VC
About
Media
Startups & Experiments
Advisory
Writing
Resources
Contact
Posts
AI & TechnologyMay 16, 2026 / 10 min

AI Due Diligence: How I Evaluate AI Startups

A practical framework for evaluating AI startups, including data advantage, workflow depth, customer ROI, defensibility, technical risk, and operating leverage.

VCVik ChadhaFounder • Operator • Investor
Expert Knowledge
Based on evaluating AI products, B2B SaaS companies, and portfolio opportunities from an operator-investor lens(Focuses on business durability, not demo quality)
The first wave of AI startup evaluation was too demo-driven. A founder showed an impressive workflow, a chatbot, or an automation layer, and investors had to decide whether it was a feature, a product, or a company. That is the wrong starting point. When I evaluate AI startups, I start with the business system around the AI, not the model itself. The model matters, but the company is built around the customer problem, the workflow, the data, the distribution, and the economic value created. AI does not make weak business fundamentals disappear. It usually exposes them faster. For a founder-side strategy framework, see AI Strategy for CEOs. For the product decision side, read Build vs Buy AI.

AI Demo Quality Is Not Company Quality

AI demos are seductive because they compress complexity. In a few minutes, a founder can show work that would have taken a team hours or days. The product feels inevitable. But a demo does not answer the hard questions:
  • Who owns the budget?
  • How often does the workflow happen?
  • What does the customer do if the AI is wrong?
  • How much behavior change is required?
  • What data advantage improves the product over time?
  • What prevents a platform, incumbent, or competitor from copying the feature?
  • Do margins improve with scale, or do inference costs pressure the business?
Those questions determine whether the startup is durable. The evaluation has to move from "Can this work?" to "Can this become a company?"

The Six Questions I Ask First

AI Startup Diligence Framework
1
Data Advantage
Does usage create proprietary learning?
Unique inputs
Feedback loops
Compounding dataset
2
Workflow Depth
Is the product embedded in important work?
Daily use
System of record
High switching cost
3
Customer ROI
Is the economic value obvious?
Time saved
Revenue gained
Quality improved
4
Defensibility
What gets stronger with scale?
Distribution
Trust
Integration
My diligence framework starts with six questions:
  1. Does the company own or create a meaningful data advantage?
  2. How deeply does the product sit inside the customer's workflow?
  3. Is the ROI obvious and measurable?
  4. Is the buyer clear?
  5. What is defensible over time?
  6. Do margins improve as usage scales?
If a company cannot answer these questions, it may still be a good product. But I am not yet convinced it is a venture-scale company.

1. Does the Company Have a Data Advantage?

The strongest AI companies get better through usage. That does not always mean they train foundation models. Most startups should not. But they should accumulate proprietary context, workflows, labels, feedback, performance data, or domain-specific patterns that make the product more useful over time. A real data advantage might come from:
  • Proprietary customer workflow data.
  • Human feedback loops that improve recommendations.
  • Domain-specific labeled examples.
  • Integration data across systems.
  • Outcome data that connects AI actions to business results.
  • A growing knowledge base that competitors cannot easily reconstruct.
The key question is whether the product becomes meaningfully better because customers use it. If every competitor can plug into the same model, use the same prompts, and produce the same result, the company needs another moat: distribution, brand, workflow ownership, regulatory trust, or superior execution.

2. How Deeply Does It Own the Workflow?

AI products that live at the edge of a workflow are easy to replace. AI products that become part of the operating rhythm are much harder to dislodge. I want to know where the product sits:
  • Is it a nice-to-have assistant?
  • Is it a step in a workflow?
  • Is it the system where work gets done?
  • Is it connected to approvals, reporting, compliance, billing, or customer communication?
The deeper the workflow integration, the more durable the product can become. This is especially important in B2B. Companies do not buy AI because it is interesting. They buy outcomes: fewer manual hours, faster response times, better conversion, lower error rates, higher retention, better forecasting, or reduced risk. The product should map to a workflow that already matters.

3. Is the ROI Obvious?

The best AI startups make the value case almost painfully clear. They save time, reduce headcount pressure, increase revenue, improve quality, reduce risk, or make a previously impossible workflow possible. The strongest companies can quantify this in the buyer's language. Weak ROI sounds like:
  • "It makes the team more productive."
  • "It helps users get insights."
  • "It automates manual work."
  • "It uses AI to streamline the process."
Strong ROI sounds like:
  • "This reduces average handling time by 35 percent."
  • "This cuts review time from three days to twenty minutes."
  • "This increases qualified pipeline by 18 percent with the same sales team."
  • "This prevents compliance errors that cost customers six figures."
  • "This lets a 10-person team operate like a 25-person team."
Founders should be able to explain the value in operational terms, not AI terms.

4. Is the Buyer Clear?

Many AI products have users but no economic buyer. This matters because AI often spreads bottom-up. A team discovers a tool, starts using it, and gets value. But if no executive owns the budget or no department feels the pain strongly enough, usage may not become durable revenue. I look for:
  • A specific buyer with budget authority.
  • A painful enough problem to create urgency.
  • A clear before-and-after workflow.
  • A path from user adoption to enterprise purchase.
  • A pricing model tied to value.
If the product is sold to operations, what metric does the operations leader own? If it is sold to sales, does it improve pipeline, conversion, speed, or account expansion? If it is sold to finance, does it improve cash visibility, forecast accuracy, compliance, or reporting speed? AI startups fail when they sell wonder instead of a budgeted outcome.

5. What Is Defensible Over Time?

Defensibility in AI is tricky because model access is broad and improving quickly. The moat is rarely "we use AI." The moat is what surrounds the AI. Defensibility can come from:
  • Proprietary data.
  • Workflow integration.
  • Distribution advantage.
  • Regulatory or compliance trust.
  • Domain expertise.
  • Brand and credibility.
  • Switching costs.
  • Network effects.
  • Operational execution.
I am skeptical of companies whose moat is only prompt quality. Prompting matters, but it is not enough by itself. The question is what compounds as the company grows. This is where operator judgment matters. A product that looks technically modest may be highly defensible if it controls a painful workflow with high switching costs. A product that looks technically impressive may be fragile if it is a thin layer over a foundation model with no customer lock-in.

6. Do Margins Improve With Scale?

AI changes the cost structure of software. Traditional SaaS companies often have very attractive gross margins once infrastructure is scaled. AI companies may face inference costs, model provider costs, human review costs, data processing costs, and support costs that do not disappear automatically. I want to understand:
  • Cost per task or transaction.
  • Model usage by customer segment.
  • Gross margin at current and projected scale.
  • Whether smaller models can handle common tasks.
  • Whether caching, retrieval, batching, or workflow design improves economics.
  • Whether pricing captures enough value to cover usage growth.
An AI product can be valuable and still have a bad business model if usage costs scale faster than revenue. Founders should know their unit economics early. They do not need perfect answers, but they need a plan.

Red Flags in AI Startup Diligence

The most common red flag is demo dependence. If the pitch depends on a controlled demo but cannot explain deployment, customer adoption, or business impact, the company is not ready. Other red flags:
  • No clear buyer.
  • No proprietary data or workflow advantage.
  • Heavy services work hidden inside the product.
  • Gross margin assumptions that ignore inference cost.
  • Generic positioning that every AI startup could claim.
  • A roadmap controlled by a model provider.
  • Customers using the product experimentally but not operationally.
  • Security and privacy treated as afterthoughts.
  • The founder cannot explain why the incumbent will not copy the feature.
None of these automatically kills a company, but each one requires a clear answer.

What Strong AI Startups Have in Common

The strongest AI startups I see share a few traits. First, they are specific. They do not sell "AI for business." They sell into a defined workflow, buyer, and pain point. Second, they measure value. They can explain what changes for the customer after adoption. Third, they understand distribution. They know how buyers discover, evaluate, approve, and expand the product. Fourth, they treat AI as a means, not the story. The story is the business outcome. Fifth, they build operating leverage into the product and into their own company. They use AI to keep teams small, move faster, and scale support without scaling headcount linearly. This connects to the broader idea in Blueprints for the 10-Person Unicorn: the next generation of high-value companies will be built with fewer people, better systems, and more automation.

The Bottom Line

AI due diligence should not start with the model. It should start with the customer, workflow, data, economics, and operating leverage. The question is not whether the product is impressive. The question is whether the product creates durable business value that gets stronger with usage and scale. The best AI startups turn intelligence into operating leverage. Everything else is a demo.

Related Reading

See What I Am Building

Explore the Scalable Ventures portfolio and active companies to see where I am applying this AI and B2B software thesis in practice.

Related Articles

Explore more insights on entrepreneurship, AI, and leadership:

Explore More

Dive deeper into related topics and resources:
On this page