Type: Independent, unaffiliated case study — illustrative product teardown of a real, public product (SciSpace/Typeset). This is not a commissioned engagement or insider account; product facts below are drawn from public sources (Capterra, vendor documentation, third-party reviews) and the strategic recommendation is my own analysis. Flagged explicitly, same as I'd flag any assumption in production work.

Domain: AI research assistants · Vertical AI agents · Academic/scientific workflow tools

SciSpace (formerly Typeset, founded 2015, rebranded 2022) is an AI research assistant with 300K+ active users, a 280M+ paper corpus, and multiple specialized agents (SciSpace Agent, Biomedical Agent, Deep Review, Systematic Research/PRISMA). It's a strong reading-and-literature-review tool with two structural gaps that cap its ceiling.

Problem

Two gaps, both visible in public reviews and vendor documentation, that box SciSpace into a single-user, single-workflow product:

1. Domain concentration risk. SciSpace is explicitly strong in biomedical/STEM literature but documented as "not suitable for in-depth research in the humanities, social sciences, engineering sciences, or physical sciences." A platform this narrow is vulnerable the moment a well-funded generalist (a frontier LLM lab shipping a research-agent feature) matches its core reading/summarization loop.
2. No integration surface for agentic workflows. SciSpace has zero API access for external agent frameworks and no MCP server — it integrates with Zotero, Notion, Mendeley and GitHub only as manual import/export destinations. As research workflows move toward agents-calling-agents (a lab's internal tooling, a grant-management system, an institutional repository), SciSpace is invisible to that layer entirely. It's a destination app in a world moving toward composable tooling.

Secondary, monetization-adjacent problem: heavy reliance on individual subscriptions ($12–160/mo tiers) with no visible enterprise/institutional sales motion — despite the natural institutional buyer (university library systems, R&D departments, journal editorial boards) sitting right next to the product.

Approach & Framework

I'd apply the same payer-segment × capability-gap lens I used on Sarvam's on-device teardown: don't chase every discipline SciSpace could serve — identify the smallest defensible expansion wedge with a payer, a workflow, and a channel.

Wedge candidate: institutional research-integrity infrastructure, not broader subject-matter coverage. Two moves:

• Trust/provenance layer as the moat. Every AI-assisted literature review or Deep Review output ships with per-claim source attribution and a confidence/hallucination-risk flag — reviews already note "occasional hallucinations" as SciSpace's main trust gap. Making that gap visible and auditable (rather than hidden) is what an institutional buyer (a university library, a journal) can actually procure against — the same logic as MyUni's clinical-advisor-reviewed MH&W Score, or Sarvam's signed-bundle trust architecture.
• An API/MCP surface, scoped narrowly (read-only paper search + cited-summary retrieval first), so SciSpace becomes callable infrastructure for institutional tooling instead of only a destination app — capturing the agentic-workflow shift instead of being bypassed by it.

Framework — RICE against the two wedges:

Scoring is qualitative (High/Medium/Low) rather than numeric — this is a cold teardown from public sources with no access to SciSpace's internal user counts, feature adoption data, or engineering capacity estimates. Qualitative RICE preserves the relative prioritization logic without fabricating precision I don't have.

Output (What a v1 Would Look Like)

• Provenance layer: every Deep Review claim gets an inline citation-confidence score and a "verify this claim" one-click jump to source — ships as an extension of the existing Deep Review feature, not a new product
• Institutional pilot package: a library/journal-editorial pilot SKU bundling the provenance layer with usage analytics for editorial teams vetting AI-assisted submissions — the natural institutional payer next to an individual-subscriber product