AI-washing: separating signal from noise
Private capital teams are seeing AI everywhere in decks, but few claims survive evidence testing. This piece lays out a 30-day diligence sprint to validate data rights, real model advantage, adoption feasibility, and delivery economics.
The question behind this piece
Every target now claims AI differentiation. Many cannot prove data rights, repeatable model performance, or scalable adoption. Traditional diligence may check architecture and feature lists, but miss AI-specific failure modes that drive post-deal disappointment. How can an investor validate AI differentiation in 30 days with evidence tests that predict real outcomes after close?
Why this matters now
AI-native has become a default sales posture. Tools are easier to access, teams can add AI features quickly, and marketing can get ahead of what the product can consistently deliver. Over-claiming is often unintentional, but the effect is the same: the deal thesis rests on assumptions that were never tested.
The economics can also surprise. AI products often carry hidden costs: data labeling, model iteration, compute, monitoring, and customer success capacity to drive adoption. A business that looks software-like can behave like a services-heavy delivery model underneath.
Defensibility is also more fragile than many decks suggest. If advantage depends on third-party APIs with shifting terms, uncertain training rights, or weak data retention, the moat may be smaller than it appears.
In AI diligence, the moat is usually data rights plus adoption, not the model.
Our perspective
AI-specific diligence should run as a sprint with four evidence tests: data rights, model advantage, adoption feasibility, and delivery economics. The goal is not a prettier thesis slide. The goal is a truth set: what is defensible, what is fragile, and what must be built post-close.
Test 1: Data rights and data quality
Ask for evidence, not explanations. Validate what is owned, licensed, or derived, and confirm contract terms. Sample customer agreements to determine whether data can be used for training, fine-tuning, and evaluation, and under what constraints. Clarify retention rights after churn and what anonymization actually allows in practice. Then assess labeling, coverage across segments, and measurable data quality, not anecdotal confidence.
Test 2: Model advantage and repeatability
Avoid benchmark theater. Run targeted tests on representative, messy data. Use holdout evaluation, segment-level performance (especially edge cases that drive customer pain), and robustness tests for drift and changing conditions. Where the product influences regulated or high-stakes decisions, test explainability and evidence requirements, and whether the system can produce defensible rationale consistently.
Test 3: Adoption feasibility and workflow fit
AI value is realized in daily work, not demos. Validate integration into customer workflows and systems, who the user is, what changes, and where friction remains. Look for sustained usage signals: retention cohorts, repeat behavior, and renewal dynamics. Quantify the “human glue” required: customer success load, customization, exception handling, and ongoing tuning to keep value delivered.
Test 4: Delivery economics
Many AI businesses hide a services and compute curve. Model compute cost per unit of value and how it scales with usage. Quantify support and escalation volume, including the cost of exceptions and investigations. Include monitoring and retraining costs with realistic staffing assumptions. Measure implementation burden per customer and time-to-value. The question is simple: do margins improve with scale, or collapse?
There are predictable failure modes worth testing explicitly: data rights that do not permit training or vary materially by customer; models that perform in curated examples but fail on edge cases and drift; adoption that depends on heavy customization or ongoing services; and gross margin that degrades as usage grows due to compute and support.
A 30-day sprint can be structured cleanly:
The output should be an investor-ready truth set: what is defensible, what is fixable, what is fragile, and the post-close plan required to achieve the deal thesis. Done well, diligence becomes a value creation blueprint, not a checklist.
The best diligence produces a build plan, because most AI differentiation must be earned post-close.
What we offer and how we can help
Diagnostic: 30-Day AI Differentiation Diligence Sprint.
Strathen Group can works alongside your deal team and run a 30-day AI Differentiation Diligence sprint. It will help you validate AI claims with evidence tests and translate findings into an actionable post-close plan. Typical output from such a spring can include:
- AI differentiation scorecard across the four tests, with red flags and confidence levels.
- Data rights and retention map based on contract sampling, with practical implications for training and moat.
- Model evaluation readout (holdout results, segment performance, robustness risks, and required fixes).
- Adoption feasibility assessment (workflow fit, integration burden, retention signals, and “human glue” costs).
- Delivery economics model (compute, support, monitoring, implementation burden, and margin trajectory).
- Value creation blueprint and risk register with prioritized initiatives for the first 100 days and first year.
If you are underwriting an AI-heavy target and want conviction based on proof, Strathen Group can help.
