Many decision-makers think AI risk is theoretical or too technical to probe deeply. But every missed question can turn into a missed opportunity—or worse, a silent liability. It’s time to separate what sounds true from what’s actually true with a proper breakdown of myths and facts around AI due diligence.
Myth 1: Due diligence in AI is just regular tech diligence with a few extra questions.
Fact: It’s a different beast altogether.
Standard technical evaluations often focus on code quality, infrastructure, and scalability. But due diligence in AI goes further—it interrogates datasets, model assumptions, bias mitigation strategies, and retraining workflows. For instance, a Head of Product might show off an accurate model—but how is it maintained? What’s the rate of model drift? What if it’s accurate but ethically problematic? Failing to ask these questions turns a powerful tool into a future liability.
Myth 2: You only need due diligence for AI if you’re buying an AI company.
Fact: If a product uses AI—even if it’s not the core product—you need it.
A corporate venture team once assessed a healthtech platform that claimed to use machine learning only for appointment recommendations. They skipped deep AI evaluation because the core functionality wasn’t AI. Months later, users flagged alarming recommendation patterns—patients were being deprioritised based on demographic assumptions baked into training data. If AI due diligence had been performed, the team would’ve known the model was using outdated datasets and no fairness filters.
Myth 3: If the AI works, it doesn’t matter how it works.
Fact: Explainability is not optional—especially in regulated industries.
The Head of Risk at a fintech firm once greenlit a credit scoring algorithm that outperformed all previous models in raw prediction. The problem? No one could explain the output to regulators or customers. When a regulatory audit came, the firm had to roll back the system entirely. Due diligence in AI would’ve flagged the model’s “black box” nature and forced conversations around interpretable alternatives. Accuracy without explainability is not a win—it’s a lawsuit waiting to happen.
Myth 4: Open-source AI tools reduce the need for diligence—they’re already vetted.
Fact: Open-source does not mean risk-free.
A Head of Engineering proudly integrated a popular open-source vision model into their SaaS product. Within months, the company faced a takedown request because the model had been trained on copyrighted datasets. The open-source license did not cover the training material. AI due diligence would have uncovered licensing inconsistencies and prompted a legal review before integration. Just because a tool is widely used doesn’t mean it’s safe—or suitable—for commercial deployment.
Myth 5: AI bias is a moral issue, not a business issue.
Fact: Bias costs money—and reputation.
An HR analytics firm deployed a candidate screening model that disproportionately favoured applicants from specific backgrounds. A sharp-eyed Head of Diversity flagged the issue, but not before social media backlash triggered a wave of lost contracts. The model was “accurate” in terms of historic hiring data—but accuracy based on biased patterns simply replicates past exclusion. Due diligence in AI ensures your models not only work but align with your company’s legal, social, and strategic commitments.
Myth 6: AI risk is mostly theoretical—it hasn’t affected us yet.
Fact: Absence of failure doesn’t mean absence of risk.
Many AI systems operate invisibly for long periods—until they don’t. A Head of Customer Success once learned this the hard way when an automated support classification model suddenly started misrouting high-value client tickets. It had drifted over time due to seasonality and shifting vocabulary. No one had monitored it. AI due diligence includes assessing how ongoing performance is tracked and what failsafes are in place.
