Technology, Media & Healthcare
How to Underwrite an AI or Machine Learning Company
Underwrite an AI software company by separating production contracts from pilots, then testing gross margin after inference and model costs and the durability of its advantage over foundation model providers. Key metrics are gross margin after compute, revenue in production deployments, net retention once pilots convert, and rights to the training and customer data that differentiate the product.
What makes ai / ml software underwriting different
AI and machine learning software companies sell models, applications built on third-party foundation models, or tooling for teams that build their own. Value comes from proprietary data, workflow integration that makes the product hard to remove, and outcomes customers can measure, such as reduced labor hours or higher conversion. Risk comes from inference costs that scale with usage and compress margins, dependence on a model provider that can change pricing or compete directly, rapid commoditization as base models improve, and revenue that is still in paid pilots rather than committed production contracts. Data rights and regulatory exposure add legal risk that typical software diligence may miss.
Lenders treat most AI companies like other venture-backed software: they lend against committed recurring revenue and equity support, with liquidity covenants, and they discount pilot revenue and usage that may not renew. They pay close attention to compute commitments, which can behave like debt. Equity investors underwrite defensibility and growth: whether the company owns data or distribution that model providers cannot replicate, how quickly pilots convert to enterprise contracts, and whether gross margin will improve as models get cheaper. They pay for momentum but increasingly ask what remains if a foundation model provider ships a similar feature.
Key metrics
Gross margin after inference and compute
Every query or job consumes compute or API spend, so margins can be far below traditional software and vary with usage patterns.
Production vs pilot revenue
Paid pilots and proofs of concept often do not convert, so only committed production revenue should be treated as recurring.
Pilot conversion rate
Shows whether the product delivers enough value to move from experiment to budgeted line item.
Net revenue retention
Expansion within customers that deploy AI across more workflows signals real adoption.
Model provider concentration
Reliance on one foundation model API creates pricing, availability and competitive risk.
Compute commitments vs revenue
Multi-year cloud or GPU commitments can exceed revenue if growth slows and function like fixed obligations.
How to underwrite it, step by step
- Classify revenue into production subscriptions, usage, paid pilots and services, and reconcile to contracts.
- Calculate gross margin by product after inference, API, hosting and human review costs.
- Review dependence on foundation model providers, including contract terms, pricing changes and fallback options.
- Diligence data rights: sources of training data, customer data usage permissions, and licensing of third-party data.
- Assess defensibility: proprietary data, integrations, workflow depth and evaluation results against base models.
- Review compute commitments, GPU leases and cloud agreements as fixed obligations.
- Evaluate regulatory, privacy and IP exposure, including output liability and industry-specific AI rules.
- Model revenue and margin under lower model prices, competitor entry and slower pilot conversion.
Due diligence checklist
- Customer contracts distinguishing pilots from production agreements
- Revenue by customer, product and deployment stage
- Inference and compute cost reports by product
- Foundation model provider and cloud agreements
- GPU leases and compute commitment schedules
- Training data inventory with licenses and consent records
- Customer data processing agreements
- Model evaluation and benchmark results
- AI governance, privacy and security policies
- Patent filings and IP assignment agreements
- Litigation or claims involving data use or model outputs
Key risks and mitigants
Inference costs compress margins
Analyze unit costs, pricing power and model routing options, and underwrite to current rather than hoped-for margins.
Foundation model provider competes or reprices
Assess multi-model capability and proprietary layers, and stress margins for higher API prices.
Pilots fail to convert
Count only production revenue toward debt sizing and track pilot conversion in reporting.
Data rights or IP challenge
Review data provenance and licenses, obtain legal analysis, and require indemnities for data claims.
Compute commitments exceed demand
Treat commitments as fixed charges in coverage analysis and review termination or resale rights.
Typical financing and deal structure
Most AI software companies are equity financed, often with large rounds relative to revenue. Venture debt is common once there is committed recurring revenue, sized relative to equity raised and runway, with warrants and minimum liquidity covenants. Companies with large GPU needs may use equipment financing or leases secured by GPUs, though lenders apply steep depreciation assumptions given rapid hardware cycles. Profitable AI-enabled software businesses can access recurring revenue loans or EBITDA-based term loans. Acquisitions by larger technology companies often include significant retention packages for researchers and engineers, earnouts tied to revenue or product milestones, and specific indemnities covering data rights and IP.
Red flags
- Most revenue from paid pilots with short terms and no committed renewals
- Gross margin falling as usage grows
- The core product is a thin interface on a single third-party model
- Training data scraped or licensed without clear rights
- GPU or cloud commitments larger than projected revenue
Frequently asked questions
How are AI startups valued?
AI startups are usually valued on revenue multiples based on growth, retention and gross margin, with heavy emphasis on defensibility. Investors discount pilot revenue and companies that depend on a single model provider. Strong proprietary data, deep workflow integration and improving gross margins typically support higher valuations.
Why do AI companies have lower gross margins than SaaS?
AI products consume compute every time they run, through GPU usage or model API fees. These costs scale with usage, unlike traditional software where incremental users cost little. Margins can improve as models become cheaper or as companies optimize routing and caching, but underwriters base projections on demonstrated cost trends.
Can an AI company get venture debt?
Yes, if it has committed recurring revenue and strong equity backing. Venture lenders size loans relative to recent equity raised and cash runway, and they discount pilot revenue. Loans typically include warrants, minimum cash covenants and monthly reporting. Companies with large compute commitments may face tighter terms.
What data rights matter in AI due diligence?
Buyers and lenders review where training data came from, whether licenses permit commercial model training, and whether customer contracts allow use of customer data to improve models. Gaps can lead to litigation or require retraining. Diligence findings often result in specific indemnities or price adjustments.
Related terms
Underwrite a ai / ml software deal in minutes
Drop a deck or data room. Underridge returns a verdict, investment memo, term sheet and a live model, with every headline figure traced to a page, a formula or a labeled assumption. Ten underwrites free.
Try Underridge freeEducational content, not investment, credit, legal or tax advice. Benchmarks are common practitioner rules of thumb and vary by lender, market and cycle.