Latest news with #PeterKellis
Forbes
05-08-2025
- Business
- Forbes
Scaling Production AI: 20 Surprising Hurdles And How To Overcome Them
Scaling AI from prototype to production is a major hurdle for many organizations. Even with skilled teams and powerful models, companies often hit unexpected bottlenecks—such as disjointed data pipelines, unclear ownership or low user trust—that slow or stall progress. To scale AI successfully, businesses need more than just technical expertise; they must invest in robust infrastructure, change management and cross-functional alignment. Below, members of Forbes Technology Council reveal common hurdles teams face when scaling production AI—and share practical strategies for overcoming them. 1. Instilling Digital Literacy And Trust Among Team Members A major bottleneck in scaling AI is organizational resistance due to fear and digital skill gaps. Many employees worry that AI will replace their roles or feel unprepared to use it. This often leads to low engagement and stalled adoption. To overcome this, companies must invest in AI literacy and upskilling. Involving staff early in the AI journey fosters trust, making scaling smoother. - Praveen Tomar, UK Civil Services (Ofgem) 2. Choosing The Right Problems To Tackle AI holds promise in enterprise restaurants—from predicting inventory to personalizing upsells. However, a common bottleneck isn't technical; it's misalignment between what's built and what drives value. Too often, teams chase flashy use cases without validating impact. Scaling AI starts with choosing the right problems and having the right data. - Peter Kellis, TRAY Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify? 3. Providing Frontline Access One key bottleneck is providing direct access for the frontline workers who would benefit from it. Providing frontline workers with AI solutions to efficiently and effectively complete their tasks is critical in today's landscape, where time is of the essence. - Mike Zachman, Zebra Technologies Corporation 4. Building Trustworthy Data Architecture Fragmented data architectures fragment business context. Without integrated enterprise knowledge, AI lacks organizational understanding. Success requires a trusted data architecture that's operationally simple yet preserves business meaning. - Louis Landry, Teradata 5. Avoiding Prompt Injection And Tool Misuse Prompt injection and tool misuse are critical bottlenecks. Attackers can trick LLMs into leaking sensitive data, performing unauthorized actions or amplifying these to widen impact. This slows deployment and can lead to costly breaches. You can mitigate risk with the same security best practices that you use elsewhere in the organization: least privilege, sandboxing, audits and adversarial testing. - Willem Delbare, Aikido 6. Managing AI-Generated Content Overload AI will generate more content than existing review and approval processes are designed to handle. Companies will need to decide how to proceed—they can either accept the risk, dramatically slowing adoption, or find and/or build tools to help them manage it. - Larry Bradley, SolasAI 7. Identifying And Managing ML Pipeline Cost And Performance Issues Companies often overlook MLOps when they take on AI-related development. ML pipelines can have very different cost and performance challenges than traditional development domains. Understanding the true unit cost of model performance is essential and can mean the difference between a model that scales well and one that doesn't. - Siamak Baharloo, Labviva 8. Linking AI Spend To Business Value One big bottleneck? The CFO calling out the inference cost line on the P&L. If teams can't clearly link AI costs to business value, there's a real risk that projects will stall. The solution? An AI Command Room. Start by building a clear business case tied to at least one normalized metric, such as labor dollars saved through AI agents and automation. Then, track and report impact—from hypothesis to daily execution. - Matt Kesby, Multiplai Tech 9. Establishing Shared Data Definitions As enterprises deploy agentic AI, agents make decisions based on data no one fully trusts. In pilots, humans catch issues. At scale, autonomous systems pull from disconnected sources with conflicting rules, and small cracks become major failures. The issue isn't the model; it's the lack of data trust. Scaling AI requires shared definitions and automated checks to ensure a reliable foundation. - Jay Limburn, Ataccama 10. Grounding Data In Real-Time, Trusted Context A key bottleneck in scaling AI is context fragmentation—models can't reason effectively without access to clean, connected and secure organizational data. Grounding data in real-time, trusted context is the real challenge. The fix is better infrastructure: unified knowledge layers, privacy-preserving compute and retrieval pipelines that align AI with the right data. - Regan Peng, PINAI 11. Addressing Data Lineage And Quality Debt One unexpected bottleneck is data lineage and quality debt. Many organizations assume that once a model is trained and performs well in testing, scaling it into production is mostly an engineering and computing problem. In reality, the biggest bottleneck often emerges from inconsistent, incomplete or undocumented data pipelines—especially when legacy systems or siloed departments are involved. - Sandeep Uthra, OneAZ Credit Union 12. Working With Outdated Infrastructure A bottleneck would be inadequate data infrastructure, which can be overcome by investing in modern, scalable data platforms and robust engineering practices. Companies should be establishing robust data governance practices and centralized data infrastructure early in the development lifecycle. - Nazih Chamtie, KMicro Tech, Inc. 13. Managing Model Drift Even though data quality is an obvious bottleneck, when we scaled AI in production, I primarily expected challenges with data and infrastructure. However, what caught us off guard was model drift. For example, in finance, market shifts rendered our fraud models stale. In healthcare, evolving patient data skewed predictions. Real-time monitoring and retraining turned out to be just as critical as building the models. - Dr. Suresh Rajappa, KPMG LLP 14. Overcoming Resistance To AI Integration Change management within the existing user experience is a significant challenge when scaling AI. While AI has enormous potential to enhance user experience, effective integration requires more than technical implementation. Organizations can proactively address this issue by increasing user adoption, reducing resistance and unlocking AI's full transformative potential. - Shivaprakash S Nagaraj, Digit7 15. Struggling To Find Skilled Architects And Operators Designing and building infrastructures that support AI initiatives requires a different set of skills. While many current IT professionals will get up to speed over time, there are significant nuances that come with AI that they may have never experienced before. Struggling to find skilled infrastructure architects and operators who already understand these nuances will be the reality for a few years. - Mike Wong, Accton Technology 16. Developing A Single Source Of Truth AI can be hindered by a lack of data access, data silos, data quality issues, missing context and so on. A lack of a single source of truth is also a common issue—especially if data connections are intermittent and complex—and AI models need to be run locally. Thinking strategically about the desired business outcomes and implementing a data fabric will improve data management and help scale AI end-to-end. - Heiko Claussen, Aspen Technology, Inc. 17. Identifying And Patching Security Gaps Security is the most significant challenge, because LLMs can't be controlled. Even foundational model providers' system prompts and training instructions have leaked. Strengthening AI security requires identifying potential vulnerabilities and testing against them, but we do not yet have an exhaustive list of those vulnerabilities, as AI systems remain a black box. - Rohit Kapoor, Tekmonks 18. Connecting AI Outputs To Decision-Making Processes A major bottleneck is the lack of integration between AI models and core business workflows. Even the best models fail to deliver value if their outputs don't connect to decision-making tools or processes. You can overcome this by embedding AI via APIs, aligning with existing systems and designing with end users to ensure seamless adoption and continuous feedback loops. - Motasem El Bawab, N3XT Sports 19. Detecting Performance Degradation Companies can build decent AI models but struggle to scale them due to poor production monitoring and deployment capabilities. Unlike deterministic code, where 2 + 2 always equals 4, AI models produce variable outputs, making it hard to detect performance degradation. Companies can overcome this by implementing robust MLOps pipelines with continuous monitoring, A/B testing, and automated alerts for model drift detection. - Mia Millette, Skyline Technology Solutions 20. Ensuring Organizational Alignment Accelerating AI initiatives without addressing data quality, readiness and strategic alignment will undoubtedly result in significant roadblocks. To scale successfully, organizations must prioritize use cases based on business impact, ensure data is properly prepared, and tightly align each initiative with clear, measurable outcomes. - Sean Nathaniel, DryvIQ
Forbes
11-07-2025
- Business
- Forbes
Why Adaptable POS Solutions Are Essential For Security And Compliance
Peter Kellis, Founder and CEO, TRAY. Businesses of all sizes face mounting challenges in maintaining security and adhering to shifting compliance regulations today. I recently wrote about Android-based point-of-sale (POS) systems, dispelling myths about their security and highlighting best practices for secure deployment and maintenance. Taking it a step further, let's discuss the broader challenges of cybersecurity and compliance, emphasizing the importance of adaptable systems, endpoint security, encryption, cloud security and regulatory navigation. Point-of-sale systems, once viewed as straightforward tools for processing transactions, now stand at the intersection of commerce, cybersecurity and regulatory scrutiny. For organizations, this convergence underscores the importance of deploying adaptable POS solutions that can keep pace with emerging threats and regulatory demands. A New Era Of Risk And Regulation The rise of digital commerce and integrated payment systems has brought unparalleled convenience and operational efficiency. However, it has also made POS systems prime targets for cybercriminals. The risks range from data breaches and ransomware attacks to malware specifically designed to exploit vulnerabilities in POS networks. Compounding these threats are dynamic compliance standards—such as PCI DSS, GDPR and CCPA—that require businesses to safeguard sensitive customer data while navigating jurisdiction-specific regulations. The complexity of these challenges is exacerbated by common misconceptions about security and compliance. One prevalent myth is that all POS systems offer a comparable level of protection out of the box. In reality, security capabilities vary widely, and a one-size-fits-all approach often leaves businesses exposed to unnecessary risks. The Case For Adaptability Adaptable POS solutions play a critical role in addressing these challenges. Flexibility in a POS system allows businesses to respond quickly to new threats and evolving regulatory requirements. Whether it's implementing tokenization to protect payment data, integrating new authentication methods or adhering to updated privacy laws, a modular and customizable POS architecture enables businesses to stay ahead of the curve. Adaptable systems also reduce the burden on IT teams by simplifying updates and configurations. For example, a system that supports remote updates ensures that security patches and compliance changes are applied uniformly across all locations, reducing the risk of human error and inconsistent policies. Key Security Considerations For Modern POS Systems POS security encompasses several layers, from endpoint protection and encrypted transactions to secure network configurations. Here are three key considerations: • Endpoint Security: POS terminals often operate as endpoints within a broader network, making them potential entry points for malware or unauthorized access. Regular software updates and robust authentication protocols are essential. • Data Encryption: Payment data must be encrypted both in transit and at rest to protect it from interception. Technologies like point-to-point encryption (P2PE) and tokenization can significantly reduce the risk of data breaches. • Cloud Security: Many modern POS systems rely on cloud-based infrastructures for data storage and processing. Evaluating the security practices of cloud providers, including data access controls and incident response capabilities, is critical. Building Resilience Through Partnership While technology is a crucial enabler of security and compliance, businesses cannot go it alone. Partnering with POS providers and technology vendors who prioritize security and regulatory adaptability is essential. These partnerships can provide access to expertise, tools and support to navigate a complex and ever-changing landscape. In addition to technical support, these partnerships often include training programs to educate staff on the latest security best practices. Employee awareness is a vital layer of defense against cyber threats, as human error remains one of the leading causes of data breaches. By investing in adaptable POS systems, businesses can protect their customers, safeguard their operations and maintain their competitive edge in an increasingly connected world. As cyber risks continue to grow, the ability to combine robust security features with adaptability will define the leaders in the modern commerce landscape. Specific to restaurants, the industry my company serves, these principles are especially critical. Restaurants process a high volume of sensitive customer payment data daily, making them attractive targets for cybercriminals. Adaptable POS systems allow restaurant operators to address industry-specific security challenges, such as multi-location compliance or integration with delivery platforms, while also ensuring data protection. By partnering with trusted providers and staying proactive about security and compliance, restaurants can build customer trust, avoid costly breaches and focus on delivering exceptional dining experiences in a competitive market. Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
