Earlier this year, I don’t think I was alone in wondering whether AI meant we would have to reinvent everything. As the year has progressed, I’ve increasingly realized how much of what we already knew as best practice remains relevant, and, in some cases, becomes more important. How we manage data, govern technology, and measure return on investment are obvious examples.
That does not mean, however, that nothing is changing. New product categories are emerging, and familiar categories are taking on new capabilities. Both are finding their way into otherwise conventional technology diligence. For investors, there is another question worth keeping in mind: what might these emerging capabilities make possible within companies already in the portfolio?
It is not unusual to be introduced to a product described as ‘cutting-edge AI’, only to find that its core functionality is a relatively straightforward application built around a hosted language model. That is not necessarily a criticism. A technically simple product can solve an important problem, improve the customer experience and make money. Complexity is not the objective, but usefulness, technical sophistication and defensibility are different things, and in assessing a company it’s important to be clear-minded on these distinctions.
A product can deliver value today without giving the business a lasting competitive advantage. Equally, its defensibility may lie less in the AI itself than in its proprietary data, the workflow it supports, its integration into customers’ operations or the relationships through which it reaches them etc.
So the diligence needs to get beneath the AI labels and marketing. What does the product actually do? What does it depend on? How easily could someone reproduce the outcome, and why would customers continue to choose this company? Those questions matter when assessing long-term value and whether any premium is justified. There are also potentially new opportunities for companies who would not traditionally have taken on these tech projects, but now sit on assets that can support these new product categories, and allow you to maximize your portfolio return.
As we see more AI within product and more agentic workflows, natural questions arise that are driving these new product categories. Which model powers it? How is the work being coordinated? How does the model or system actually act in the world? What does it know and how does it remember things over time? Where does it actually run? How is it checked and controlled? The image below shows the new categories which are emerging as a result of systems having to answer these questions.
These questions are creating a new layer of infrastructure around the model. Some of these capabilities are genuinely new product categories; others are familiar software infrastructure being reshaped for AI and agentic systems.
Unpacking the categories
Products do not always sit cleanly in one of these categories but may be trying to answer multiple questions as a vertical solution. So the categories aren't always as clean as the image portrays, but each one is worth understanding on its own, as a working solution typically requires all of them.
Models, gateways, and routing
The model itself is increasingly becoming one component of a larger system. Products may use OpenAI, Anthropic, Google or open-weight models, sometimes switching between them depending on cost, latency or the task, with the switching or optimizing potentially giving the company the edge. This has created a growing category of infrastructure for accessing, routing and managing models. Examples include: OpenAI, Anthropic, OpenRouter.
Agent frameworks, harnesses, and orchestration
As AI systems move beyond a single prompt and response, something needs to coordinate the work. There is significant innovation in this space and value to be gained. Agent frameworks and harnesses provide the structure around the model, while orchestration manages sequences of tasks, tool calls and decisions, including workflows that may pause, fail and resume. This area is becoming an important part of where product logic and differentiation can sit. Examples include: LangGraph, Claude Code, Temporal.
Tool integration and connectivity
For an AI system to do more than generate an answer, it needs ways to act on other systems. This might mean searching a database, updating a CRM, sending a message or calling another service. A growing category of products provides these connections, authentication and tools, with standards such as MCP making it easier for models and agents to interact with external systems. It's becoming expected to provide this kind of access into products, as agents become users. Examples include: Composio, Zapier, Pipedream.
Search, retrieval and agent memory
Models only know what is available in their training data and current context, so useful products often need to bring in company, customer or task-specific information. Search and retrieval provide relevant information when it is needed, while memory allows systems to retain useful context across interactions and longer-running work. For many products, the quality and uniqueness of this context may be more defensible than the underlying model. Existing companies may have valuable data and context to bring to this area. Examples include: Pinecone, LlamaIndex, Mem0.
Agent hosting, code execution and sandboxes
Once agents begin carrying out tasks, they need somewhere to run. Creating the need for infrastructure for hosting longer-running agents and providing isolated environments where they can safely execute code, manipulate files or use browsers. Much of this resembles existing cloud infrastructure, but it is being adapted around the particular requirements and risks of autonomous software. Examples include: E2B, Modal.
Evaluation and observability
AI systems are probabilistic and we are giving greater levels of agency to them than standard software, which makes conventional software testing and monitoring insufficient on their own. Evaluation and observability products help teams understand what an AI system did, which models and tools it used, whether its output was good enough and how its behavior changes over time. This means that existing observability tooling has to expand to capture a wider range of logging and new specialist tools are emerging. Which is increasingly important as AI moves from generating content to taking more autonomous actions. Examples include: LangSmith, Braintrust, Arize.
Agent identity and access management
If an agent is going to act on behalf of a person or organization, systems need to know who or what it is, whose authority it is operating under and what it is allowed to access. This is driving new approaches to agent identity, permissions and delegated access, particularly as businesses move from isolated AI tools towards agents operating across multiple corporate systems, and act on our behalf. Examples include: Aembit, Descope, Microsoft Entra Agent ID.
Runtime safety and guardrails
As agents gain greater freedom to use tools and take actions, controls are needed around what they can do while they are running. Guardrails can constrain actions, validate inputs and outputs, enforce policies or require human approval before higher-risk steps. The challenge is increasingly not just making an agent capable of completing a task, but ensuring it stays within acceptable boundaries while doing so. Examples include: Lakera, NVIDIA NeMo Guardrails.
AI and agent governance platforms
Sitting across all of this is the broader question of how organizations govern AI and agentic systems. As their use grows, organizations need to know which models and agents are operating, who owns them, what data and systems they can access, what policies apply and where human accountability sits. These agents are creating a growing category of platforms for maintaining an inventory of AI systems, managing risk and approvals, monitoring compliance and providing an audit trail. Governance itself is not new, but agents make the need more acute. The more agency we give the systems the potentially more useful they are, but the higher the risk if not managed well. Examples include: Credo AI, Holistic AI.
What does this enable?
These capabilities enable AI systems that can move beyond generating content to more effectively surfacing information, remembering context, using tools and acting within defined controls, enabling much more powerful applications. This is already enabling new types of products, including coding and application-building agents, research and knowledge products, voice agents, vertical agents in areas such as legal, finance and customer service, and increasingly sophisticated workflow automation.
Importantly, these products are also becoming easier to build. Companies can combine foundation models and increasingly mature infrastructure with their own data, expertise and existing systems, rather than developing every component themselves. This potentially changes the economics of building new technology products, including for companies that would not traditionally have considered themselves technology businesses.
Much of the recent discussion has focused on the threat this poses to existing software businesses, as AI makes some functionality easier to build or replicate. But the same change can be viewed from the opposite direction: if building useful software and intelligent services is becoming easier, what does that enable companies that already have valuable assets, expertise and customers to do?
What does this mean for existing portfolio companies?
The immediate AI opportunity for many portfolio companies has understandably been framed around efficiency: where can we automate work, reduce cost or allow existing teams to do more? Those opportunities remain important, but the emerging product categories suggest it is worth looking beyond them. For investors, this may be a useful point to reassess existing portfolio companies: not simply for where AI can improve efficiency, but for assets and capabilities that could now support new products, services or sources of revenue.
Many businesses that would never consider themselves technology companies have accumulated potentially valuable assets: proprietary data, documents, specialist knowledge, established workflows, customer relationships and distribution. Historically, exploiting those assets in new ways may have required too much specialist capability or investment. The emerging AI product stack may change what is practicable.
A company with years of proprietary data might be able to create a new knowledge or data product. Operational and maintenance data could support predictive services. Scarce expertise could be made available to many more customers through a vertical or voice agent. Existing distribution could provide an advantage in bringing an AI-enabled product to a market that a technology startup would struggle to reach.
This suggests three questions worth asking across a portfolio:
- Efficiency: Where can these technologies make the existing business cheaper or faster?
- Improvement: Where could they materially improve the existing product or customer experience?
- New opportunity: What assets does the company already possess that could support a product, service or business model that was not previously viable?
The third is perhaps the easiest to overlook. The opportunity is not necessarily to turn every portfolio company into an AI company, but to reassess whether AI has changed the value of what those companies already know, own and do.
That may uncover opportunities for value creation that are considerably more interesting than simply doing the same things a little more efficiently.