This article was sponsored by NLP Logix and was written, edited, and published in accordance with NLP Logix policies. Emerj Sponsored Content Guidelines. Learn more about our thought leadership and content creation services. Emerj Media Services Page.
Government and industry leaders increasingly agree that governance now matters. basic Not AI option Because generative and predictive systems already shape key decisions in the public sector.
Guidance on generative AI from the Colorado Department of Information Technology shows why. Nearly a quarter of organizations reported inaccurate output and 16% reported cybersecurity issues, highlighting how implementation can outpace governance.
A recent OECD report argues that fragmented data, legacy systems, and weak impact measurements often keep government AI in pilot programs. The report goes on to argue that governance needs to define accountability and measurement early on.
NLP Logix defines AI governance across ethics, policy, and testing. In practice, such a policy means documenting models, forcing human review in sensitive workflows, and performing standardized bias/robustness testing before and after deployment. This perspective positions governance as both risk management and enabling scalable and trustworthy AI.
In a special series sponsored by NLP Logix, Emerj editorial director Matthew DeMello speaks with Naveen Kumar, head of insider risk, analytics and detection at TD Bank, Matt Berseth, co-founder and CIO of NLP Logix, and Russell Dixon, strategic advisor at NLP Logix, to explore how organizations can effectively deploy AI tools, balance innovation and governance, and measure real business impact.
Their argument emphasizes that AI efforts stall when control, training, and measurement are treated as consequentials. This article analyzes three core insights for successful AI adoption, with a focus on robust governance, measurable business outcomes, and strategic deployment.
- AI governance as a built-in control layer: Enforce role-based access, strict data classification, phased rollouts, and mandatory human oversight for more secure deployments.
- Plan, manage, train, and measure AI. Deploy AI tools with a clear strategy, defined use cases, upfront governance, user training, and measurable adoption to ensure effective outcomes and ROI.
- Power your strategic planning and metrics for AI success. Plan your AI deployment with clear goals, metrics, and usage tracking to prevent tool creep and drive measurable value.
AI governance as an embedded control layer
Episode: Managing AI for Fraud, Compliance, and Automation at Scale – Co-authored with Naveen Kumar of TD Bank
guest: Naveen Kumar, Head of Insider Risk, Analytics and Detection, TD Bank
Expertise: Regulatory compliance, fraud and threat detection
Easy recognition: Naveen has over 16 years of experience in AML, insider risk, fraud and sanctions. Previously, he worked at PwC and Stellaris Health Network. She holds a Master of Science in Data Modeling from Rochester Institute of Technology.
In an interview, Kumar argues that AI governance starts with traceability, knowing what data is being used, who has access to it, and how AI interacts with it.
“I think role-based AI is like a polite bodyguard. It only provides information based on the role. If there’s an insider investigation going on, the financial institution doesn’t know anything about it. You feed it into the AI and it shouldn’t give you anything back. Guardrails are invisible forces. These are the It’s a rule that you can never break, no matter what prompts you receive. It prevents people from asking a series of questions or revealing things that the attacker shouldn’t know.”
– Naveen Kumar, TD Bank Insider Head of Risk, Analytics and Detection
He describes balancing innovation with customer obligations and constraints around regulatory and security compliance as an exercise that requires significant time and deliberate trade-offs. His recommendation is a gradual rollout, starting with a narrow range of use cases and minimal data access, and expanding privileges and sources only after controls are established.
Classification also plays a central role, with Naveen recommending that leaders label data as safe, sensitive, and critical, and exclude important data from early iterations. In his view, this structured, step-by-step approach helps organizations navigate the tension between usefulness and risk.
He also emphasizes that how The use of AI is highly domain-dependent, with a clear distinction between compliance and retail use cases. On the retail side, where the goal is to acquire customers, it may make sense to use AI more aggressively. However, compliance requires the opposite approach. This means organizations need to be much more conservative.
Kumar uses the example of a suspicious activity report to explain that while AI can support the process, it should not be run end-to-end without human review.
The challenge is to balance automation and monitoring. To address this, Naveen suggests thinking in terms of speed and accuracy. Automate low-risk alerts and route high-risk cases to human reviewers. Ultimately, he says, the right balance will depend on your domain and use case. In some situations, AI should be positioned as an efficiency layer or first draft rather than a fully autonomous end-to-end solution.
Next, we provide a series of practical steps to advance AI in a controlled manner. We recommend starting with a secure sandbox, building a complete inventory of internal and vendor AI, and engaging early with compliance. The goal is to gain visibility into what models exist, what data is accessed, and how it is managed.
Plan, manage, train, and measure AI to realize ROI
Episode: Making Microsoft Copilot and ChatGPT Enterprise Work for You – Co-authored with Matt Berseth and Russell Dixon of NLP Logix
guest: Russell Dixon, NLP Logix Strategic Advisor
Expertise: Technology innovation, business transformation, information technology
Easy recognition: Dixon is a strategic advisor at NLP Logix, specializing in global operations and business transformation. With over 20 years of experience in the information technology field, he advises organizations on implementing AI solutions and cloud technologies. Russell’s expertise includes enterprise sales and business automation with a focus on identifying high-value use cases to drive ROI.
During his podcast appearance, he argued that tools like ChatGPT and Microsoft Copilot are nearly universally applicable, but are only effective if implementation includes training, guardrails, and measurement of adoption and productivity.
Without that structure, Russell warns that simply releasing AI tools into an organization will not yield results or ROI. Instead, users may become dissatisfied and look for alternatives, or worse, conclude that the tool has no real value.
Therefore, governance must be defined before deploying AI tools. He argues that without realistic use cases, deployment plans, and user training strategies, organizations won’t get the results they want.
“There’s also a governance issue here: How will we use this tool? What guardrails will we put in place around it to ensure internal and client data is protected? Finally, we need to ask ourselves how we will measure productivity: Will we rely on user feedback, or will we introduce more formal measurement tools and processes along the way to measure usage as the project unfolds?”
– Russell Dixon, NLP Logix Strategic Advisor
For Dixon, the success of an AI project is closely tied to how well the use cases are defined. The more common your use case, the higher your chances of success. For example, implementing a tool like Copilot or ChatGPT to support general workplace productivity should come with some pretty high expectations, especially if the goal is to significantly increase productivity across common office tasks.
In contrast, very specific use cases carry greater risk, he says. The narrower the scope of the solution, the more likely it will not produce the desired results. He agrees with colleague and fellow podcast guest Matt Barseth, co-founder and CIO of NLP Logix, that it’s reasonable to aim for a success rate of around 80%, and that some failure is to be expected and necessary when organizations drive innovation.
But Russell emphasizes that early signals are important. If implementation and results are not immediate, organizations should pause and reevaluate. In his view, technology itself has potential. When a project fails, the root cause is more likely to have to do with user behavior or a mismatch between the tool and use case, rather than AI limitations.
Strengthen strategic planning and metrics for AI success
Episode: Making Microsoft Copilot and ChatGPT Enterprise Work for You – Co-authored with Matt Berseth and Russell Dixon of NLP Logix
guest: Matt Berseth, Co-Founder and CIO, NLP Logix
Expertise: AI, data science, software engineering
Easy recognition: Berseth is the co-founder and CIO of NLP Logix, where he leads the delivery of advanced machine learning solutions for industries such as healthcare, logistics, and finance. He has over 20 years of technical leadership experience and previously held engineering and architecture roles at Microsoft and CEVA Logistics. He is an adjunct professor and holds a master’s degree in software engineering from North Dakota State University.
Matt explains that successful AI deployments are not just deployed and counted, but are measured, deeply understood, and continuously enhanced. He distinguishes between adoption and value, saying usage patterns are more important than license counts and recommends combining user feedback with telemetry about who is using the tool, how often, and in what workflows.
Recruitment, he says, is just a “body of tools.” What really matters is usage patterns: how different users, teams, and departments are applying AI to high leverage.
He also explains what can go wrong without proper planning and warns of a new phenomenon commonly referred to in development circles as “tool creep.”
“I think what’s happening now is tool creep. These tools become one of those things that you don’t know how to use. You buy licenses, but you don’t see the value in them. At home, I use ChatGPT and I like the interface better, but at work, I don’t want to learn new tools that keep changing. The real problem is that these tools become enterprise AI It’s something you need to see as a strategic part of your strategy. You need a plan, clear goals, metrics, and a way to drive adoption across the organization. If you do, you’ll get there. If you don’t, you’ll be back in three or six months to fix a rollout that started off in the wrong direction.”
– Matt BarsethCo-Founder and CIO of NLP Logix
In contrast, he emphasizes that a certain amount of failure is necessary to promote innovation. Almost 80% of his team’s AI POCs have reached production and stayed for a year, demonstrating that the technology is capable. When a project struggles, the problem is often not the tool, but a poorly selected use case. Today, with accessible AI like ChatGPT, creating value is easier than ever. Organizations just need to choose the right problem to solve.
Barseth argues that when organizations claim a 100% proof-of-concept success rate, they may actually be avoiding risk rather than driving innovation. When a team tests new ideas, some failure is expected and even desirable.
On the governance side, Matt reiterates that successful AI implementation requires a structured plan, clear goals, and defined metrics. To ensure responsible, effective, and measurable use of AI tools such as ChatGPT and Microsoft Copilot, he emphasizes that organizations need to strategically select use cases, monitor implementation, and track both tool-level and business-level results.
