AI Due Diligence: Models, Data, GenAI Risk & Moats

Many targets now claim “AI-powered” growth, margin, or defensibility. AI due diligence is the work that decides whether those claims survive contact with model cards, data contracts, eval suites, inference bills, and customer liability. It is not the same as technology due diligence (stack and scalability), product due diligence (roadmap and fit), cybersecurity diligence (security posture), or SaaS metrics diligence (ARR/NRR). AI diligence underwrites the intelligence layer: what is proprietary, what is rented, and what can break the model at scale.

AI vs technology vs product vs cyber diligence

Six pillars of AI diligence

1. Model inventory & system map

Catalog every model in production and near-production: proprietary trained models, fine-tunes, embeddings, classical ML, third-party foundation APIs, and open-source weights. For each, document task, owner, latency SLA, fallback path, versioning, and whether it sits on the critical path of revenue or safety. Separate “AI in the deck” from “AI in the product path.” Map orchestration (agents, tools, RAG pipelines) so buyers see the full system, not a single model name.

2. Data rights, lineage & quality

Trace training, fine-tuning, evaluation, and retrieval corpora: customer data, partner feeds, licensed datasets, scraped web, synthetic data, and employee-created labels. Confirm contractual rights to train, improve, and commercialize; deletion and opt-out obligations; cross-border transfer limits; and whether licenses (including open-source dataset terms) conflict with the product. Assess data quality, labeling process, drift monitoring, and PII handling — and connect findings to data privacy diligence and IP diligence.

3. Evaluation quality & model risk

Demand reproducible evals: offline benchmarks, golden sets, human preference or expert review, online A/B or shadow tests, and task-specific metrics (not only generic leaderboard scores). For GenAI, test hallucination rate on customer-critical tasks, refusal behavior, jailbreak/prompt-injection resistance, and regression gates on model upgrades. Require model cards or equivalent documentation, known failure modes, and a change-management process when the provider ships a new base model.

4. GenAI product, safety & liability risk

Map where AI outputs reach customers, regulators, or automated actions without human review. Inventory safety incidents, customer complaints, content moderation, logging of prompts/outputs, and data leakage paths (training on tenant data, shared context windows, plugin tools). Align product claims with reality: copilots vs autonomous agents, accuracy warranties, and insurance implications (see insurance diligence). Red-team or at least sample adversarial prompts on high-risk flows.

5. Unit economics, vendors & scale path

Build cost-to-serve for inference, embedding, storage, GPU/CPU, evaluation, human-in-the-loop review, and vendor markups. Stress token or request growth vs pricing power. Test concentration risk on a single model provider, region, or GPU supplier and the switching plan (second provider, open weights, distillation). Separate gross margin impact of AI from marketing claims. For platforms selling AI features, verify packaging, usage caps, and whether AI is margin-accretive or a growth giveaway.

6. Moat, talent, IP & governance

Decide whether defensibility is proprietary data, distribution, workflow lock-in, model performance, brand trust, or none of the above. Map key ML/product talent, documentation quality, and bus factor. Review patents, trade secrets, open-source outbound obligations, and ownership of fine-tunes and outputs. Confirm AI governance: policies, risk tiers, board/IC reporting, regulatory horizon (sector AI rules, consumer protection, sector-specific model risk). Tie people risk to people diligence and compliance to regulatory diligence.

Stage sequencing (IOI to close)

Red flags

Cost & timeline (traditional vs first-pass)

50-point AI diligence checklist

• Written inventory of all production and near-production models
• Each model mapped to product feature and revenue criticality
• Owner, version, and rollback path documented per model
• Proprietary vs third-party vs open-source weights labeled
• RAG / agent / tool orchestration diagram available
• Training and fine-tuning data sources listed with lineage
• Contractual rights to train / improve / commercialize confirmed
• Customer data use and opt-out / deletion obligations mapped
• Cross-border data transfer and residency constraints known
• Licensed and scraped dataset licenses reviewed
• Synthetic data generation process and bias risks noted
• Labeling quality process and inter-annotator checks exist
• Offline eval suite with golden sets and pass thresholds
• Online / shadow evaluation or A/B framework in place
• Hallucination / error rates measured on customer tasks
• Prompt-injection and data-exfiltration tests performed
• Regression gates for provider model upgrades defined
• Model cards or equivalent documentation for material systems
• Known failure modes and human escalation paths written
• Where AI outputs reach customers without human review mapped
• Incident log for AI safety, accuracy, and customer complaints
• Logging retention for prompts/outputs and privacy controls
• Content moderation / abuse handling for GenAI surfaces
• Product marketing claims vs measured accuracy reconciled
• Inference unit cost and gross margin impact modeled
• Token / request growth scenarios vs pricing power stress-tested
• Embedding, storage, and evaluation costs included
• Human-in-the-loop cost in cost-to-serve included
• Primary model provider concentration assessed
• Second-source / open-weight switching plan documented
• GPU / cloud capacity constraints and commitments known
• Open-source outbound license obligations reviewed
• Ownership of fine-tunes, embeddings, and outputs clear
• Patent / trade-secret position and freedom-to-operate skim
• Material AI-related litigation or claims inventory
• Key ML / applied science talent map and bus factor
• Documentation quality sufficient for new hires
• AI governance policy and risk tiering exist
• Board / IC reporting pack for AI risk agreed
• Sector regulatory horizon (AI-specific rules) reviewed
• Insurance coverage for AI / cyber / media liability checked
• SPA data/IP/AI rep draft topics listed
• Escrow or source access plan for critical AI assets
• Day-1 access to model accounts, keys, and data stores
• No silent production model swap policy post-close
• Customer concentration of AI-heavy accounts understood
• Competitive moat narrative stress-tested (data vs wrapper)
• Integration plan if buyer will combine models or data
• Kill criteria documented for each material AI system
• Public comps / peer AI cost and performance cited where used

How deal teams use a first-pass pack

Before LOI, buyers use structured public research to pressure-test whether an AI story is even plausible: product claims vs demos, hiring and research signals, patents and open-source footprint, pricing and packaging, customer logos, and competitive density of similar wrappers. After LOI, the same hypotheses drive the data-room request list — model inventory, data contracts, eval reports, cost telemetry, talent map — so advisors do not spend weeks on marketing slides. The pack is screening research, not a substitute for model audits, red teams, or legal IP opinions.