In this episode of the Influencer Series, Ravi sits down with Manjeet Singh, Head of Enterprise AI at Salesforce (and formerly ServiceNow), to explore why so many enterprise AI initiatives fail to make it past flashy demos—and how AI agents may transform the future of AI adoption in the enterprise. 

The Influencer Series is an intimate, invite-only gathering of influential, good-energy leaders. The intent is to have fun, high-impact, “dinner table” conversations with people you don't know but should. The Influencer Series has connected over 4,000 participants and 15,000 influencers in our community over the last decade.

These roundtable conversations provide a space for prominent VC funds, corporate leaders, start-up founders, academics, and other influencers to explore new ideas through an authentic and connective experience.

Why Most AI Projects Fail—and How AI Agents Could Change Everything

It's no secret that Silicon Valley is experiencing its latest AI gold rush. But amid the excitement, a peculiar pattern has emerged. While companies eagerly showcase their AI capabilities, many find themselves stuck in an endless loop of proofs-of-concept that never actually make it to production.

The enterprise AI landscape is littered with stalled projects and shelved demonstrations, revealing a fundamental disconnect between impressive demos and production-ready systems.

In this episode of the Influencer Series, Ravi Belani sits down with Manjeet Singh, Head of Enterprise AI at Salesforce (and formerly ServiceNow), to explore why so many enterprise AI initiatives fail to make it past flashy demos—and how AI agents may transform the future of AI adoption in the enterprise.

Key takeaways

• Non-deterministic AI systems require new frameworks for testing, monitoring, and improvement beyond traditional software practices.

• Enterprise AI demos typically achieve 60% accuracy, far below the 90%+ needed for production through systematic refinement.

• AI agents offer built-in frameworks for handling hallucinations, bias, and enabling autonomous workflow automation.

• Successful deployment requires monitoring systems that detect subtle response quality degradation and maintain security guardrails.

• Organizations must balance speed with thorough evaluation, using phased rollouts while maintaining market responsiveness.

The Era of Flashy AI Demos

When ChatGPT made its dramatic entrance, enterprises transformed overnight. Companies that once needed extensive data science teams to experiment with AI suddenly found themselves able to validate ideas with remarkable speed and minimal technical overhead.

These early demonstrations pack quite a punch in the boardroom. With relatively simple prompts, teams achieve roughly 60% accuracy—enough to generate genuine excitement among leadership teams eager to demonstrate their AI initiatives. The excitement is particularly palpable when board members, many facing pressure to show concrete AI strategies, see these capabilities firsthand.

But here's the thing: beneath the surface of these compelling demonstrations lurks a more complex reality. The apparent ease of creating AI demos has created a dangerous illusion, leading many organizations to underestimate the profound challenges that emerge when attempting to move beyond the controlled environment of a proof of concept.

Why Enterprise AI Projects Stall After the Demo

At the core of enterprise AI's implementation challenge lies a fundamental characteristic that many organizations initially overlook: these systems are inherently non-deterministic. Ask the same question twice, and you might get two different answers—a feature that creates immediate trust issues in enterprise environments where consistency often equals reliability.

Legal and reputational concerns quickly surface as teams dig deeper into deployment preparations. In the healthcare sector, a biased response could affect patient care. For financial services, AI-generated advice could trigger regulatory scrutiny. These stakes force leadership teams to pause and reassess their appetite for AI-related risks.

Security vulnerabilities present another sobering reality check. Unlike traditional software systems with well-understood attack vectors, AI models introduce novel security challenges that often only become apparent when preparing for broader deployment.

Most organizations lack the sophisticated evaluation frameworks needed to systematically measure, improve, and test AI performance. While simple prompt testing might suffice for demonstrations, it falls woefully short when preparing for production deployment at scale.

The Challenge of AI System Evolution

Moving an AI system from its initial 60% accuracy to production-grade performance requires a methodical approach that many organizations underestimate. Rather than simply fine-tuning prompts or adding more training data, the process demands a fundamental rethinking of how we measure and improve system performance.

The traditional software development playbook proves inadequate for AI systems that evolve based on user interactions and changing data patterns. Where conventional applications follow predictable paths, AI systems exhibit emergent behaviors that require continuous observation and adjustment.

In Silicon Valley's rush to embrace AI, many companies overlook the ongoing resource commitment required for continuous testing, monitoring, and improvement. What starts as a straightforward implementation often evolves into a complex, resource-intensive program requiring dedicated teams, sophisticated tooling, and ongoing maintenance to maintain acceptable performance levels.

Building Effective AI Evaluation Frameworks

For effective AI development, robust testing demands more than clever prompts. Organizations need comprehensive datasets that systematically probe system limitations across multiple dimensions of performance, from accuracy, consistency, to bias and security vulnerabilities.

Before writing a single line of code, successful teams establish clear accuracy thresholds for production readiness. These objective criteria replace the subjective assessments that often drive early AI projects, creating a clear path from development to deployment.

The most effective evaluation approaches leverage batch testing methodologies that examine system performance across thousands of scenarios. This systematic evaluation provides far more realistic deployment readiness metrics than limited manual testing can achieve, revealing edge cases, potential failure modes, and other issues that might otherwise remain hidden until production.

How AI Agents Are Changing the Implementation Paradigm

AI agents represent a fundamental shift in enterprise automation. Unlike passive models that simply respond to queries, these autonomous systems actively perform tasks, make decisions, and execute complex workflows with minimal human intervention.

The agent architecture naturally accommodates the guardrails and monitoring mechanisms that enterprises require. When implemented properly, these systems can incorporate sophisticated checks against hallucination and bias, addressing key concerns that often derail traditional AI implementations.

In practice, multi-agent systems demonstrate remarkable problem-solving capabilities. Picture a product proposal review where specialized agents—each embodying different stakeholder perspectives like finance, technology, and operations—collaborate to provide comprehensive feedback.

Salesforce's Agent Force platform exemplifies how enterprise-focused agent implementations can overcome traditional deployment challenges. The platform's architecture directly addresses the monitoring, security, and reliability concerns that typically slow enterprise AI adoption.

These autonomous systems open new frontiers in workflow automation. Where previous AI technologies required constant human guidance, agents can independently navigate complex business processes while maintaining appropriate guardrails and controls.

Monitoring and Observability for Production AI

Traditional monitoring tools face a steep learning curve in the age of generative AI. While platforms like Datadog and Splunk excel at conventional application monitoring, they require significant adaptation to track the unique behaviors and failure modes of modern AI systems.

Security considerations for AI deployments extend far beyond traditional application security. Novel threats like prompt injection attacks and model-specific vulnerabilities demand new approaches to system protection and monitoring.

The subtle ways AI systems can fail present unique challenges for monitoring teams. Unlike traditional applications, where errors typically trigger clear alerts, AI degradation often manifests as a gradual decline in response quality that traditional monitoring might miss.

Effective AI systems require sophisticated feedback loops that capture both explicit corrections, implicit signals of system failure, and adaptation mechanisms. These mechanisms create the foundation for continuous improvement, allowing systems to learn from their mistakes and adapt to changing conditions.

Organizations implementing AI at scale need comprehensive red-teaming protocols. These systematic attempts to break AI systems reveal vulnerabilities before they can impact customer experience or business operations.

Practical Steps for Moving from POC to Production

Open-source frameworks have democratized AI development. Tools like LangChain provide robust foundations for building production-ready systems, allowing organizations to move forward without waiting for enterprise vendor solutions. Within days, small teams can create working prototypes that demonstrate real business value.

Smart organizations adopt a measured approach to AI deployment. Rather than attempting a big bang release, they gradually expand from limited use cases to broader deployment. Each phase builds confidence while providing valuable insights for future expansion.

Although AI development moves rapidly, speed remains crucial. Successful organizations find ways to balance thorough testing and market responsiveness, often running parallel tracks for different aspects of system development and deployment.

The Future Belongs to Adaptive AI Systems

Organizations that master the art of comprehensive AI testing, monitoring, and improvement will establish insurmountable leads over competitors focused solely on initial capabilities. These leaders recognize that the true value of AI lies not in its first deployment, but in its ability to continuously evolve and improve.

As AI agents mature, they will fundamentally reshape enterprise software. Rather than simply augmenting existing systems, these autonomous agents will increasingly handle complex workflows independently. In this new paradigm, human workers will shift from executing routine tasks to strategic oversight and exception handling, marking a profound transformation in how enterprises operate.

Follow the Alchemist Accelerator Influencer Series on:

Spotify
Apple
YouTube

Thank You to Our Notable Partners

BASF Venture Capital

Investing globally since 2001, BASF Venture Capital backs startups in Decarbonization, Circular Economy, AgTech, New Materials, Digitization, and more. Backed by BASF’s R&D and customer network, BVC plays an active role in scaling disruptive solutions.

WilmerHale

A premier international law firm with deep expertise in Corporate Venture Capital, WilmerHale operates at the nexus of government and business. Contact whlaunch@wilmerhale.com to explore how they can support your CVC strategy.

FinStrat Management

FinStrat Management is a premier outsourced financial operations firm specializing in accounting, finance, and reporting solutions for early-stage and investor-backed companies, family offices, high-net-worth individuals, and venture funds.

The firm’s core offerings include fractional CFO-led accounting + finance services, fund accounting and administration, and portfolio company monitoring + reporting. Through hands-on financial leadership, FinStrat helps clients with strategic forecasting, board reporting, investor communications, capital markets planning, and performance dashboards. The company's fund services provide end-to-end back-office support for venture capital firms, including accounting, investor reporting, and equity management.

In addition to financial operations, FinStrat deploys capital on behalf of investors through a model it calls venture assistance, targeting high-growth companies where FinStrat also serves as an end-to-end outsourced business process strategic partner. Clients benefit from improved financial insight, streamlined operations, and enhanced stakeholder confidence — all at a fraction of the cost of building an in-house team.

FinStrat also produces The Innovators & Investors Podcast, a platform that showcases conversations with leading founders, VCs, and ecosystem builders. The podcast is designed to surface real-world insights from early-stage operators and investors, with the goal of demystifying what drives successful startups and funds. By amplifying these voices, FSM supports the broader early-stage ecosystem, encouraging knowledge-sharing, connectivity, and more efficient founder-investor alignment.

Alchemist connects a global network of enterprise founders, investors, corporations, and mentors to the Silicon Valley community.

Alchemist Accelerator is a global venture-backed accelerator focused on accelerating seed-stage ventures that monetize from enterprises (not consumers). The accelerator invests in enterprise companies with distinctive technical founders and provides founders a structured path to traction, fundraising, mentorship, and community during the 6-month program.

AlchemistX partners with forward-thinking corporations and governments to deliver innovation programs worldwide. These specialized programs leverage the expertise and tools that have fueled Alchemist startups’ success since 2012. Our mission is to transform innovation challenges into opportunities.

Join our community of founders, mentors, and investors.