Few IT executives feel the pace of developments in artificial intelligence (AI) as acutely as Manu Narayan. Some nine months into his role as the first chief information officer (CIO) at GitLab, the software development platform that has surpassed $1 billion in revenue and employs over 2,000 people, Narayan is tasked with turning the company into a proving ground for the very technologies its customers use. The challenge is immense, given that AI is evolving faster than any previous technology wave, requiring constant recalibration of strategies.

“The AI space in general is changing so rapidly that we’ve constantly had to revisit our goals and things that we want to accomplish,” he said in a recent interview. Narayan’s mandate is mostly internal: modernizing the business application stack, user support, as well as data and analytics. But instead of bolting AI onto existing workflows, his goal is to rebuild operations from the ground up.

“When I was revisiting our AI strategy a few months ago, the focus was not on how we introduce AI,” he explained. “The focus was to rethink the nature of work internally, leveraging AI. It’s thinking about processes from first principles and then using agentic AI to drive them.” This first-principles approach is central to GitLab’s AI transformation, separating it from companies that merely add AI features to legacy processes.

Rejecting ‘tokenmaxxing’

As AI adoption increases across the enterprise, CIOs naturally grapple with cost control and measurement. However, Narayan is wary of strategies such as “tokenmaxxing,” where developers and employees are encouraged to maximize the number of AI tokens they use. This gamification tactic has been promoted by some AI vendors as a way to drive usage, but Narayan sees it as fundamentally flawed.

“We’ve specifically avoided and don’t want to do tokenmaxxing,” said Narayan. “Gamification can help drive outcomes, but I think it drives the incorrect behaviour. We’re not looking for purely context-in, context-out as the measure of success. It’s really hard to know if somebody’s gaming the system. Are they just sending excessive content because they don’t actually know what they’re doing?”

Instead of tracking token burn, GitLab tracks daily active usage across its tech stack to ensure its workforce is building sustainable habits. For calculating hard return on investment (ROI), Narayan insists on anchoring AI deployments to traditional business metrics. For an AI agent assisting a sales development representative, success isn’t measured by the number of prompts generated, but by standard key performance indicators: outbound messages, meetings scheduled and sales pipeline conversion. This approach aligns with broader industry recognition that AI must deliver measurable business outcomes rather than vanity metrics.

The concept of tokenmaxxing emerged from the world of large language models (LLMs), where usage is often billed per token. In some organizations, leaders have encouraged employees to use AI for every possible task, believing that higher token consumption correlates with higher productivity. But Narayan argues that this overlooks the quality and relevance of AI interactions. “If someone generates 10,000 tokens a day but produces no actionable output, what have we really gained?” he asked rhetorically.

GitLab’s rejection of tokenmaxxing is part of a broader cultural shift toward intentional AI adoption. The company has implemented a hub-and-spoke operating model to manage AI deployments. A central AI enterprise team handles governance, technical building and guardrails, while dedicated “AI transformation owners” embedded in individual divisions identify time-consuming, repeatable work that is ripe for automation.

Agentic AI in practice: transforming the employee experience

The hub-and-spoke model has already delivered tangible results in GitLab’s internal employee support network. The company built AI agents to assist its 120 internal support staff across IT, people operations and sales, helping them instantly pull context or deflect routine tickets entirely. Before the agents were deployed, support staff spent an average of 30 minutes per ticket searching through multiple systems for relevant information. Now, AI agents aggregate data from customer relationship management, data warehouses, chat channels and knowledge bases in seconds.

Pointing to a customer success manager (CSM) as an example, Narayan noted that the purpose of the role is to build deep client relationships, yet CSMs spend hours on administrative tasks such as building quarterly business review slides for clients, transcribing notes and hunting for context across disparate systems. By deploying AI agents to handle the grunt work, GitLab is looking to free up its workforce to focus on high-level strategy. “We want all of our team members to focus on what matters most: the core purpose of their role,” said Narayan. “We’re leveraging AI for tasks that can help them scale out in a more linear way, more than just a 10-15% increase in productivity.”

This vision of agentic AI goes beyond simple chatbots or copilots. GitLab’s agents are designed to take autonomous actions within predefined guardrails, such as updating records, sending notifications, or generating reports. The agents learn from human feedback and improve over time, creating a continuous cycle of automation and optimization. For instance, an IT support agent can automatically reset passwords, provision accounts, and escalate complex issues to human technicians, dramatically reducing resolution times.

Build vs. buy and the future of SaaS

As AI lowers the barrier to building internal tools, some industry observers have suggested that the days of off-the-shelf software-as-a-service (SaaS) applications are numbered. Narayan views this as vastly overstated, particularly from a governance and compliance perspective. “We may see more custom interfaces and the disaggregation of systems of interaction from systems of record,” he said. “But the underlying governance controls in core SaaS tools aren’t going anywhere.”

Narayan also pointed to the hidden costs of bespoke software development. “It’s easy to get to 90% of an application you develop in-house. That last 10% – the role-based access controls, auditability, immutable logging, which are things you need as a public company or as a company that deals with regulated customers – is incredibly complex to build.” This sentiment echoes the experience of many enterprises that have attempted to build custom AI applications. The initial development is often quick thanks to low-code platforms and LLM APIs, but production-grade requirements escalate costs exponentially.

GitLab itself has built several internal tools using AI, including a custom knowledge base query system and an automated code review assistant for internal development. However, the company continues to rely on commercial SaaS for core functions like human resources, finance, and customer relationship management. Narayan believes that the future of SaaS will be hybrid: platforms will offer more extensibility and AI agents that can connect with internal systems, but they will retain their governance and security foundations.

Governance and data classification

To ensure safety across custom and supplier tools, GitLab grounds its AI governance in a strict data classification standard. Public data flows through self-service platforms, while proprietary or customer data requires deeper security reviews before interacting with large language models. The company has implemented automated workflows that check data sensitivity before any AI processing occurs, reducing the risk of data leakage.

This governance framework is critical as GitLab expands its use of AI into more sensitive areas. For example, when the company considered using AI to analyze employee performance data, the central AI team conducted a thorough privacy impact assessment and implemented differential privacy techniques to prevent reidentification. The hub-and-spoke model ensures that each division’s AI transformation owner works closely with the central team to maintain consistency across governance standards.

Despite strong executive backing and budget, change management remains a challenge for Narayan. Bridging the gap between AI-forward employees and those who are slower to adapt requires a mix of departmental centres of excellence and internal AI hackathons. GitLab has organized several “AI days” where employees from all departments come together to prototype new AI applications, fostering a culture of experimentation and learning.

Yet, for a CIO, the greatest pressure is the ticking clock. “The thing that keeps me up at night is whether we’re moving fast enough,” said Narayan. “In the AI era, our decision-making needs to happen in days and weeks, not months and quarters. But I still worry about whether we are driving the right initiatives that are going to have the right long-term ROI for us.”

Broader industry context and implications

GitLab’s approach offers a blueprint for other organizations navigating the AI transformation. The rejection of tokenmaxxing aligns with a growing recognition that AI adoption must be measured by outcomes, not activities. Industry analysts have noted that many enterprises are falling into the “AI theater” trap, where they invest heavily in AI tools without rethinking underlying processes. GitLab’s first-principles approach addresses this directly.

The hub-and-spoke operating model also resonates with best practices in enterprise AI governance. By centralizing technical infrastructure and guardrails while decentralizing domain-specific innovation, organizations can balance speed with control. GitLab’s experience demonstrates that this model works even in a company that builds AI tools for customers, as it prevents siloed efforts that lead to fragmentation and duplication.

Looking ahead, Narayan sees the next frontier as integrating AI agents across the entire business application stack. GitLab is already experimenting with agents that can automatically hand off context from sales to customer success to support, creating a seamless flow of information. This end-to-end agentic approach could eventually lead to fully autonomous business processes, where humans only intervene when exceptions occur or strategic decisions are needed.

However, achieving this vision requires robust data infrastructure and a culture that embraces continuous learning. Narayan emphasizes that AI is not a one-time implementation but an ongoing journey. “We’re learning as we go, and we need to give ourselves permission to experiment and sometimes fail,” he said. “The key is to fail fast, learn, and iterate.”

Source: ComputerWeekly.com News