Executives from Sapphire Ventures Share Insights on AI Use Cases, R&D Spending, and the impact of Agentic AI on the enterprise
In 2025, AI startups captured the majority of venture capital funding. According to new data from Pitchbook, AI startups now account for 63% of U.S VC investments, up significantly from a year ago when they accounted for 40%. AI is now far outpacing other technologies for VC investments, including fintech, social software, and crypto.
ASUG Executive Exchange recently caught up with Rami Branitzky, Partner, Portfolio Growth, and David Carter, Partner, Engineering Excellence, at Sapphire Ventures to discuss the opportunities they are seeing in AI investments and how they are evaluating the fast-paced Agentic AI landscape. In this Q&A, they share insights on where the strongest enterprise use cases are emerging, how R&D spending in AI is evolving, where technology leaders are making AI investment mistakes, and the growing impact of Agentic AI on the enterprise.
Sapphire Ventures is a global venture capital firm specializing in growth-stage investments, particularly in enterprise software companies with the potential to become category leaders. Founded initially as SAP Ventures, the venture arm of SAP, the firm spun out to become independent in 2011; in 2014 SAP Ventures became Sapphire Ventures.
Today, Sapphire Ventures oversees billions in assets, with a track record that includes more than 30 IPOs and over 80 successful exits.
This interview has been edited and condensed for length and clarity.
Q. Sapphire Ventures has invested in several AI and GenAI startups. Where do you see the strongest enterprise applications emerging, and which use cases are delivering a measurable ROI for customers?
A. We continue to see enterprise value concentrate around four primary use cases: (1) software engineering, (2) employee productivity and knowledge management, (3) customer and employee support, and (4) marketing. Measuring ROI across these diverse use cases is still very much an uneven and evolving practice.
On the software engineering front, measuring ROI is achievable at the individual or workstream level (e.g., % of code contributed by AI), but understanding both organizational (e.g., can we increase quality and velocity without increasing headcount) and business (e.g. tying R&D efforts with increased conversions, user retention, etc.) wide impact remains elusive.
For areas such as help desk and customer support, ticket deflections and auto-resolutions are common metrics we see. This same type of metric can blend into the knowledge management use case, where tickets might be opened less frequently as a result of more digestible documentation.
On the desktop productivity front, many teams run on a pragmatic “this feels more efficient” while more rigorous metrics lag.
And lastly on the marketing front, performance is often subjective and so, again, difficult to quantify and often requires a human in the loop to oversee brand messaging and ensure authenticity.
Beyond these core four, enterprises are investing significant R&D efforts into building autonomous agents (non-deterministic systems with tools, guardrails, memory, and the ability to pursue long-lived goals), capable of owning end-to-end tasks (rather than single prompts), across a broad range of use cases and domains. This shift from more assistive copilots to autonomous agents will shape enterprise software engineering for the next decade.
In this scenario, effective use case scoping is key; start with workflows where partial automation is valuable, design human in the loop checkpoints, and gradually increase autonomy over time as accuracy is validated.
Q. What are the most significant gaps and barriers you're seeing between companies in GenAI pilot’s vs those already in production? From your investment perspective, what are the critical building blocks enterprises need to move GenAI from pilot to production at scale?
A. The gap between “pilot” and “production” is often less about model capabilities and more about operating discipline.
Enterprises that effectively scale GenAI pursuits stack three foundations.
- First, mandates and incentives: a top-down call to change, coupled with both KPIs intended to incentivize adoption and the instrumentation needed to measure utilization (e.g., unused licenses) and ROI.
- Second, culture and structure: flatter organizations tend to move faster; they budget for bottoms-up experimentation, share patterns across silos, and avoid designing every decision by committee.
- Third, data readiness: prior investments in data lakes, transformation and streaming tools, and data catalogs, etc., become the foundation for enabling higher-quality contextual data to be safely and accurately accessed by AI agents.
On the cyber front, CISOs must strike a balance between enforcing guardrails across a massive sprawl of new GenAI tools, each exposing novel threat vectors, while not stifling experimentation and roadmaps by defaulting to the department of no.
AI engineering skills remain scarce, so winners often invest in upskilling and identifying internal champions rather than trying to hire GenAI unicorns.
Software costs are rarely the gating factor, given AI remains a board-level priority, though contract terms still skew shorter.
From a startup ecosystem PoV, we believe platform capabilities make up the critical building blocks—agent frameworks and builder platforms, identity providers and data access controls, guardrails, evaluation harnesses, usage analytics, and observability tooling —combined with empowered teams. Get these pieces right, and scaling from proof of concept to production becomes a more repeatable pattern.
Q. How is the startup market addressing some of these challenges?
A. Startups are helping enable enterprise adoption across three core areas:
Streamlining onboarding and third-party oversight
- Forward-deploy engineers to implement initial use cases
- Off-the-shelf, prebuilt connectors, templated “first mile” use cases, and sandbox environments so teams can prove value quickly under real governance constraints
- Adherence to emerging AI risk and control frameworks (e.g. ISO 42001) to simplify third-party oversight reviews
Providing the foundational building blocks of the modern enterprise AI stack
- Core data platform components, such as unstructured ETL, data storage, metadata catalogs, etc., necessary for making data ‘AI-ready’ and accessible to models and agents
- Cyber for AI platforms which protect against AI model supply chain risks, enforce proper entitlements for agents, scan for and actively protect against prompt injection and other LLM-specific attack vectors, etc.
- Agent infrastructure platform components, including agent orchestrators, memory, model routers and gateways, GPUs for hosting, underlying foundational models, and more
- LLMOps tooling to measure performance, accuracy, handling routing, and control costs across models and workloads
Delivering telemetry to help measure ROI
- Process intelligence and workforce analytics tools (e.g., Skan, ActivTrak) to quantify time saved and flow efficiency
- Domain specific telemetry - for example, developer productivity and code quality intelligence to quantify software engineering gains
Q. Based on your portfolio companies and discussions with enterprise technology leaders, what AI investment mistakes do you see companies make repeatedly?
A. Some common pitfalls include:
- Security teams can sometimes default to blanket ‘deny all’ policies that impede experimentation and learnings.
- Many buyers over-license early, then underutilize.
- “Analysis paralysis”: insisting on perfect accuracy in probabilistic systems can stall momentum.
- We also see too much centralization: heavyweight governance models that require approval by committee and suppress grassroots experimentation.
- Finally, organizations routinely underinvest in upskilling. Training, prompt clinics, hackathons, and internal communities of practice create the muscle memory to infuse genAI efficiently.
- The pattern that works is simple: start with a few high-leverage workflows, instrument them to capture outcome metrics, and reinvest savings into enablement. That flywheel improves both adoption and safety while keeping costs aligned to realized value.
Q. What impact do you foresee Agentic AI having on enterprise IT operations and workforce models?
A. Agentic AI will materially change IT operations. We’re seeing systems that classify and resolve tickets, orchestrate runbooks across ITSM tools, and propose or execute remediations in a true ‘self-healing’ manner. Agents are able to synthesize logs, system configurations and topologies, correlate incidents, and open pull requests for humans to review proposed fixes. As confidence grows, these handoffs will shrink, and agent autonomy will increase.
The broader impact on the workforce is really a reshaping of responsibility matrices. AI creates a blending of duties across traditional role boundaries. For example, designers and product managers can now contribute directly to software engineering with next-gen agents. Conversely, developers can spend more time as system architects and “agent managers.” Support teams can now update knowledge docs with support from text gen and summarization models. Sales and marketing roles are blending research, content, and analytics, and new hybrid personas such as the ‘GTM Engineer’ are rapidly emerging.
Organizationally, we’ve seen a shift away from highly centralized teams of data scientists to more of a federated model. Successful organizations will often construct a centralized AI platform team responsible for delivering a common AI builder stack, associated design patterns, and expert consulting, while AI engineers are federated and embedded within each individual product team.
Q. How are you evaluating investments in Agentic AI?
A. We evaluate Agentic AI startups across a 5D framework:
- Design: End-to-end UX for non-deterministic systems: chain of thought style feedback in the UX, approvals and other human ‘in the loop’ or ‘on-the-loop’ patterns, autonomous error recovery, multi-modal (voice, video) experiences
- Data: Ability to both provide and unlock proprietary data, turning data procurement and management into key differentiators
- Domain Expertise: Depth in the target workflow (e.g., finance ops, ITSM, supply chain) reflected in templates, data sets, and team pedigree and experience
- Dynamism: Ability to provide a more adaptable and personalized experience tuned to each specific workflow and end user
- Distribution: Flexibility in pricing and deployment in ways that can more tightly align with value creation