The Virtual Lab in this issue's framework section sits at the ambitious end of "AI in science." Most working researchers, and most executives trying to understand what is happening in their R&D function, will meet AI through something much more modest first.
SciSpace is a good example of such a tool. I have used it on and off for research. The product started out as Typeset.io, a writing and formatting tool for academics, and pivoted to AI-first literature research later. The value proposition is to search across roughly 280 million academic papers, ask questions across a corpus, get cited answers, extract structured data from a set of papers, and generate literature reviews.
What is in the box in 2026
SciSpace has quietly grown into more than a single tool. It now bundles a stack of things worth knowing about.
-
Library with Zotero import. The fastest way to make the rest of the product useful. Pull your existing reference library in, and the whole platform starts working against your corpus rather than the public one.
-
Chat with a paper, chat with the whole library. The library-level chat changes what feels possible compared to a year ago. Asking a question of fifty papers you already trust, with citations back into the exact passage, is closer to a conversation with your own bibliography than a search.
-
Agent gallery (~550 agents at the moment). Niche, prebuilt task agents: literature review, latex proofing, regulatory document assistance, bioinformatics reporting. These prebuilt agents offer a stronger leverage than the raw chat.
-
AI writer with templates. Research proposal, literature review, abstract, thesis statement. Inline citations as you write. Export as a Word doc.
-
Chat with PDF + highlight-to-explain. Highlight a paragraph or a math expression or a table, and ask for an explanation grounded in cited sources.
-
Smaller utilities. Paraphraser, AI detector, citation generator. None of them revolutionary on their own, but useful where they sit.
Where SciSpace fits in the new economics of research
The framing I use with leaders comes back to the compute-time vs calendar-time point I make in this issue's framework piece. The Virtual Lab spends compute to ask hundreds of questions in an afternoon. SciSpace spends a small fraction of that to ask one researcher's questions across hundreds of papers in a session. Same direction of travel. Very different entry cost.
That is the right way to think about the tool's role. Not as a competitor to a serious research stack, but as the place where most leaders, most researchers, and most knowledge workers will first feel what it is like to have a corpus respond to questions. The discipline I covered in this issue's leadership section on Project Hail Mary applies here directly: divide the work according to what each side is built for. SciSpace is much better than you at scanning 280M papers. You are much better than SciSpace at deciding which question is worth scanning them for.
One practical caveat
SciSpace runs on a credit system on top of the subscription. The cheap tier ($20/month) goes faster than you would expect once you start running the agents. The free 100-credit tier is mostly a trial. Budget accordingly and start with the prebuilt agents that match your actual workflow, rather than burning credits on exploratory chat.
My recommended starting workflow
If you have not used a tool like this before, the order matters more than the feature list. Three steps will get you most of the value with the smallest spend.
- Import your existing reference library from Zotero or a folder of PDFs. The first session has to be against your corpus, not the public one. Otherwise you are doing a Google search with extra steps.
- Pick one prebuilt agent that matches a task you currently do badly or slowly. Literature review and regulatory document assistance are the two I see executives get value from fastest. Run it once on a small input. Read the output critically.
- Use the AI writer for one short artifact, not a full proposal. An abstract, a section, a methodology paragraph. Watch where the inline citations point and read the original papers. That last step is the one most people skip, and it is the one that builds the fluency.
Your action step
If you have a research-adjacent function (R&D, strategy, market research, regulatory, medical affairs, competitive intelligence) and you have never put your own corpus into a tool like SciSpace, that is the week's experiment. Two hours. One library. One prebuilt agent. One artifact you would have written anyway.
The point is not the artifact. The point is to feel, in your own hands, what changes when the corpus answers back. If you feel it, you will start asking different questions of every other AI tool you evaluate this year. That is the leverage worth $20 and a Saturday morning.
If you are mapping where AI fits across your research, strategy, or knowledge-work functions and want a working session to prioritise the highest-leverage entry points, that is the conversation I have in AI strategy advisory engagements and working backwards sessions with leadership teams.