AI sprawl is becoming a pressing challenge for UAE enterprises, warns Sid Bhatia, Area VP & GM at Dataiku. Yes, without proper governance, agentic AI can multiply risks and inefficiencies across organizations.

As a renowned early adopter of emerging technologies, the United Arab Emirates (UAE) is routinely among the first to encounter their downsides. During the COVID pandemic, the country’s enterprises were among the first to embark on the Great Cloud Migration. In preserving business continuity and economic progress, organizations soon discovered the resultant lack of control over their own systems architectures. What was once proprietary became shared, fractured, and ill-defined. IT sprawl was born. Today, the UAE faces a similar sprawl with the latest emerging tech – agentic AI.

While AI has market momentum, we should be cleareyed about its maturity. AI agents need a strong hand on the tiller if organizations are going to chase opportunities into choppy waters. Robust governance must mandate care and collaboration in experimentation; it must demand 360-degree visibility through the elimination of operational silos; and it must insist upon meaningful metrics that guide future governance. Agent sprawl arises from the absence of these provisions, and like IT sprawl before it, it will be a sinkhole for resources and a multiplier for risk.

The problems

If we are to avoid inefficiency and risk, we must consider how IT sprawl came to dominate some enterprise stacks. Redundant SaaS apps, for example, led to rising costs and off-radar security vulnerabilities. Agent sprawl leads to similar cost inefficiencies by duplicating AI workflows, and it becomes a compliance risk by using the wrong data in the wrong way.

Governance is, more than anything, a way of having all departments row the boat in the same direction. The finance and HR teams may both be trying to eliminate paper. IT sprawl would have seen each pursue separate solutions to the same problem. Now that AI has become more accessible, it is more likely that team silos will lead to similar shadow AI agents.

Every hour a shadow agent sits idle is an hour of wasted GPU cycles. Every idle agent represents many wasted development hours. Meanwhile, infrastructure bills continue to pile up while a range of hidden costs escalate. Shadow agents can slip the notice of security audits, just like shadow apps have. When operating outside a sphere of governance, agents could access sensitive data outside mandated controls. This not only increases risk but multiplies it across agent instances.

The solutions

Just as governance works for IT sprawl, it can help with agent sprawl. First, governance teams must bring formerly isolated stakeholders together to define what business use-case each agent satisfies. They must be clear how the agent operates and decide who will take ownership. By establishing these things at the outset, the organization guarantees accountability, while ensuring that duplication of effort will be unlikely.

The stakeholder council should meet regularly to ensure all operational agents are still required. Those that are not can be retired, repurposed, or merged with others. Governance standards will dictate that only agents that have delivered measurable value will be scaled. To make these decisions possible, it is crucial that stakeholders have access to enterprise-level dashboards that allow them to establish the right metrics and act upon them easily.

Strong governance will introduce standardization in agent lifecycles, pipelines, lineage tracking, and audit logs. This will protect investments in agentic AI by ensuring agents can be reproduced on demand, but also that they remain compliant. Compliance is the bedrock of trust, and a lack of trust can spell failure for much more than just the agentic AI journey.

A life well lived

The lifecycle of an agent must be carefully, and centrally, managed. While prototyping itself should be effortless, all agents must be validated before they are pressed into service, after which they must operate under the appropriate access permissions. The metrics used to evaluate agentic AI must cover four areas: efficiency in terms of the usage of computing resources; accuracy and error rates; the presence of redundant, idle, or unowned agents; and productivity in terms of labor saved. These metrics serve as the core KPIs of agentic AI, allowing the best decisions to be made on the future of each agent. The best-performing ones can be shared across teams and deployed to orchestrate core workflows. Others can be retrained or retired.

These anti-sprawl processes will gradually increase AI capital and have agents contribute sustainable value to the business. Over time, its uncluttered AI infrastructure will grant the enterprise what it desires most, a competitive edge.

By Sid Bhatia, Area VP & General Manager – Middle East, Turkey & Africa at Dataiku

Related post

View all