Autonomous Research Tools: Market Map Digest
TL;DR
Last cycle's central question — destination vs. backend — got its answer: the market bifurcates, and both layers are real, with everyone racing to be MCP-native in both directions. Meanwhile, the pricing picture hardened into a $12–20/mo commodity band for individuals and custom-quoted, compliance-gated enterprise deals above it. The wedge that's left isn't autonomy — it's control, auditability, and governance, and nobody in the consumer tier ships those natively.
The destination-vs-backend question resolved: both layers win, but not the way anyone expected
The bifurcation is now documented, not hypothetical: host platforms absorb the casual end of the destination layer while specialists become callable backends. Dynamic Business's 33-tool guide (Sept 15, 2026) shows ChatGPT's Projects "maintains persistent context and custom instructions across conversations so a user's ongoing research or writing project accumulates memory across sessions," and Perplexity ships "a Spaces feature for collaborative research workspaces with shared sources" — meaning $20/mo hosts already give away the lightweight workspace destination-vs-backend-research-layer.
"the general-purpose model supplies reasoning and interface, while the domain layer supplies the standardized method and the grounded corpus." — Cypris, Prompts vs. Agents
Cypris is the proof case for the backend side: 500M patents plus 270M papers, an MCP server, and Cypris Q for Microsoft Copilot launched in August 2026 so teams can call its agents from inside their existing Microsoft environment destination-vs-backend-research-layer. The striking detail is that the bottom-up destination (Panel) is itself a shell over an agent harness — Claude Code only — and its first users immediately asked to call it as a backend: "In my opinion it can be more usefol if I can add it with an mcp server, so I can use with my agent and test it in real workflows" (Gabry848). MCP in both directions isn't a strategy anymore; it's table stakes.
What to watch: The registered tripwire — a host shipping a full native research workspace (files + notebooks + citation library beyond chat Projects), which would compress the workspace layer from above destination-vs-backend-research-layer.
Pricing: the individual tier is a commodity band; enterprise money hides behind compliance
The individual research market has price-compressed into a narrow band where differentiation is by job, not by model. Mid-2026 published prices: Elicit Pro $12/mo, Consensus Pro $15/mo, Perplexity Pro $20/mo, ChatGPT Plus and Claude Pro at $20/mo, with enterprise platforms like Glean and Dust ($39/user/mo) custom-quoted research-tool-pricing-segmentation-2026.
"The most consequential quality difference between AI knowledge tools in 2026 is not the sophistication of their AI model but the reliability of their grounding." — Dynamic Business, 33-tool guide
CASRAI's five-class procurement taxonomy — literature discovery, general LLM assistants, AI reference managers, review screening, and domain Q&A over a corpus — is the buyer-side segmentation taking shape, and its own Ask CASRAI product shows the intermediary template: RAG over a defined corpus that "cites a source for every factual claim" and admits when it doesn't have something, sold as "a paid feature of a Regulatory Radar subscription, with no free tier" research-tool-pricing-segmentation-2026. The Cypris-drawn academic/enterprise split — freemium academic tools versus custom-quoted enterprise platforms with security certifications and patent+paper corpora — is vendor-argued, but the pricing structure confirms it.
What to watch: Whether any consumer-tier tool ships native compliance features (provenance logs, decline behavior, disclosure reports) before enterprise vendors bundle them in — the NIH/NSF/Cochrane disclosure rules define a feature set none of the $12–20 tools currently offer research-tool-pricing-segmentation-2026.
The capability ceiling is quantified — and it's the sales pitch for an orchestrator
The trust axis now has hard numbers, and they're damning for the freemium tier. A 2025 Cochrane study found Elicit averaged roughly 39–40% sensitivity against about 94.5% for traditional systematic-review searches — "precision was better; coverage was not" — and an independent feasibility study found Elicit's supporting quotes reproduced across accounts only about 46% of the time research-tool-pricing-segmentation-2026.
"A retrieval system without a decline behaviour is a chatbot with extra steps." — CASRAI, Best AI for Research
This extends last cycle's Jaccard-based reproducibility findings from a new angle: it's not just that results are unstable, it's that coverage itself is silently incomplete — "they return a cleaner but much smaller slice." The Cypris essay supplies the vocabulary that turns this gap into a value proposition: "A prompt is a single instruction... An agent is an encoded workflow with a defined scope, a defined corpus, defined evidence standards, and defined output structure... The difference between them is not sophistication. It is control" destination-vs-backend-research-layer. An orchestrator that treats retrieval completeness as a first-class, auditable property has a quantified enemy: 39–40% sensitivity.
What to watch: Whether the first vendor publishes an auditable coverage metric — whoever does it first converts the Cochrane data from an indictment of the category into a benchmark they own.
The governance gap is the enterprise wedge
Enterprise adoption is pulling back precisely because governance hasn't caught up, and that's the entry point. KPMG tracked enterprise agent deployment rising from 11% to 42% over 2025 before falling back to 26% in the fourth quarter, attributed to leaders shifting from pilots toward professionalizing their agent systems; Deloitte's 2026 research found only 21% of organizations have a mature AI-agent governance model while roughly 75% intend to deploy agentic AI within two years destination-vs-backend-research-layer.
The regulatory scaffolding is already written: NIH prohibits AI tools in peer review, NSF bars reviewers uploading proposal content to non-approved tools, and the joint Cochrane/Campbell/JBI/CEE statement requires disclosing any AI use "that makes or suggests a judgement — eligibility, extraction, risk of bias, certainty of evidence... with the system name, version, purpose and known limitations" research-tool-pricing-segmentation-2026. Cypris's own positioning language confirms the pitch shift: "The value proposition of agents in R&D is usually pitched as autonomy... That is real but secondary. The primary value is standardization" destination-vs-backend-research-layer
. The 75%-intend-but-only-21%-governed gap is a feature list waiting for a product.
What to watch: Whether the first research tool ships a governance module (audit trails, evidence standards, disclosure logs) as a paid tier — the buyer is the enterprise team stuck between the 75% mandate and the 21% readiness.
What surprised us
- Even destination users want to be backends. Panel's very first substantive feature request was MCP-server exposure — users of a workspace product want to call it from their own agents. "Build a destination" and "be a backend" stopped being an either/or the moment users demanded both.
- The Panel retention doubt is still unanswered. "letting the agent build its own panes is a neat idea. do people keep the layouts it creates?" (quietraster) — the persistence question from last cycle remains open, and it's the make-or-break for the whole agent-authored-UI pattern destination-vs-backend-research-layer
.
- Enterprise deployment went down. The KPMG 42%→26% Q4 pullback is the first hard evidence that the agentic adoption curve isn't a smooth S — governance debt is causing active retreat, which paradoxically strengthens the case for orchestration tooling.
- The recall number is worse than the citation number. 46% quote-reproduction is bad, but 39–40% sensitivity against 94.5% for traditional searches means AI review tools miss most of the literature while looking clean doing it — a failure mode buyers can't see without an audit layer.
Open threads worth a vote
- [watch] Host platform ships a full native research workspace (files + notebooks + citations beyond chat Projects) — the tripwire that would force pure orchestration differentiation destination-vs-backend-research-layer