GenAI, particularly third-party GenAI, has several risks. It invents. Biased and poisonous. It may violate copyright. This study by the MIT Sloan Management Review and Boston Consulting Group found that third-party AI technologies cause over 55% of AI failures in enterprises.
It’s hardly unexpected that some firms are hesitant to use the technology yet.
What if GenAI was warranted?
Entrepreneur and electrical engineer Karthik Ramakrishnan had that company concept as a senior manager at Deloitte some years ago. After co-founding two “AI-first” firms, Gallop Labs and Blu Trumpet, he realized that trust and risk quantification were holding back AI adoption.
“Right now, nearly every enterprise is looking for ways to implement AI to increase efficiency and keep up with the market,” Ramakrishnan emailed Eltrys. Many rely on third-party suppliers and AI algorithms without fully understanding product quality. AI is progressing so quickly that dangers and effects are continually changing.”
Ramakrishnan founded Armilla AI alongside Dan Adamson, a search algorithm specialist and two-time company founder, to provide corporate clients with AI model warranties.
How can Armilla achieve this when most models are black boxes or locked behind licenses, subscriptions, and APIs? I had the same question. Ramakrishnan suggested benchmarking and deliberate customer acquisition.
Armilla uses worldwide AI regulatory data to “verify its quality” for open-source or proprietary models. Using its in-house evaluation system, the organization evaluates for hallucinations, racial and gender prejudice, fairness, robustness, and security across a variety of theoretical applications and usage scenarios.
If the model passes, Armilla’s warranty reimburses the customer for any usage fees.
“What we really offer enterprises is confidence in the technology they’re procuring from third-party AI vendors,” Ramakrishnan added. Businesses may ask us to evaluate suppliers. AI penetration testing is similar to new technology penetration testing.
Ramakrishnan was questioned about whether Armilla will not test models for ethical grounds, such as a face recognition algorithm from a provider with poor business practices. He said:
“To produce assessments and reports that give clients and society false confidence in AI models that are problematic is against our ethics and business model, which is based on trust. We won’t take on clients for models that are prohibited by the EU or have been banned, such as some facial recognition and biometric categorization systems, but applications that fall into the EU AI Act’s ‘higher risk’ category, yes.”
This writer was astonished to learn that AI warranties and insurance coverage aren’t new. Munich Re launched aiSure in 2018 to defend against losses from possibly faulty AI models using Armilla-like benchmarks. Many providers, including OpenAI, Microsoft, and AWS, offer copyright protections for AI tool deployments.
Ramakrishnan says Armilla’s method is unique.
He continued, “Our assessment covers KPIs, processes, performance, data quality, and qualitative and quantitative criteria, and we do it at a fraction of the cost and time.” “ We evaluate AI models based on the EU AI Act, NYC Local Law 144, and state regulations like Colorado’s proposed AI quantitative testing regulation or New York’s insurance circular on AI in underwriting or pricing. We’re also ready to analyze new legislation like Canada’s AI and Data Act.
Swiss Re, Greenlight Re, and Chaucer-backed Armilla commenced coverage in late 2023 and claims 10 clients, including a healthcare provider using GenAI to handle medical information. Ramakrishnan said Armilla’s customer base has grown twice every month since Q4 2023.
Ramakrishnan said there are two key audiences: corporations and third-party AI suppliers. Our guarantee protects third-party AI suppliers contracted by enterprises. Third-party suppliers utilize our guarantee to prove their product’s reliability, shortening their sales cycles.”
AI warranties make sense. But I worry whether Armilla can keep up with fast-changing AI laws (e.g., New York City’s hiring algorithm bias statute, the EU AI Act, etc.), which might leave it on the line for large sums if its evaluations and contracts aren’t impregnable.
Ramakrishnan ignored this.
He stated, “Regulation is rapidly developing in many jurisdictions independently, and it’ll be critical to understand the nuances of legislation around the world. No ‘one-size-fits-all’ global standard exists; therefore, we must sew it all together. This is difficult, but it gives us a’moat’.”
Toronto-based Armilla, with 13 employees, raised $4.5 million in a seed round led by Mistral (not the AI startup of the same name) and including Greycroft, Differential Venture Capital, Mozilla Ventures, Betaworks Ventures, MS&AD Ventures, 630 Ventures, Morgan Creek Digital, Y Combinator, Greenlight Re, and Chaucer. Ramakrishnan said the $7 million collected would be used to increase Armilla’s warranty and introduce additional goods.
Insurance will play the largest role in tackling AI risk, and Armilla is leading the way in designing insurance products to help firms use AI solutions securely, Ramakrishnan added.
Armilla offers AI warranties for enterprises.
