Generative AI in asset management Report Interpretation

Overview

Bernstein summarizes a discussion with Perplexity's Head of Computer Capabilities on how Perplexity Computer is evolving from AI search into an enterprise work platform for financial institutions. The central proposition is that multi-model orchestration, connected data, reusable workflows and auditable outputs can automate research-intensive processes while retaining human-defined objectives and decision boundaries.

Core views

The report frames Perplexity as an AI orchestrator rather than a single-model provider. It says the platform currently coordinates 24 frontier and open-source models, allocating subtasks to models suited to their complexity. This approach is intended to reduce vendor lock-in while improving the trade-off among performance, governance and cost: lower-cost models can handle simpler work, while frontier models are reserved for more demanding reasoning. The report argues that the growing diversity of available models makes this orchestration layer increasingly important. Perplexity Computer, launched in February 2026, is described as the operating framework behind this transition from AI-enabled search to end-to-end work execution. Bernstein identifies four pillars: Trusted Data, Work Execution, Auditable Answers and Take Work Everywhere. Users can connect licensed internal or external sources through more than 600 connectors or use pre-negotiated datasets. The platform can produce emails, spreadsheets, presentations, documents, PDFs, markdown files and research reports, while MCP-based integrations allow it to interact with connected applications. Perplexity states it answers more than 1.5 billion questions globally each month. Auditability and enterprise control are central to the report's financial-services use case. For quantitative outputs, users can inspect original sources, filings or transcripts and reconstruct calculation paths for derived metrics. Administrators can control permissions, authentication, user access and approval requirements. The platform also provides visibility into intermediate reasoning steps, code, artifacts, task progress, data sources, connectors, skills and credit consumption. Bernstein characterizes this as a "trust, but verify" design intended to reduce black-box behavior in high-stakes workflows; users can inspect or interrupt tasks without disrupting the overall workflow. The platform's Brain memory layer and reusable Skills are presented as mechanisms for institutionalizing knowledge. Brain incrementally stores user preferences, project context, communication styles and work patterns after it is enabled, so prompts may become shorter as the system gains context. Skills convert preferred examples and proven workflows into version-controlled, shareable templates that can be managed at individual, project or enterprise level. Examples include investment-committee memos, portfolio reviews, due-diligence frameworks and internal reporting standards. The report argues that pre-built infrastructure plus initial configuration of projects, connectors, skills and workflows can create reusable organizational assets rather than requiring firms to build every capability internally. Bernstein highlights automation as the most impactful application for investment teams. Scheduled or event-driven workflows can monitor portfolios daily, track company news and earnings releases, follow macro events, generate alerts and produce research briefs. A "living thesis" can continuously update an original investment thesis using connected news, filings, earnings transcripts, industry developments and internal research; it records changes and preserves version history and source traceability. The intended result is less manual review, more centralized investment knowledge, real-time portfolio intelligence and faster decision support, while human users remain responsible for objectives and decision boundaries. The report also describes Model Council, in which multiple models independently evaluate a scenario and identify consensus and disagreement. This is positioned as useful for judgment-intensive, high-conviction decisions because it exposes differing assumptions and implications. Other showcased functions include live dashboards, task assignment through email or Slack, and automated artifact generation. On infrastructure, Hybrid Compute is cloud-first and routes privacy-sensitive tasks to local devices when appropriate; it is currently available for Mac users. Portable Computer is local-first on NVIDIA DGX Spark hardware and escalates to cloud resources when more computing power is needed. The report says this extends orchestration beyond model selection to deciding whether work should run locally, in the cloud or in a hybrid configuration, with the goal of balancing security, performance and cost. The Q&A attributes the move from AI search to operational AI systems to more capable foundation models, wider tool integration and live-data access, and advances in the harness layer that manages memory, files, skills and workflows. Perplexity recommends beginning deployments in areas with high manual effort, operational friction and fragmented information, then expanding after measurable productivity gains. It also expects both closed frontier models and open models to remain important: closed models for high-performance reasoning and open models for economical simpler workloads.

Analysis framework

Bernstein summarizes a discussion with Perplexity and explains the platform through its product architecture, operational controls and practical asset-management use cases. The analysis links multi-model routing, connected data, memory, reusable workflows and local/cloud deployment to intended benefits in productivity, cost control, traceability and governance.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Perplexity

  Featured enterprise-AI platform discussed as an orchestrator for asset-management research and workflow automation.

  Strengths

  Multi-model orchestration, connected data, auditable outputs, persistent memory, reusable skills, automation and hybrid local/cloud processing.

  Comparison

  Closed frontier models are positioned for complex reasoning, while open models offer more attractive economics for simpler workloads.

  Risks

  The report emphasizes the need for governance, traceability and human-defined objectives in high-stakes decision workflows.

Key data

• Models orchestrated24Perplexity is described as coordinating frontier and open-source models according to task requirements.
• Enterprise connectors600+The platform supports more than 600 connectors for licensed data sources and enterprise systems.
• Monthly questions answeredMore than 1.5 billionReported global monthly activity for Perplexity.
• Perplexity Computer launchFebruary 2026The report identifies this as the launch timing for the work-platform product.
• Max versus Pro creditsAbout 10x more credits per monthPerplexity stated that Max-license users receive roughly ten times the monthly credits of Pro-license users.

Impact & implications

The report suggests that investment organizations may obtain the greatest initial benefit from automating repetitive monitoring and research workflows while retaining source traceability, approval controls and human decision boundaries. It portrays orchestration across models and local/cloud resources as a way to reduce dependence on one vendor and manage the quality-cost-security trade-off.