Evaluate an AI startup by testing its claims against customer behavior, product performance, operating economics, dependencies, and risk controls—not by treating a polished pitch deck as proof. The framework below is for investors comparing companies at different stages; it identifies evidence to request, not universal pass/fail thresholds. Diligence can clarify a company’s strengths and risks, but it cannot predict success.
Start by turning the pitch into testable claims
For each major statement in the deck, write down what would have to be true for it to matter. “Customers love the product,” for example, needs evidence of sustained use or renewal; “our AI is better” needs a defined evaluation and a relevant comparison. Ask for the underlying records, definitions, and time periods—not just a summary slide.
As an Amazon Associate I earn from qualifying purchases.
Keep observed evidence separate from management’s explanation, your assumptions, and your investment thesis. When a measure is immature or unavailable, record what is missing and what evidence could resolve it. Do not estimate a number from a chart or treat an unverified claim as a fact.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Check whether customers have a durable need
Identify the task the product improves, how important or costly that task is, who uses the product, who pays, and what changes in the customer’s workflow after adoption. A user, buyer, and budget owner may be different people; confirm each role rather than assuming that active users imply a paying market.
#1 Best Overall
Look for customer-level evidence over time: repeated use, renewal, expansion, churn, contract duration, revenue concentration, and documented outcomes. A large signup count or polished demonstration does not establish recurring value. Ask whether customers keep using the product after experimentation and whether a particular use case has become important to their work. CRV’s March 5, 2026 investor guide highlights lasting use, expansion, and indispensable use cases; these are investigation points, not proof that a company has achieved them.
Test product performance in conditions that resemble deployment
Request a product demonstration, but do not stop there. Ask the team to show how it evaluates the system on the tasks it claims to handle, what examples are included, how those examples were selected, and whether they represent the intended users and operating conditions.
- Review task-specific measures, test-set provenance, and the baseline used for comparison.
- Inspect examples of uncertainty, incorrect output, edge cases, and known limitations—not only successful examples.
- Ask how the team decides when the system should abstain, escalate to a person, or fail safely.
- Check whether evaluation results can be reproduced or reviewed independently, and whether they reflect conditions similar to deployment.
- Ask how performance is monitored after release and who handles incidents, complaints, and changes in system behavior.
NIST’s AI Risk Management Framework (AI RMF) organizes voluntary lifecycle work around Govern, Map, Measure, and Manage. Its guidance covers context, documented evaluation, deployment-relevant testing, monitoring, and continuing risk management. NIST describes the framework as voluntary, not a certification or proof of product quality. Its framework page reports that AI RMF 1.0 is being revised, so check NIST’s current version when using it.
Recommended Free Tools
Determine whether the advantage is durable—and what it depends on
Ask what makes the product difficult to replace. Possible sources of an advantage include workflow integration, properly licensed or proprietary data, accumulated feedback, distribution, a specialized model or system, and customer switching costs. The team should show how the claimed asset creates a practical advantage; describing the AI as “proprietary” is not enough by itself.
Trace dependencies through third-party models, data, software, cloud infrastructure, and hardware. For material dependencies, ask what rights the company has, how provenance is documented, how resilient access is, and what happens if a provider changes prices, access, terms, or capabilities. NIST’s AI RMF calls for mapping third-party software and data risks, including potential rights infringement. Its July 8, 2026 ICT supplier quick-start guide also discusses ownership and control, provenance, resilience, cybersecurity practices, and supply-chain tiers in the context of supplier assessments. That guide is scoped to ICT supplier diligence, not a universal investment scorecard.
Reconstruct economics rather than accepting headline ratios
Ask the company to define its key measures and show the underlying records used to calculate them. Reconcile recurring revenue, gross profit, customer acquisition cost (CAC), customer lifetime value (LTV), payback, burn, and retention to financial data and cohort assumptions. Make the sales channel and customer segment visible: self-serve, product-led, and enterprise-sales motions can have materially different acquisition costs, service demands, and cash timing.
For AI products, check whether delivery costs include inference, hosting, customer-specific training, onboarding, support, and other usage-related work. A product that attracts more usage may also incur higher costs to serve; examine how those costs affect gross and contribution margin as customers expand. CRV’s March 5, 2026 AI SaaS guide calls attention to inference, hosting, and customer-specific training costs. Its July 23, 2026 Series A article discusses CAC, LTV, payback, margin, and retention together, with assumptions and segments made visible. These are investor perspectives, not universal benchmarks.
Do not use a single payback threshold as a verdict. CRV’s March 2026 article discusses investor rules of thumb, while its July 2026 article emphasizes that acceptable payback depends on the sales model, contrasting self-serve with enterprise economics. The useful test is whether acquisition cost can be reconciled to gross profit, cash timing, and retention for the relevant customer cohort.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Review the team, governance, and operational readiness
Assess whether the team can execute across the technical, product, commercial, and domain work the company needs. Compare roadmap promises with what has shipped and with customer evidence. A credible team should be able to explain tradeoffs and limitations, not only describe an ambitious product vision.
Find out who is responsible for model evaluation, privacy, security, incident response, customer complaints, and human oversight. Ask whether those responsibilities are documented and resourced. Review data rights, access controls, vulnerability handling, third-party risk, and the legal or regulatory duties that apply to the specific use case and jurisdictions.
NIST’s AI RMF emphasizes governance, documented roles, context-specific risk mapping, evaluation, feedback, monitoring, and ongoing management. Renaissance Capital’s AI company checklist also identifies regulatory readiness, privacy, security, and governance as evaluation areas. Neither source establishes that a particular startup complies with every applicable law; assess the company’s actual activities and jurisdictions against current requirements.
Compare alternatives on consistent evidence
When weighing multiple startups or approaches, use the same definitions, time windows, and evidence standard for each. This comparison framework helps expose where one company has stronger proof and where another simply has more polished claims.
| Comparison area | Evidence to compare |
|---|---|
| Customer value | Importance of the use case, verified outcomes, repeat use, renewal, expansion, and customer concentration. |
| Product quality | Task-level performance, reliability, failure modes, fit with deployment conditions, and human oversight. |
| Economics | Gross and contribution margin, inference and service costs, acquisition channel, payback, cash need, and retention. |
| Defensibility and resilience | Data and intellectual-property rights, workflow integration, vendor dependence, compute access, switching costs, and contingency plans. |
| Risk readiness | Relevant privacy, security, fairness and safety testing, governance, monitoring, incident response, and jurisdiction-specific obligations. |
| Execution | Team capability, delivery progress, evidence quality, and whether milestones connect to customer and operating outcomes. |
Use a diligence sequence that ends with explicit uncertainty
- Translate each material pitch claim into a question and request its underlying evidence.
- Validate the problem, user, buyer, workflow change, and measurable outcome with customer-level evidence.
- Inspect product behavior with task-specific tests under realistic conditions; record limitations and failures.
- Map model, data, software, compute, and cloud dependencies, including rights and fallback plans.
- Reconstruct cohort retention and unit economics using fully loaded costs, including AI-variable delivery costs; separate sales motions.
- Review team execution, governance, security, privacy, monitoring, and incident practices.
- Write up the evidence, assumptions, unresolved questions, and downside cases separately from the investment thesis.
This process is a way to make a comparison more disciplined, not a formula for eliminating uncertainty. The sources cited here offer frameworks and investor checklists rather than a single authoritative dataset of industry-wide startup thresholds.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
