Evaluate an AI agent startup on four things: a specific job it finishes end to end, measured reliability on real customer work, a reason it gets stronger (not obsolete) when the next model ships, and unit economics that survive inference costs. A great demo is not evidence; repeat usage, paid expansion and live deployments are.
Why AI agents are a different kind of software bet
A chatbot answers. A copilot suggests. An agent acts: it reads the inbox, updates the CRM, files the claim, books the meeting or writes and runs the code, across several tools, with a human checking in rather than driving. That one difference changes almost everything an angel should look at.
The budget changes. Classic SaaS sells seats to a team. An agent often replaces hours of work, so the money comes from a labor or services line, and pricing moves toward per task, per resolution or per outcome. That can mean much bigger contracts, and it also means the buyer compares the agent to a person, not to another tool.
The risk changes too. When software only suggests, a wrong answer costs a few seconds. When software acts, a wrong action can email the wrong customer or change the wrong record. Reliability, permissions and audit trails stop being nice-to-haves.
Some context on where I sit. I've made 234 angel investments, Lenny's Newsletter ranked me the #1 most active angel investor in 2024, and it described my investing as mostly AI and B2B SaaS. Seven of the twelve companies highlighted on my angel portfolio page are AI companies, including Artisan AI (AI sales automation) and Brainbase (an enterprise AI deployment platform), alongside OpenAI, Anthropic, Scale AI, Hugging Face and Pika. I'm also a technical founder and an 8x startup CTO, which is why the technical questions below get the most space.
How should angels evaluate an AI agent startup in 2026?
Six questions cover most of the diligence. None of them needs a data room; a founder who has done the work can answer each one in a few minutes.
1. What job does the agent finish, and who signs the check?
Ask the founder to name the job in one sentence, the person who used to do it, and the budget line that pays for the agent now. "Helps sales teams" is a category. "Researches inbound leads, writes the first email and books the call, paid from the SDR budget" is a business. The tighter the job, the easier it is to prove value and the harder it is for a general assistant to absorb it.
2. How reliable is it on real work, not the demo?
Demos are curated. Ask what share of real tasks the agent completes without a human fixing it, how that number is measured, and what happens on the tasks it fails. Good teams have an evaluation set built from customer work, track it on every release, and design the hand-off to a human as a feature rather than an embarrassment. "Accuracy" with no denominator is a warning sign.
3. What happens when the next frontier model ships?
Every agent company runs on models that improve every few months. The right answer is that a better model makes the product better: more tasks completed, lower cost, same workflow, same customers. The wrong answer is that the product's main value was working around a weakness the next model removes. Proprietary workflow data, deep integrations into systems of record and customer trust all survive a model upgrade. A clever prompt does not.
4. Do the unit economics survive inference costs?
Agents call models many times per task, so cost per completed task matters more than cost per token. Ask for gross margin today and the plan for it: smaller models for easy steps, caching, fewer retries. Falling model prices help everyone, including competitors, so price per outcome only holds up if the outcome is hard to copy.
5. Can a customer trust it with permissions?
An agent that acts needs access. Ask how it scopes permissions, logs every action, handles prompt injection from emails or web pages it reads, and lets an admin reverse a mistake. Enterprise buyers will ask the same questions in security review; a startup that already has answers sells faster.
6. Is the team built for a moving target?
The stack under an agent changes every quarter. Look for founders who can explain their failure modes plainly, who ship weekly, and who have someone deep in evaluation and infrastructure, not only in prompts. For the questions to ask about model and data claims, see how to test an AI startup's technical claims as an investor.
Where the agent opportunity sits: a simple market map
Vertical agents. One industry, one painful workflow: legal intake, insurance claims, healthcare admin, finance operations, sales development, support. Domain data and regulation make these harder to copy. The case for them is in the rise of vertical AI agents for enterprise.
Deployment and orchestration layers. Tools that help large companies build, run and govern many agents across departments.
Agent infrastructure. Evaluation and observability, memory, browser and computer-use tooling, identity and permissions for agents, and payments that agents can initiate.
Model labs. The base layer. Rounds are large and access is limited, which changes how an angel can take part; angel investing in frontier AI companies covers that side.
Incumbents matter in every row. Large software suites are shipping agents into the customers they already have, Salesforce's Agentforce being one well-known example. A startup wins where the incumbent lacks the workflow, the data or the cross-system reach, or where the job is too narrow for a platform to care about yet.
Red flags in agent pitches
Pilots and proofs of concept reported as revenue.
A single success rate with no description of how it was measured.
Gross margin that is "not a focus yet."
No answer to what the product does when the agent is unsure.
Value that depends on one model provider keeping its current pricing and limits.
A team that can describe the vision but not the last three things that broke.
How to size and structure an AI agent check
Valuations in AI are high, and the category will produce many companies that look alike for a year and then diverge fast. That argues for smaller checks spread across more companies, reserves kept for follow-on rounds in the ones that pull ahead, and close attention to the valuation cap on a SAFE. For what I'm watching across the category this year, read my notes on AI agent trends for 2026 and what the rise of AI agents has taught me. More guides like this one are in the angel investing section.
Frequently Asked Questions
Q:Are AI agent startups too risky for a new angel?
They are riskier than they look, because the technology underneath changes fast and many teams chase the same jobs. A new angel can still take part with small checks, a portfolio spread across several companies, and a firm rule to invest only in teams that can show real usage. Treat every check as money that could go to zero.
Q:What is the difference between an AI agent and an AI copilot?
A copilot suggests and a human decides each step: a draft email, a code completion, a summary. An agent takes the steps itself across tools and asks a human only at checkpoints. That difference moves agents toward per-task pricing and raises the bar for reliability and permissions.
Q:How can I check an agent startup's claims if I'm not an engineer?
Ask for a live run on a task you pick, not a recorded demo, and ask how they measure completion on customer work. Talk to two customers about what the agent does without supervision today. If you can, bring in a technical co-investor for one call; good founders welcome it.
Q:Should angels back model labs or agent applications?
Most angels can only reach model labs through large rounds or special vehicles, with little influence on terms. Agent applications are where early checks still fit, but they carry platform risk from the labs and the incumbents. Many investors hold some of both and size each by how much access and conviction they really have.