DueDig

Deep AI Due Diligence

For investors backing companies that buildAI. When the thesis rests on the model itself, the question is simple: is the AI a genuine, defensible advantage - or a thin wrapper around someone else's? We answer that with evidence, across the full model lifecycle.

~40%

of European 'AI' startups showed no evidence of material AI use (MMC Ventures, State of AI, 2019)

>80%

of AI projects fail - about twice the rate of non-AI IT projects (RAND Corporation, 2024)

>280x

fall in the cost to run a GPT-3.5-quality query in ~18 months - model access is commoditizing (Stanford AI Index, 2025)

50-60%

typical AI gross margins, vs. 70-90% for classic SaaS (a16z, 2020)

Is the AI defensible?

Defensibility - moat or wrapper?

  • Whether core value depends on proprietary models, proprietary data, or just a public API
  • Strength of any proprietary data advantage and how it compounds with usage
  • Real switching cost for a competitor to replicate the capability
  • Model-provider concentration and lock-in - exposure to their pricing, terms, and rate limits

IP & data provenance - who owns it?

  • Ownership of models, weights, prompts, agents, and fine-tuning datasets
  • Provenance and licensing of training data, including open-weight licence obligations
  • Copyright and IP exposure arising from training-data sources
  • Rights to use customer data for training, and the consent and privacy basis behind it

Is the ML real?

Datasets & data pipeline

  • Sourcing, curation, labeling quality, and deduplication
  • Train/eval contamination and leakage checks
  • Coverage, representativeness, and known gaps
  • PII/safety filtering and reproducibility of the data pipeline

Training & alignment

  • Base-model strategy: pre-train from scratch vs. adapt an existing model - and whether the spend/claims are credible
  • Supervised fine-tuning (SFT) data and method; parameter-efficient tuning (LoRA/QLoRA)
  • Preference optimization & alignment: RLHF (reward model + PPO), RLAIF, DPO
  • Whether the team has the preference data and infra to sustain a custom-model story

Evaluation & metrics

  • Offline benchmarks and task-specific metrics
  • Hallucination/factuality and safety evaluations
  • Human eval rigor and inter-rater agreement
  • Regression testing / CI for models, plus online metrics and A/B testing

Inference, cost & MLOps

  • Serving stack, latency, and quantization/optimization
  • Cost per request/token and how it scales - direct margin impact
  • Experiment tracking, data/model versioning, and retraining pipelines
  • Drift and production monitoring, and the data-flywheel/feedback loop

Deep AI Due Diligence is a specialist engagement, scoped to the target's stack and stage. Findings land in a board-ready report and red-flag list. Book a call and we'll scope it with you.

Investing in a company that simply uses AI? See SaaS Due Diligence.