Artificial intelligence has become increasingly embedded in everyday services, and the conversation is shifting from what AI can do to whether it can be trusted to do it reliably, safely and responsibly. This is even more important in legal technology, where accuracy, transparency and accountability are essential when helping individuals navigate complex legal matters. Building AI systems that users can genuinely trust requires far more than deploying powerful language models; it demands rigorous engineering, robust governance frameworks, privacy safeguards and continuous oversight throughout the product lifecycle.
In this exclusive interview with TechBullion, Haytham Allos, Chief Technology Officer at Vikk AI, shares his perspective on the technical and ethical challenges of developing trustworthy, customer-facing AI for regulated industries. He discusses the practical realities of reducing AI hallucinations, implementing effective governance, protecting user privacy and creating resilient AI systems that deliver consistent, dependable outcomes. Haytham also explains why engineering discipline, rather than model capability alone, will define the next generation of enterprise AI applications and what organisations must do to deploy artificial intelligence responsibly at scale.
Please introduce yourself to our readers and tell us about your role as Chief Technology Officer at Vikk AI?
My name is Haytham Allos, and I’m the co-founder and CTO of Vikk AI. I have about 25 years of experience across software engineering, cloud computing, cybersecurity, AI, and blockchain. My main responsibility here is to serve as the technology strategist, overseeing architecture, engineering, security, and compliance. Vikk AI is a consumer product that helps people access and understand legal information, so my role is really about balancing innovation with accountability. I need to make sure our AI is capable and enjoyable to use, but also responsible and secure in this regulated environment.
Artificial intelligence is becoming increasingly accessible to consumers, particularly in legal services. What inspired Vikk AI to focus on this space, and what problems are you aiming to solve for individuals seeking legal guidance?
Legal information is difficult to understand, intimidating, and expensive to access. Many consumers don’t initially know what type of legal issue they have or what their next steps should be. Vikk AI helps users describe their situations in everyday language, identify the relevant legal category, and organize their thoughts.
We’re not here to replace attorneys, as they have a huge and important role. We’re here for that first mile: to triage and help consumers through their legal issues. We provide assistance, not legal guidance, and we’re careful to stay on the right side of unauthorized practice of law.
Trust is one of the biggest challenges facing AI today. From your perspective as CTO, what are the essential technical and organisational foundations required to build AI systems that users can genuinely rely on?
Trust is built on several foundations. You need clear boundaries and users need to know what the AI can and cannot do. You need evidence-based, reliable responses. And you need the system to be tested, private, and accountable.
For us, that means clearly defining what our AI can do, ensuring responses are based on reliable information, and protecting sensitive data throughout the entire customer journey. Trustworthiness isn’t just a feature you add. It’s embedded in our controls and our culture. We continuously monitor production, maintain clear escalation paths, and keep evaluating what trust means in AI.
AI hallucinations continue to attract significant attention across many industries. Why do you believe they should primarily be viewed as an engineering challenge rather than simply a limitation of large language models?
Hallucinations are a characteristic of generative models, but here’s the thing: when we design products, we determine how much users are exposed to them. As engineers, we have control over the models, prompting, guardrails, and how we ground the AI.
Not every question needs to generate an answer. In a well-designed system, especially in regulated environments, we express uncertainty when it exists and ask for clarification. At Vikk AI, we narrow responses by asking about jurisdiction, which lets us route our knowledge base appropriately and reduce hallucinations. So hallucinations are really an engineering and product design challenge, not just a limitation of the model.
Legal technology demands a particularly high standard of accuracy and accountability. How do you approach designing AI systems that can provide useful assistance while recognising the boundaries of their capabilities?
One of the first things we did was distinguish general legal information from individualized legal advice. AI shouldn’t be a substitute for attorneys. Rather it should be on-demand assistance for initial guidance.
We use clear disclosures, recognize urgent issues and emergencies, and maintain firm boundaries around the system. This ensures users understand both the value and the limitations of our AI.
Many organisations are rushing to integrate generative AI into customer-facing products. What common engineering mistakes do you see businesses making when deploying AI into production environments?
I’ve seen, and have made several mistakes. First, relying on just one model or prompt without broader controls. You need layered safeguards. Second, testing only normal cases and missing edge cases and adversarial behaviors. Third, giving agentic systems too much autonomy without guardrails.
Another big one: collecting more data than you need. We deliberately limit sensitive information to only what the AI needs. And finally, launching without clear accountability. You need incident response procedures, monitoring, and the ability to roll back quickly if something goes wrong.
Privacy is particularly important when dealing with sensitive legal information. How do you balance the need for intelligent AI capabilities with the responsibility to protect users’ personal and confidential data?
The sequence is: minimize, protect, and govern. First, collect only what you need. For us that’s first name, last name, and email. Then use encryption, access controls, and retention policies. Make sure you’re audited and compliant with standards like SOC 2, GDPR, and HIPAA.
The whole goal is to not give AI access to everything. Give it only the minimum information required to perform its task safely. That’s how you balance capability with privacy.
Beyond selecting powerful AI models, what engineering practices play the greatest role in improving reliability, consistency and user confidence in AI-powered applications?
Implement input and output guardrails, version your prompts, and do regression testing. For us, compliance and privacy are deeply ingrained because we’re in the legal field. You also need defense-in-depth and multiple layers of safeguards so a single failure doesn’t compromise everything.
Make sure you can recover from failures, maintain low latency, and monitor for unusual behavior. It’s about building an ecosystem where your model selection, implementation practices, and safeguards all work together.
Continuous testing has become an important part of modern software development. How does AI testing differ from traditional software testing, and what additional challenges does it introduce?
Traditional software is deterministic, with expected inputs and outputs and clear pass/fail criteria. AI is probabilistic. The same question might generate different but equally valid answers.
When testing AI, you need to evaluate accuracy, relevance, safety, privacy, tone, and consistency. That requires expert judgment, not just simple pass/fail assertions. You’re asking: are the responses safe, relevant, and consistent with our boundaries? That’s different from conventional testing, and it requires more work.
Looking ahead, which emerging technologies or engineering practices do you believe will have the greatest impact on making AI more trustworthy and dependable over the next five years?
There’s a lot of movement in creating guardrails around AI systems bounded by safety and privacy. With agentic AI, new protocols are being developed to embed defense systems so agents don’t do something misaligned with consumer safety. And there’s evolution toward more controlled, secure models for mission-critical applications.
It’s really amazing how the field is moving toward more secure and controlled approaches to AI.
For organisations planning to introduce AI into highly regulated sectors such as legal services, healthcare or financial services, what practical advice would you offer before they begin their AI transformation journey?
My best advice is to start narrow. Define a clearly defined problem, not a general desire to use AI. Classify the risks involved and make sure you have proven controls before expanding.
Involve the right people like engineers, security personnel, and domain experts. Understand what happens when your AI goes wrong. Have clear acceptance criteria, start with limited pilots, and only expand when the evidence supports it. This measured approach helps you navigate the complexity while managing risk appropriately.
