswered, and a list of suggested follow-ups.

All of it turns up afterwards in the report. The "Insights" section joins the ones you already knew and gathers the themes for each open question, the sentiment per question, the Q&A summary, a handful of anonymised standout quotes and the participation figures. It is generated once and then frozen, so every download of the archived report is identical to the last. If you curated the themes after the session you can regenerate it by hand, and the report records that it was regenerated and when.

Which question types it can propose

The AI can propose all 14 interactive types in the catalogue. The two left out are display-only — the join code and the spinner wheel — and collect no answer, so there would be nothing to "ask" with them. These six are the ones you see most often in an analysis.

A worked run: thirty minutes of mandatory training, start to finish

It reads better as a complete example. Take an internal session on the new information-security policy: 24 slides that already exist, 40 people on a video call and half an hour. Preparation runs to about eight minutes.

Look at the last column. All of the AI's work happens before and after; during the half hour that is live, what runs is the same engine as always. It was built that way on purpose: with forty people watching, we don't want anything hanging on a call to a model.

Where it fits best

L&D and HR

It fits neatly because learning and development already works with documents: policies, handbooks, regulation. Document generation turns a twenty-page PDF into an eight-question assessment with explanations, and the deck analysis seeds intermediate checks along the way, so the final assessment isn't the first time somebody discovers they weren't following.

Schools, colleges and universities

A lecturer prepares the same module year after year and the deck already exists. Analysing it by ranges, topic by topic, produces quick checks anchored where they belong. The writing assistant also earns its keep on INSET and CPD days, where the same question often has to work for two very different groups without being rewritten twice.

Conferences and events

Response analysis is what matters here. In rooms of several hundred, an open question produces a volume no chair can process live. Theme grouping and the Q&A summary turn that volume into something usable in the question slot on stage and in the report that goes back to the sponsor.

All-hands and town halls

An all-hands lives or dies on its Q&A. The executive summary — topics, most-upvoted unanswered questions, suggested follow-ups — is more or less the note somebody was going to write up by hand the next morning. And aggregate sentiment gives a reading of the mood without pointing at anyone, which is the condition for people writing what they actually think.

Governance and data: what happens to your content

AI at work has a problem that isn't technical: you have to be able to explain to whoever signs off what gets processed, when and where. That is why governance shipped with the first version rather than in a later round. Your organisation's admin settings hold four independent switches.

Enterprise starts with all four switched off deliberately. In a corporate purchase, AI usually has to clear a security review and a data-processing addendum before anyone turns it on; arriving switched off makes that a box you tick when the time comes rather than a blocker.

On data use there are three commitments, and they appear inside the product itself, one click away from any point where the AI is invoked. No model is ever trained on your organisation's content or your customers'. Content is processed only when a member explicitly invokes an AI feature, never in the background. And inference runs in the EU region, in line with the platform's data residency — which is what makes the UK GDPR and EU GDPR conversation short rather than long.

One last detail that only shows up when people genuinely work together: if you are co-editing a deck across two organisations, the invocation needs the feature switched on in both — the stricter policy always wins — and the generation is charged to the organisation that owns the deck. If something is blocked, the message says which of the two blocked it.

What each plan includes

On the Free plan the remaining count sits in plain view at every entry point, so you don't find the limit by walking into it. The counter is careful about what it charges for: reviewing, editing, accepting, discarding and inserting are all free, and so is an analysis that finds nothing, a generation that fails, or a document with no readable text. If you do run out, everything already generated stays reviewable and insertable; only the next invocation is blocked.

Response analysis comes with the plan, though not only with it: a session run on an event pass includes it too, for that session. If you have already paid for the event, running into an upgrade prompt halfway through it makes no sense.

The limits, stated plainly

A guide that only covers what goes well is no use for deciding. These are the tool's edges today.

• Five output languages

  English, Spanish, French, German and Italian. A deck in another language is still analysed, but you pick which of the five you want the suggestions in and the result is labelled indicative.

• 15 suggestions per analysis, one per slide

  A 200-slide deck doesn't return 200 proposals, it returns the 15 most relevant. To cover a long deck, analyse it in ranges.

• Correct answers need checking

  The correct marking is a proposal. The notice on the draft is there because a wrong right-answer in front of a class is a hard thing to fix live.

• Themes need ten responses

  Below that threshold you get the usual response list and a note. Four sentences don't make groups: that would be inventing a structure that isn't there.

• Heavily mixed-language sessions

  Responses in several languages are grouped by meaning, but the quality of the grouping drops. The panel says so, so you don't discover it mid-session.

• No text, no analysis

  A deck of pure images, or a scanned PDF, gives it nothing to work with. You get an explicit empty result with a recommendation, not an invention.

How to start in five minutes

• Import a deck you already use

  Nothing special to prepare. The more real text it carries, the better the suggestions.

• Press "Votinova AI" in the editor toolbar

  Choose the whole deck or a range and run the analysis. It takes less time than you would expect, even on long decks.

• Accept two or three suggestions

  There is no need to accept eleven. Three well-placed questions already change the session, and you will see straight away whether the anchoring convinces you.

• Present and watch the results

  Accepted questions behave like any other: you activate them from the remote, people vote from their phones with a code or a QR, and results come up live.

Frequently asked questions

Where is our content processed, and is anything used for training?

Inference runs in the EU region, in line with the platform's data residency, which keeps both the UK GDPR and EU GDPR conversation short. No model is ever trained on your organisation's content or your customers', and content is processed only when someone explicitly invokes an AI feature — never in the background.

Does Votinova AI write my slides?

No, and that is a product decision. Votinova AI neither generates nor designs slides: it reads the ones you already have and proposes interactive questions anchored to them. Accepting a suggestion creates no new slide and changes nothing in yours.

Can anything be published without my approval?

No. There is no auto-publish path: no AI-generated question is inserted, activated or shown to a participant without you explicitly accepting it. Reviewing, editing, accepting and discarding also cost no AI quota.

How many generations does the free plan include?

20 a month per organisation, resetting monthly, with the counter visible at every entry point. Pro, Team and Enterprise have unlimited generations. An analysis that finds nothing, a failed generation or a document with no readable text don't count against it.

What does the AI need to analyse my deck?

An imported deck with readable text on the slides. The analysis works on the deck's structured content (text, structure and narrative order), not on screenshots. Slides with no text are skipped and you are told how many.

Which languages does Votinova AI generate in?

English, Spanish, French, German and Italian. It detects your deck's dominant language and generates in it; if the original is in another language, it asks you to pick one of the five and labels the result indicative.

Can the AI see who answered what?

No. Response analysis processes content in aggregate only: sentiment is never attributed to a person, not even in identified or authenticated sessions, and the report's standout quotes carry no name or identifier. Participants don't see themes or sentiment on their devices either.

What happens if the AI service is unavailable?

AI entry points show an explicit unavailable state with a retry, and the rest of the product carries on as normal: importing, editing, presenting, voting and generating reports don't depend on the AI. A live session is never affected.

Can an admin switch the AI off for the whole organisation?

Yes, and feature by feature. There are four independent switches: deck analysis, document generation, writing assistant and response analysis. Changes reach every member in under a minute and are written to the audit log. On Enterprise they arrive switched off until an admin turns them on.