A recent experience highlighted a critical concern for every AI agent startup and anyone building AI agents. We were deploying Agentforce, our latest sales agent, to engage over 1,000 ghosted leads. Training it on our specific messaging, tone, and process—a task that typically consumed weeks—was surprisingly simple.

Instead, we simply copied existing prompts from another outbound AI tool, made minor adjustments, and it immediately performed as expected. It wasn't a partial success or something requiring extensive tweaking; it simply worked, and worked remarkably well. The quality, personalization, and results—including an impressive 72% open rate—were identical.

If you're building or investing in AI agents, this revelation should make you uncomfortable. For more details on this deployment, see SaaStr + Agentforce: Early Results Are Strong.

The Uncomfortable Truth About AI Agent Moats

We now operate over 20 AI agents across our entire Go-to-Market (GTM) stack, sourced from different vendors, serving various use cases, and at diverse price points. Some were custom-coded, others are specialized platforms, and some originate from incumbents like Salesforce.

Six months of deploying these agents has taught us a crucial lesson: the actual differentiation among AI agents is far narrower than commonly perceived.

Yes, there are distinctions:

  • Some offer a superior user experience (UX).
  • Some integrate more natively with existing tools.
  • Some boast more specialized features.
  • Some provide better deliverability infrastructure.

However, when it comes to core intelligence and the actual AI processing, if trained effectively, leading agents deliver comparable performance. Essentially, similar AI agents can often leverage the same prompts and training methodologies.

The Copy-Paste Revelation

Let's delve into what transpired with Agentforce. Our existing outbound AI agent (let's call it Agent A) had been operational for months, having undergone considerable training:

  • Tone of voice tailored for different audiences.
  • Proof points for various products.
  • When to adopt an aggressive versus a consultative approach.
  • Strategies for handling multi-threaded questions.
  • Addressing edge cases and potential failure modes.

All of this was meticulously documented in detailed prompt instructions, comprising hundreds of words of carefully tuned guidance.

When Agentforce was introduced, we simply opened Agent A's prompt builder, selected all content, copied it, then opened Agentforce's prompt builder and pasted it. We changed "outbound cold prospecting" to "following up with warm leads who raised their hand" and adjusted a few product-specific details. Upon saving, it worked immediately.

It didn't just "work" in the sense of not crashing; it generated emails indistinguishable in quality from Agent A, demonstrating excellent contextual understanding, appropriate personalization, and consistent brand voice.

Why This Matters for AI Agent Companies

If prompts can be copy-pasted between entirely different AI agent platforms—from a startup to Salesforce—and yield comparable results, what truly constitutes a defensible competitive advantage? The uncomfortable truth is: very little.

Here's what we initially believed would be strong moats:

  • Proprietary training data: This is largely ineffective. Every agent can ingest existing company data (Salesforce, website content, past emails). The sophistication of underlying Large Language Models (LLMs) like Claude and GPT-4 means the incremental value of "proprietary training" is minimal.
  • Custom models: While potentially useful for highly niche applications, for general Go-to-Market (GTM) functions, base LLMs already excel at sales and marketing copy. Fine-tuning offers marginal gains, often comparable to the impact of well-crafted, transferable prompts.
  • Unique workflows: Though more significant, these are often replicable. Once a campaign structure is understood in one tool, similar workflows can be recreated elsewhere.
  • Superior AI: This is a misnomer, as most platforms leverage the same foundational LLMs. The "wrapper" around the AI often holds less significance than commonly believed.

Instead, true defensibility stems from:

  • Network effects: Where an agent improves its intelligence through aggregated data from all users, not just individual clients.
  • Deep integration: Native integrations, such as with Salesforce, offer a genuinely superior experience compared to bolted-on solutions.
  • Specialized infrastructure: Critical elements like email deliverability, compliance, and domain management.
  • Workflow lock-in: The inherent cost of migrating when numerous campaigns are already operational within a platform.
  • Speed of innovation: The ability to rapidly deploy new features faster than competitors can replicate their underlying prompt logic.

Notably absent from this list is the core quality of the AI itself.

Universal Principles for AI Agent Training

After deploying over 20 agents, we were somewhat surprised to find that the same fundamental concepts apply universally. Whether training an outbound, inbound, support, or marketing agent, the process involves roughly the same steps:

  1. Define the goal clearly: Whether it's "book meetings," "answer questions," or "generate leads," a precise objective is paramount.
  2. Provide business context: Ingesting documentation, website content, and past conversations.
  3. Establish tone and voice: For instance, professional yet friendly, or data-driven.
  4. Define boundaries: Specifying when to escalate to human intervention and what information to withhold.
  5. Offer examples: Including successful emails, less effective ones, and edge cases.
  6. Iterate based on results: Continuous monitoring, refinement, and repetition.

While user interfaces and terminology may differ (e.g., 'coaching,' 'instructions,' 'system prompts'), the underlying process remains fundamentally consistent. This implies that mastering one agent allows for remarkably rapid deployment of others.

What This Means for Startups

If you're building an AI agent company, this should be a sobering realization. Your AI is not your moat; it is merely table stakes.

The companies that will succeed are not those with the "best AI" (whatever that truly means). They will be the ones that:

  • Go vertical: Instead of building a generic "AI sales agent," focus on highly specific niches, such as "the AI SDR for enterprise SaaS selling to IT buyers." Depth consistently outperforms breadth.
  • Build network effects: Can your agent leverage aggregated customer data to become smarter? Without this, you're merely a superficial wrapper.
  • Own infrastructure: Essential, albeit less glamorous, elements like email deliverability, domain reputation, and compliance represent genuine, defensible moats.
  • Integrate deeply: Native integrations with platforms like Salesforce and HubSpot are demonstrably superior to simple OAuth connections. Prioritize these.
  • Move with extreme speed: If prompts are easily transferable, companies must ship new features faster than competitors can manually replicate their functionality.
  • Create data lock-in: Accumulating significant campaign data within your platform creates real migration costs. The challenge, however, is retaining users long enough to build that data.

The narrative of "we have better AI" is no longer sufficient. Good AI is ubiquitous; the critical question is what additional value you provide.

What This Means for Buyers

For those purchasing AI agents, this is actually excellent news:

  • Reduced vendor lock-in: Buyers are less constrained than they might believe. If a vendor underperforms, their functionality can often be replicated elsewhere with relatively low switching costs, as prompts, training, and accumulated learnings are largely portable.
  • Conduct targeted bake-offs: Instead of evaluating numerous vendors simultaneously, select two, provide them with identical prompts, and compare their performance. The difference may be surprisingly small.
  • Negotiate assertively: Many AI agent companies secure high valuations based on "proprietary AI" that is often less proprietary than advertised. Buyers possess more leverage than they realize.
  • Prioritize non-AI differentiation: Factors like superior deliverability, native integrations, or exceptional onboarding are more impactful than debates over which underlying LLM (e.g., GPT-5 vs. Claude) is employed.
  • Anticipate commoditization and parity: Expect prices to decrease and features to converge. Plan your strategy accordingly.

The Exception: Genuinely Specialized AI Agents

There is one significant caveat to these observations: highly specialized agents with unique workflows.

Our inbound agent, Qualified, for instance, performs functions that are exceptionally difficult to replicate:

  • Real-time website visitor tracking
  • Intent signal aggregation
  • Complex routing logic
  • Integration with our calendar system
  • Pre-meeting context aggregation

Could we theoretically copy-paste prompts and rebuild this? Perhaps. However, the intricate workflow and supporting infrastructure are genuinely differentiated. Similarly, specialized agents leveraging unique datasets or custom models for narrow applications—like a medical coding or legal contract review agent with years of dedicated training—present a much higher barrier to replication. Conversely, for generic Go-to-Market (GTM) agents in sales, marketing, or support, the moats are notably weak.

What We're Doing About It

Understanding this landscape, here's how we're approaching our agent stack:

  • Invest in specialized tools: We utilize distinct platforms for outbound, inbound, and warm follow-up, recognizing that each offers unique advantages beyond its core AI capabilities.
  • Embrace strategic redundancy: While consolidation to a single platform is possible, the low switching costs allow us to optimize for best-in-class solutions for each specific use case.
  • Maintain prompt portability: We curate a library of our most effective prompts, tone guides, and training materials, ensuring they function across various platforms with minimal adjustments.
  • Continuous evaluation: We regularly assess whether each agent remains the optimal choice, as the threshold for switching has significantly lowered.
  • Cultivate founder relationships: In an environment of weak moats, the quality of the product team becomes paramount. We seek partners who innovate rapidly and are highly responsive.

AI Agents: Evolving into Core Infrastructure

We believe AI agents are increasingly transitioning into foundational infrastructure, akin to email service providers, payment processors, or cloud hosting. While indispensable and immensely valuable, their differentiation is progressively shifting towards reliability, integration, and service, rather than the core underlying technology.

The "AI" component is rapidly commoditizing, or at least achieving functional parity at a foundational level. What remains uncommoditized includes:

  • Distribution
  • Integration
  • Domain expertise
  • Customer success