Artificial intelligence has changed how startups build products. It is now changing how investors should evaluate them. Adrian Vanzyl has long argued that good investing depends on looking past surface momentum and understanding the system beneath a company. With AI-native startups, that system is often harder to inspect because the product, operations, and go-to-market engine may all depend on agents working behind the scenes.
That creates a new diligence problem. A polished demo can look impressive, but it may hide weak reliability, fragile data pipelines, expensive human workarounds, or agents that only perform well in narrow conditions. Investors need a sharper framework for evaluating what is real, what is defensible, and what is merely theatrical.
Why AI-Native Startups Need Different Diligence
Traditional software diligence often focuses on product-market fit, unit economics, customer concentration, competitive positioning, technical architecture, and team quality. Those still matter. But agentic products add another layer: the company may be promising a system that can act, decide, coordinate, or complete workflows with less human involvement.
That promise changes the questions investors need to ask. Does the agent actually complete the task in production, or only in a controlled demo? How often does it fail? What happens when it is uncertain? Is there an audit trail? Does the customer know when a human has intervened? Is the company’s margin profile dependent on hidden manual labor? Can the system improve with usage, or does every new customer require custom implementation?
These questions are not technical details. They go directly to defensibility, scalability, risk, and valuation.
The First Question: What Is the Agent Actually Doing?
Many AI startups describe themselves as agentic, but the word can mean several different things. In diligence, investors should force clarity. Is the agent retrieving information, drafting content, making recommendations, executing transactions, coordinating across tools, or making decisions with real business consequences?
The risk profile changes dramatically depending on the answer. A research agent that summarizes market data is different from an agent that changes pricing, sends customer communications, updates production systems, or moves money. The more consequential the action, the stronger the need for permissions, review gates, rollback, logging, and customer trust.
A practical diligence step is to map the product into three layers:
- Inputs. What data, context, documents, tools, and permissions does the agent rely on?
- Reasoning and workflow. What steps does the agent take, and where does human review enter the process?
- Outputs and consequences. What does the agent change, send, approve, recommend, or commit?
If a company cannot explain these layers clearly, it may not yet understand its own risk.
What Defensibility Looks Like
In AI, investors often look for proprietary models or unique datasets. Those can matter, but many durable companies will be defended by workflow depth rather than model ownership. If a startup deeply understands a painful business process, integrates into the customer’s systems, learns from repeated usage, and becomes part of daily operations, it may build a stronger moat than a company with a more impressive model demo.
Agentic defensibility usually comes from several forces working together:
- Workflow specificity. The product solves a narrow, expensive, recurring job better than a generic agent can.
- Data advantage. The system learns from structured outcomes, not just prompts and documents.
- Trust infrastructure. Customers can see what happened, why it happened, who approved it, and how to reverse it.
- Operational embedding. The agent becomes part of how the customer runs a process, not just a side tool.
- Evaluation discipline. The company measures performance against real tasks and failure modes, not only internal optimism.
The Reliability Question
Investors should ask for reliability evidence early. Agent products are probabilistic by nature, but customers do not buy probabilities in the abstract. They buy outcomes. A company selling an agent into finance, healthcare, legal, enterprise operations, or customer communications needs to show how it handles uncertainty.
That means diligence should include evals, production logs, escalation patterns, human review rates, rollback mechanisms, and customer-visible controls. It is not enough for the founder to say the model is improving. Investors should ask which errors matter, how often they occur, how quickly they are caught, and what the company has changed because of them.
A strong team will usually have a precise answer. A weak team will point back to the demo.
Red Flags in Agentic Startups
There are several warning signs investors should treat seriously. The first is demo-only autonomy. If the product only works when the founder drives it in a prepared environment, the company may be earlier than its story suggests. The second is hidden services work. If humans are manually completing tasks while the company presents the output as autonomous software, margins and scalability may be weaker than advertised.
Another red flag is vague accountability. When an agent makes a mistake, who owns it? The vendor, the customer, the human reviewer, or nobody? Serious buyers will ask that question. Investors should ask it first.
Finally, be careful with companies that cannot explain permissions. Agentic systems often need access to email, calendars, CRMs, code repositories, files, payments, or customer data. If the access model is loose, the product may carry risks that are not visible in the sales deck.
A Practical Investor Checklist
- Ask the team to show a real customer workflow from input to output.
- Separate what the agent does from what humans still do behind the scenes.
- Review reliability data, not just product claims.
- Inspect how permissions, audit trails, and escalation gates work.
- Understand whether each new customer makes the system stronger or simply adds implementation burden.
- Test whether the product could be replaced by a generic AI tool plus a disciplined operator.
Looking Ahead
Agentic due diligence is not about being skeptical of AI. It is about being specific. The best AI-native startups will be able to explain how their agents work, where humans remain accountable, what data makes the system better, and why customers will trust it inside important workflows.
For Adrian Vanzyl, that kind of clarity is central to good investing. AI may change the product surface, but the underlying discipline remains the same: understand the system, identify the leverage, test the risk, and separate durable advantage from temporary excitement.