.svg)
How can you assess incoming investment opportunities efficiently while keeping the evidence behind each assessment open to scrutiny? For an emerging fund manager, this question belongs alongside fund structure, governance and operational readiness.
At Sapphire Capital Partners, our Knowledge Transfer Partnership (KTP) with Queen’s University Belfast connects academic expertise with practical fund management. Our DealsFlow research examines how AI can support venture capital deal screening, where human verification remains necessary, and how those findings relate to the infrastructure behind fund incubation.
Key findings for emerging fund managers
- Sapphire’s research with Queen’s University Belfast evaluates evidence-gated AI screening, connecting academic methods with practical questions about venture capital operations.
- In a small study, analyst time fell from 44.7 to 18.9 minutes per application, with additional unattended machine processing time.
- The evidence audit found 84.2% of cited claims fully supported, while 70% of reports still required at least one correction.
- Historical evaluation showed modest discrimination between outcomes, but hindsight effects and funded-only selection limit what you can conclude about future performance.
- For emerging managers, the research offers questions to ask about source verification, documented decisions and human oversight when assessing fund infrastructure.
Why does screening matter when incubating a fund?
Screening determines how your team organises incoming opportunities and decides which deserve further investigation. If you are establishing a fund, you need a repeatable process for reading applications, checking claims and recording why an opportunity progressed or was rejected.
Sapphire helps managers establish and operate venture capital funds. Screening research addresses one part of that operational picture. It does not replace the separate work of choosing a fund structure, preparing documentation or meeting regulatory responsibilities.
For context on the vehicles involved, our guide to SEIS and EIS funds explains their structures and operating considerations. The DealsFlow study concerns assessment processes, rather than eligibility for tax relief.
What the Queen’s Belfast partnership contributes
The KTP brings academic expertise into Sapphire’s development of data and AI capabilities. Our earlier partnership announcement explains the wider programme, including data integration, portfolio monitoring and investor reporting.
The DealsFlow research is a specific contribution within that collaboration. Its authors are Ming Ze Tang, Dr David Cutting, Boyd Carson and Vasiliki Carson. You can find more about the firm’s practitioners on our team page.
The study asks how to constrain a language model’s assessment of startup information so that evidence, scoring rules and human responsibility remain visible. Its findings should be read as research on deal triage, not validation of every aspect of fund management.
How evidence-gated deal screening works
DealsFlow places explicit controls around the language model’s ratings before producing an assessment. The model reads and rates information; code combines ratings and applies limits or reductions.
Sources, including deck information, Companies House data and capped web search results, are gathered and numbered before scoring. A stage-aware ceiling limits unverified factual scores. A second, sceptical pass can lower a score by up to 20 points, but must explain why.
The design records evidence and interventions for audit. A parallel evaluation component is also described in the poster, but was not evaluated in this study. For your own process, the relevant question is whether you can inspect the evidence and controls behind an assessment.
What did the analyst-time study find?
The 2026 research poster reports a 58% reduction in analyst time across 20 applications reviewed by two non-independent analysts. Manual triage took 44.7 minutes per application; DealsFlow-assisted work required 18.9 minutes of analyst time, plus 6.1 minutes of unattended machine time.
The reduction concerns analyst effort in this small comparison. It does not establish that every fund will achieve the same saving, nor that investment decisions or returns improved.
Roughly half the remaining analyst effort involved verifying sources and correcting errors. If you assess an AI screening process, include those activities in your workload comparison rather than measuring report generation alone.
Why verification remains necessary
The claim audit found 202 of 240 cited claims fully supported by their sources, equivalent to 84.2%. Another 22 were partly supported, 10 were unsupported and six had inaccessible sources.
Across the 20 reports, 70% required a correction. A citation therefore did not automatically establish that the associated claim was correct or adequately supported.
For your investment team, the practical lesson is to retain source checking and correction within the workflow. An organised report can reduce reading effort while still requiring substantive professional scrutiny.
What historical outcomes can and cannot establish
The retrospective evaluation found modest separation between outcomes among 142 SEIS/EIS holdings at least three years old. The poster reports an AUC of 0.71 for identifying write-offs, compared with 0.68 for an unconstrained language model. AUC is a discrimination measure, not a percentage accuracy score.
The full pipeline’s ranking also had a Spearman correlation of 0.451 with eventual value, compared with 0.381 for the unconstrained model. The poster notes overlapping intervals with a simple regression baseline.
Interpret these results cautiously. Undated search information and model knowledge may introduce hindsight effects. Only funded deals were included, so missed opportunities are invisible. The study has not established an improvement in live investment decisions or future fund performance.
Where human responsibility sits
DealsFlow retains a human approval gate: the system may reject below a threshold, but does not approve investments. Progression requires Investment Committee action, while overrides require written justification and are logged.
The poster describes record-keeping intended to support accountability, while explicitly stating that the system does not by itself make a firm compliant. You should assess technology within your wider governance arrangements and obtain specialist advice on regulatory responsibilities.
For managers comparing operational partners, the different fund strategies supported by Sapphire illustrate why a clear mandate and accountable decision process matter alongside shared infrastructure.
Taking the research to Milan
The research will be presented as a poster at AI4F, the third Workshop on LLMs and Generative AI for Finance, at Bocconi University in Milan on 14 November 2026, within ICAIF ’26.
The accepted paper is titled “DealsFlow: An Evidence-Gated Agentic Harness for Venture Capital Deal Triage”. The workshop provides a forum to discuss the methods, findings and limitations with researchers and practitioners. Acceptance should not be interpreted as endorsement of a commercial service or proof of investment performance.
What to ask your fund incubation partner
You should ask how your incubation partner develops, tests and governs the processes supporting investment work. This research provides a concrete basis for that discussion, rather than a promise that technology can identify future winners.
Ask which sources support assessments, how errors are corrected, who authorises progression and what evidence exists for claimed efficiency gains. Also distinguish research findings from capabilities available under your particular service arrangement.
Sapphire provides fund incubation and management support; the KTP demonstrates one way we are investigating its supporting technology. To discuss your fund’s operational requirements, contact Sapphire for guidance tailored to your fund objectives.
Frequently asked questions
Does the research prove better investment returns?
No. The study reports analyst-time findings, factual-support checks and retrospective outcome comparisons. It does not establish improved investment decisions or future returns.
Does AI replace the Investment Committee?
No. The design retains Investment Committee action for progression and records justified overrides. AI-assisted analysis remains subject to human checking and responsibility.
Is DealsFlow included in every incubation arrangement?
Yes. Basic access to DealsFlow is provided when you set up a fund with Sapphire, subject to the final agreed terms of your fund incubation arrangement.
This article discusses research findings and fund operations. It is not an offer or invitation to invest.
