A quick checklist to unmask the truth behind the hype and determine if a startup is building real tech or just dressing up an API.

Follow these steps to peer under the hood and spot a true AI innovator before the smoke and mirrors fade.

1. The One-Killer Question

“If OpenAI (or Anthropic etc.) shut off your API access tomorrow, what still works?”

Real AI company:

• Talks about own models, pipelines, data, on-prem fallback, other providers, fine-tuned checkpoints.
• Mentions pain, but has a plan.

API wrapper:

• Mumbles something about:
• “We’re provider-agnostic”
• “We’ll just switch vendors”
• “We’re more of an orchestration layer”
• Translation: If API dies, we die.

2. Ask: “What is your real IP?”

“What’s your moat, excluding your UI and excluding the base model (GPT, Claude, etc.)?”

Green flags:

• Domain-specific datasets
• Labelling pipelines, evaluators
• Custom ranking / scoring systems
• In-house tools / agents / retrieval infra
• Clear evaluation framework (benchmarks)

Red flags:

• “Our prompts”
• “Our UX”
• “Our workflow builder”
• “Our brand”
• “Our templates marketplace”

Prompts are not IP. They’re seasoning.

3. Follow the Money: Infra & Team

Quick checks:

• “What’s your monthly spend on GPUs / inference infra?”
• “Who on your team has actually trained or fine-tuned a model at scale?”

Red flags:

• No GPU bills, only “OpenAI usage”.
• “We don’t really need ML engineers yet.”
• CTO is a full-stack dev, no real ML depth.

If nobody has suffered through:

• CUDA errors
• exploding gradients
• data cleaning hell
  …it’s probably an API wrapper.

4. The Latency Fingerprint Test

Ask them to:

• Run the product live
• Try a few unprompted, weird queries
• Notice:
• Response time
• Style
• Failure modes

If it:

• Feels exactly like ChatGPT/Claude
• Has similar delay patterns
• Hallucinates in the same style

…you’re basically watching re-skinned ChatGPT.

5. Ask for the Architecture Diagram

“Show me your technical architecture, from data ingestion to model output.”

Green flags:

• Separate blocks for:
• Data ingestion
• Preprocessing
• Vector DB / retrieval
• Model(s)
• Evaluation / monitoring
• Feedback loop / retraining

Red flags:

• Big box: “LLM provider”
• Arrow to: “Our app”
• Lots of arrows and buzzwords, no data flow clarity.

If the entire brain is one SaaS logo, it’s a wrapper.

6. Ask About Evaluation

“How do you measure model quality? Show me your benchmarks.”

Real AI team:

• Talks about:
• Accuracy, F1, BLEU, ROUGE, win-rates
• Custom eval datasets
• Regression tests
• A/B experiments

Wrapper team:

• Talks about:
• “Users love it”
• “Great feedback”
• “Engagement is high”
• “We’re iterating fast”

No eval pipeline = no depth.

7. Model Ownership Question

“Which parts of your system are fully under your control, and which are just vendor dependencies?”

You’re looking for:

• In-house models or at least adapted models
• Own embedding / retrieval / ranking stack
• Ability to move between providers without rewriting the whole product

Red flag answer:

“We’re built deeply on OpenAI, but we have a lot of optimizations on top.”

That’s like saying:

“We own a restaurant. Our IP is Swiggy.”

8. Data Story Interrogation

“Walk me through your data pipeline.”

Good answer includes:

• Where data comes from
• How it’s cleaned
• How labels are created
• How it’s stored
• How it’s used for:
• Fine-tuning
• RAG
• Evaluations

Red flags:

• “We don’t really need data, the foundation model is so good.”
• “Clients bring their own data and we just plug it in.”
• “We store it in a vector DB and… magic.”

No data thinking → glorified front-end.

9. Ask for a Local / Air-gapped Story

“Could you run a version of this entirely on-prem or air-gapped, if a bank or hospital required it?”

Real AI company:

• Says “yes, but expensive”, and explains:
• Containerization
• Self-hosted models
• Security considerations

API wrapper:

• “We’re cloud-native.”
• “Our value is in the cloud.”
• “Security is handled by OpenAI/AWS/etc.”

Translation: No control, no depth.

10. The “Non-LLM Feature” Trap

“Show me a feature your product has that would still be valuable even if LLMs disappeared tomorrow.”

If they can’t name:

• Workflows
• Integrations
• Dashboards
• Analytics
• Domain-specific tools

…then the only asset is “access to someone else’s model”.

That’s not a startup. That’s a skin.

11. Contract & Pricing Smell Test

Red flags:

• Pricing is purely usage-based on tokens with a fat margin.
• No:
• Implementation fee
• Customization
• Managed service component
• Value prop is:
• “We make ChatGPT safer/easier for your team.”

That’s basically:

“We are a UI tax on OpenAI.”

12. Ask Them to Draw the Boundary

“Draw a line between what the LLM does and what your system does.”

Good founders:

• Explicitly separate:
• Reasoning
• Retrieval
• Guardrails
• Business logic
• Orchestration
• Post-processing

Wrappers:

• Handwave:
• “The LLM handles that.”
• “We use AI agents.”
• “We orchestrate tools.”

Every time you hear “agentic”, mentally replace it with “glorified prompt chain”.

Quick Checklist (Investor Mode)

Print this in your head:

• ❓ Can they survive 6 months without OpenAI/Anthropic?
• ❓ Do they have any real in-house ML talent?
• ❓ Is there a proper data + eval pipeline?
• ❓ Is their infra more than: frontend → API → LLM?
• ❓ Do they own anything you can’t recreate in 3 months with a dev and a credit card?

If the honest answer is “no” across the board → API wrapper.