No, ChatGPT and Claude still can't build you a finished, consulting-grade PowerPoint deck on their own — and that's not a knock on either model. Both got genuinely good in 2026 at the job they were actually built for: drafting a narrative, turning a pile of notes into structured bullets, writing speaker notes, catching a weak argument before a partner does. OpenAI's ChatGPT for PowerPoint add-in and Anthropic's Claude for PowerPoint add-in both shipped this year and both do that content work well. What neither one does is the second, separate job a consulting deck also needs: taking a framework slide where the real content lands unevenly — eight risks and one dependency, six KPIs on one scorecard perspective and two on another — and sizing the boxes to match. That's not a drafting problem. It's a layout problem, and it's the one both tools, by their own documentation, still hand back to a human.
"Make me a PowerPoint" sounds like one request. It's actually two, stacked on top of each other. The first is content: what does this deck need to say, in what order, backed by what evidence, written at what level of detail for this audience. The second is layout: given that content, how does it sit on the slide — which boxes get taller because they have more in them, where the alignment guides go, whether a six-item list and a two-item list next to it look like two honest findings or one broken template. Consulting work has always separated these, even before AI entered the picture: an associate drafts the storyline, a different pass — often the same associate, in a different mode — builds the actual grid.
Both ChatGPT and Claude have gotten dramatically better at the first job in 2026. Anthropic's Claude for PowerPoint add-in, which moved from a research preview in February to general availability on May 7, reads a deck's existing slide master — its fonts, colors, and layouts — and uses them when it drafts new slides. It can turn a URL, a PDF, or an Excel file into a starting deck, translate an entire deck into another language, and generate speaker notes. OpenAI's ChatGPT for PowerPoint add-in, which launched two weeks later on May 21, does four things by its own description: build a deck from a brief, update existing slides, understand an existing deck well enough to flag a weak narrative or a question an executive is likely to ask, and polish for editorial tone. Neither of those is a small advance. A content draft that used to take an associate ninety minutes now takes five.
Neither add-in, by its own maker's account, finishes the second job. Anthropic's own documentation for Claude for PowerPoint says the tool isn't recommended for final client deliverables without human review, and isn't a substitute for human judgment on design and narrative flow. OpenAI's help documentation for its add-in says generated or edited slides "may not always match a preferred style perfectly," and lists advanced chart, shape, and slide-management formatting as still limited or in active development. That's two vendors, in their own words, drawing the same line — somewhere past drafting and short of finished layout — independently of each other.
Within the content job, ChatGPT and Claude aren't identical, and the difference matters for which one to reach for on a given deck.
| Axis | ChatGPT for PowerPoint | Claude for PowerPoint |
|---|---|---|
| Launched | GA May 21, 2026 | Research preview Feb 2026, GA May 7, 2026 |
| Core strength | Breadth — live integrations (Gmail, Outlook, SharePoint), fast iterative editing, image generation | Depth — reads slide master/fonts/colors, strong first-draft reasoning on long or complex source material |
| Distinctive feature | Deck-analysis mode — flags weak narrative flow, missing data, likely executive questions | URL/PDF/Excel-to-deck, full-deck translation, Skills and Connectors for repeatable workflows |
| Template fidelity, by the vendor's own account | "May not always match a preferred style perfectly" (OpenAI help docs) | Reads and reuses the existing slide master, but advanced PowerPoint features aren't all supported yet |
| Governance posture | Recommends duplicating files first — edits can delete existing content | Not yet in Enterprise audit logs or the Compliance API; carries a prompt-injection warning on untrusted templates |
| Uneven-content layout (boxes sized to item count) | — not addressed in current docs | — not addressed in current docs |
Independent side-by-side testing through October 2026 lines up with what each vendor claims about itself. Claude's current model tends to produce the stronger out-of-the-box consulting-style deck from a single prompt, with cleaner writing that needs less rework — but it has been reported to struggle with exportable, editable charts. ChatGPT tends to win on iterative back-and-forth editing, image generation, and pulling in live data through its wider set of integrations. Several reviewers land on the same practical workaround: use one tool for research and analysis and the other for drafting, rather than asking either one to run the whole process solo.
It's worth being specific about why the uneven-content row in that table reads the same for both vendors, rather than assuming it's a gap next year's model closes. A chat-based add-in is fundamentally generating and editing text and inserting shapes from a fixed set — that's the operation both companies built, and it's the operation both are visibly still refining, as the "advanced chart, shape, and formatting" caveats in each vendor's own docs show. Sizing a column's height to match a variable item count, while keeping four separate columns' header rows aligned to a shared top edge regardless of how tall each one grows, is a different kind of operation — closer to constraint-based layout than to drafting — and it's not obviously the kind of problem a better chat model solves just by getting better at writing. It's the kind of problem a tool built around PowerPoint's actual shape geometry solves, which is a different product category than either add-in currently occupies.
Here's where that gap shows up in practice, not in a vendor's limitations page. Ask either add-in to build a RAID log slide for a program with eight open risks, three assumptions, two issues, and one dependency — a realistic, lopsided split, the kind that shows up constantly in status reporting and steering-committee decks. Both tools will write the content correctly: eight sensible risk statements, three assumptions, two issues, one dependency, each with an owner and a short description. That's the part they're good at now, and it would have taken a human twenty minutes to draft from a rough list of notes.
What comes back visually is a four-column layout where the columns are either forced to equal height — so the eight-item Risks column overflows while the one-item Dependencies column sits mostly empty — or, more often with the current add-ins, the content gets dropped into a generic card or bullet layout that doesn't attempt column-level sizing at all, leaving a human to rebuild the grid by hand in PowerPoint once the draft lands. Neither failure is a content problem. The risk statements are accurate and well-written. The layout simply isn't reasoning about the fact that a column with eight entries needs roughly four times the vertical space of a column with one — because sizing shapes to match a variable item count, while keeping a shared header row aligned across all four columns, isn't a text-generation task. It's a geometry task, and it's the exact job a grid-based PowerPoint add-in, built to write and edit text and insert pre-set shapes, isn't architected to do.
A note from experience: I ran exactly this test at a global consulting firm, back when the first wave of chat-based slide add-ins started showing up in client-facing teams — not with a RAID log, but with a 3C analysis for a market-entry study where the Competitor research had come back heavy and Company had come back thin. The assistant's narrative was good enough that almost none of the bullet text needed rewriting. What it handed back was three equal-width, equal-height panels, with the six-item Competitor list running past the box border in a font that had clearly been auto-shrunk to force a fit, and the two-item Company panel floating in a mostly empty box the same size next to it. Fixing the writing had taken the tool ninety seconds. Fixing the grid took a human twenty minutes — the same twenty minutes it would have taken without the tool at all, just moved to a different part of the process.
The same wall shows up on every slide type where content naturally lands unevenly: a 3C analysis where the Competitor panel has six findings and the Company panel has two, a capability map where one domain has four L2 capabilities and another has fourteen, a maturity model where one dimension needs a two-line description and another needs five. Ask ChatGPT or Claude to draft the content for any of these, and the content is often genuinely good. Ask either one to also deliver a slide where the boxes are sized to match, and that's the request both vendors are quietly telling you, in their own documentation, to check yourself before it goes to a client.
If the deck doesn't exist yet and you're starting from a blank page, either add-in is a reasonable first move — Claude if the source material is long or messy (a PDF, a set of interview notes, a prior deck to translate) and you want a cleaner first draft with less rework; ChatGPT if you're going to be iterating heavily in conversation, pulling in live data from Gmail or SharePoint, or want its deck-analysis pass to pressure-test the narrative before a review. Teams running both report the best results from splitting the work along those lines rather than forcing one tool through the whole process.
If the deck is going into an existing branded template you need preserved, Claude's slide-master reading is the more deliberate feature of the two — it's explicitly built to notice and reuse fonts, colors, and layouts already in the file — though OpenAI's own caveat about style matching applies here too, and both still recommend a human pass before anything ships.
If the deck includes any framework slide where the real content lands unevenly — a RAID log, a scorecard, a capability map, any multi-panel comparison — plan on a layout pass after either chatbot drafts the content, because that's the specific gap this piece has been describing, and no amount of better prompting closes a gap the tool's own geometry engine doesn't model. That's also, not coincidentally, the exact problem a PowerPoint-native layout tool is built to solve once the content exists.
The calculus also shifts with team size. A boutique firm founder working solo gets the most out of either add-in precisely because there's no associate to hand the layout pass to — the fifteen minutes saved on drafting matters more when it's the founder's own fifteen minutes. A corporate strategy team with a standing deck-production process gets less marginal value from the drafting speed-up, since a competent analyst was already fast at writing bullets, and more value from whatever closes the layout gap specifically, since that's the step that was eating hours regardless of who or what wrote the words.
Duplicate before you let either tool edit an existing file. OpenAI's own guidance for its add-in is to work on a copy, because an edit pass can delete content that was already there — a precaution worth taking regardless of which tool you're using, since an "edit this deck" request is a riskier operation than a "draft from scratch" one.
Treat an unfamiliar template as an untrusted file, not just an old one. Anthropic flags this directly for Claude for PowerPoint: a deck or template from outside your own organization can carry hidden instructions a model will follow without flagging them to you, which matters more than usual once an add-in has edit access to a live file rather than just chat output to copy-paste.
Don't assume enterprise controls cover the add-in yet. As of their current documentation, chat history for Claude's add-in isn't retained between sessions, and the add-in isn't yet included in Enterprise audit logs or the Compliance API — worth confirming against your own firm's data-handling requirements before a regulated client's numbers go through either tool.
None of this is an argument against using ChatGPT or Claude to draft slide content — both are clearly faster and, on the writing itself, often better than what an associate produces under deadline pressure. The argument is narrower: content and layout are different jobs, both vendors say so in their own words, and a deck that's going in front of a client still needs someone, or something, to do the second one. Advisio, listed on Microsoft AppSource, is built specifically for that layout pass — taking the actual content, however many items land in however many categories, and sizing native, editable PowerPoint shapes to match, so a RAID log with eight risks and one dependency comes out aligned instead of overflowing. For a broader look at how eight AI slide tools compare on this exact question, the AI tools for consulting-grade slides roundup tests them side by side.