Most investors ask “is this company using AI?” The better question is “does AI compound here?” Six evaluation dimensions and five questions a polished deck cannot survive.
The AI startup investment thesis shifted in 2025. The tools are commoditized - every founding team has access to the same frontier models on the same day. What cannot be commoditized is Domain Expertise accumulated over years, a founder who knows precisely who they are building for, and an organization that runs on AI agents rather than around them. After running 80+ Tech Due Diligences and advising at our AI Venture Studio, we consistently see which founders build moats and which build features. The difference is visible early - if you know what to look for.
“AI-powered” is the new “blockchain-enabled.” It is a signal that requires a filter, not excitement. In 2023, any AI mention raised rounds. In 2026, investors who use the same filter - “is this company using AI?” - are funding feature sets that GPT-5 will commoditize before the Series A. The question has moved. It is no longer whether a startup uses AI. It is whether AI compounds in that company: whether the product gets better as agents run, whether the team moves faster as the playbook matures, and whether the moat widens rather than flattens.
PwC estimates that 20% of companies will capture 74% of AI-era value. The companies in that 20% are not the ones with the most AI features. They are the ones that built the infrastructure, the organizational design, and the domain depth that allows AI to compound. Most are already identifiable at pre-Series A.
The most important structural shift in AI startups is the move from SaaS to AaaS. AaaS - Agent as a Service - is defined as a software product where the primary deliverable is autonomous agent work, not access to a dashboard. The company does not give the customer a tool; it gives them an agent that executes on their behalf.
This changes everything about what you are buying as an investor. The pricing model shifts from per seat to per outcome. The switching cost shifts from data and workflow to agent context and institutional memory - which is significantly higher. The moat shifts from features and integrations to domain-trained agents and proprietary data loops. The risk shifts from churn to agent quality degradation.
| SaaS | AaaS | |
|---|---|---|
| What the customer gets | Access to an interface | An agent that executes |
| Pricing model | Per seat / per month | Per outcome or usage |
| Switching cost | Data + workflow migration | Agent context + institutional memory |
| Moat | Features + integrations | Domain-trained agents + proprietary data |
| Primary investor risk | Churn | Agent quality degradation |
| Upside driver | Seat expansion | Outcome volume + NRR |
The companies most worth backing in 2026 are building toward AaaS even if they are not there yet - meaning the product architecture, the pricing model, and the data loop are oriented toward autonomous output rather than passive dashboard access.
In our Three Forces framework, the startups that become genuinely hard to compete with share exactly three properties. Domain Expertise is the moat - the accumulated knowledge, data, and relationships that a general model cannot replicate without years of operation in that specific vertical. Clarity is the bottleneck - a founder who knows precisely who they are building for and what problem they are solving, before the agent writes a line of code. AI-Native Velocity is the multiplier - the gap that opens between a Level 4 team and a Level 2 team compounds with every sprint.
“Missing one force makes the other two irrelevant. A fast team without direction is just failing faster. A clear team without domain depth is building what a general model will replace.”
Use this as your first filter. If you cannot clearly identify all three forces in a pitch, do not assume they are implicit. They are almost certainly absent.
Horizontal AI is won. Foundation models, general-purpose platforms, developer tooling - OpenAI, Anthropic, Google, and Microsoft have that market. Competing there at pre-Series A is a poor use of angel capital.
The opportunity is vertical: companies with 5-10 years of domain expertise in a specific professional workflow - legal, medical, logistics, construction, compliance - that general models cannot replicate without proprietary data accumulated in that domain. The moat is not the model. The moat is the training data, the customer relationships, and the institutional memory the agent builds as it runs.
This window is closing. As frontier models improve, the bar for what counts as “proprietary” rises. The companies that build the proprietary data loop now - and structure their product to make the agent smarter with every customer interaction - will be significantly harder to displace in 24 months. The ones who don't will be offering a thin wrapper on capabilities that get cheaper every quarter.
The Investor Six-Dimension Scorecard is defined as the six criteria we use to evaluate AI startups at pre-Series A. Each dimension is scored 1-3. A company scoring below 2.5 average is either not fundable or requires a specific condition before round close.
| Dimension | Score 1 | Score 2 | Score 3 |
|---|---|---|---|
| Domain Depth | < 2 years in the vertical | 3-5 years, strong network | 5+ years before AI existed, proprietary relationships |
| Data Moat | Uses public data only | Some proprietary data, no loop | Product generates proprietary data that improves agents over time |
| AI Maturity (Team) | Using tools (Cursor, ChatGPT) | Agents in workflow, partial playbook | Level 4: agent pods, shared playbook, clear human/agent boundary |
| ICP Clarity | Broad market description | Clear segment, 1-2 named customers | Named accounts, documented workflow, measurable outcome |
| AaaS Architecture | Dashboard-first, LLM as feature | Autonomous output in one workflow | Architecture oriented toward agent-delivered outcomes at scale |
| OpenAI Question | OpenAI entering the market would be terminal | Some differentiation, unclear durability | Proprietary data + domain depth = survives a general AI entrant |
Use this alongside our 4 Levels of AI Maturity framework to assess both the product and the team before making your decision. A Level 4 team scores 3 on AI Maturity by definition. Anything below Level 3 in the team means the company is not operating at the velocity the market will require.
A compelling pitch deck proves nothing about Domain Depth or Clarity. These five questions reliably separate real companies from well-packaged ones.
These questions are not trick questions - they are diagnostic ones. The founders who answer them well typically also score high on the Six-Dimension Scorecard without knowing it exists.
Both founders are pitching AI-powered legal workflow tools. Both have a working demo. Both mention Claude and GPT in the same breath.
Founder Aspent 11 years as a contract lawyer before starting the company. The product captures negotiation patterns across 4,000 contracts reviewed by the team and trains agents on that corpus. The ICP is in-house legal teams at Series B-D tech companies. Net Revenue Retention is 118% after 18 months because the agent improves as it processes each new contract. The OpenAI Question answer: “Our training data is proprietary. OpenAI would need 4,000 contracts and 11 years of review experience to replicate what we have. We train on the delta between what the agent flagged and what the lawyer actually changed.”
Founder Bcame from product management at a SaaS company and saw the market opportunity after GPT-4 launched. The product wraps a legal review prompt around a foundation model. The ICP is “SMBs and law firms.” There is no proprietary data loop. The demo is excellent. The OpenAI Question answer: “We'd compete on UX and pricing.”
Same market. Same demo quality. The difference is visible in a 30-minute conversation if you ask the right questions.
The AI investment thesis for 2026-2028 is vertical depth, not horizontal breadth. The companies worth backing are those where domain expertise generates proprietary data, proprietary data trains better agents, better agents deliver better outcomes, and better outcomes create the NRR (aim for 110%+) that makes the unit economics compelling at Series A.
The Five Levels of Company Brain framework gives you one more lens: companies at Level 3+ (Semantic Search / RAG) have the infrastructure to compound. Companies at Level 1-2 are operating on individual prompt sessions with no organizational memory. The latter are building a feature, not a company.
In our Tech Due Diligence Manifesto, we describe how investor-grade AI assessment goes beyond the demo. The same rigour applies here: what the founder says in a pitch and what the code, the data architecture, and the team's daily workflow reveal are often different. The investors who close the gap between pitch and reality win.
AaaS - Agent as a Service - is a software product where the primary deliverable is autonomous agent work, not access to a dashboard. The company gives the customer an agent that executes on their behalf, changing the pricing model (per outcome), the switching cost (agent context and institutional memory), and the moat (domain-trained agents and proprietary data).
A real AI company is one where AI-Native Velocity creates a compounding gap - the product improves, the team moves faster, and the moat widens the longer the company runs. An AI-washed company uses an LLM API call as a feature and could remove it without changing the core value proposition. The test: if the frontier model improves by 10x next quarter, does this company win or get disrupted?
Pre-Series A vertical AI is the angel sweet spot right now. Horizontal AI is won. The opportunity is vertical: companies with 5-10 years of domain expertise in a specific professional workflow that general models cannot replicate without proprietary data. This window is open and closing as frontier models improve.
Ask the founder to describe how their team uses AI internally. Level 4 founders answer fluently - they name specific agents, describe handoffs between them, and articulate what the human is responsible for versus what the agent handles. Level 2 founders say “we use Cursor and ChatGPT.” Level 4 founders describe an organizational system, not a collection of tools.
The OpenAI Question is: “If OpenAI launches a vertical product in your market in 12 months, does your company survive?” It stress-tests moat depth. Companies that survive this question have domain data OpenAI cannot replicate quickly, customer relationships built on outcomes rather than features, and an agent that has learned from proprietary workflows. Companies that fail it are selling access to a general capability with a thin layer on top.
Above The Clouds runs Tech Due Diligence and AI Venture Studio engagements across Europe. If you are evaluating an AI startup - or building one - we apply the same rigour to both sides of the table. Get in touch to discuss your company or portfolio.