Why do most firms struggle to scale AI despite widespread experimentation? The key barriers and what sets high performers apart in agentic AI deployment.
According to McKinsey's 2025 State of AI report, 88% of organisations now use artificial intelligence regularly in at least one business function. The technology has moved from novelty to necessity in less than three years. But adoption and impact remain two different things.
Yet approximately two-thirds remain in experimentation or pilot phases. At enterprise level, only 39% report any earnings impact from AI, and most attribute less than 5% of EBIT to these initiatives.
The pattern holds for agentic AI, the category of systems capable of planning and executing multi-step workflows autonomously.
The gap between experimentation and scaling is not closing quickly. The question worth examining is what separates the minority who scale from the majority who remain stuck.
Before examining why scaling proves difficult, it helps to clarify what these systems do.
Agentic AI refers to systems built on foundation models that can act autonomously in workflows, not just respond to prompts. Unlike chatbots that answer questions, agents plan sequences of actions, execute steps, and adjust based on outcomes. An agent might read an email, categorise it, check company policy, draft a response, route for approval, and send it without human intervention at each step.
This capability distinguishes agents from earlier generations of automation. Traditional workflow automation follows fixed rules. Agentic systems interpret context, make decisions within parameters, and handle exceptions without requiring hardcoded logic for every scenario.
The technology enables automation in areas previously considered too variable for rules-based systems: customer service resolution, document processing, research synthesis, code review, procurement workflows. These tasks involve judgement, context interpretation, and multi-step reasoning that earlier automation could not reliably handle.
Early adopters concentrate agent deployment in IT service desks, knowledge management, and software engineering. These domains offer bounded problem spaces, rich data, and relatively clean toolchains that reduce implementation complexity.
Organisations experimenting with agents face a consistent set of barriers when attempting to scale beyond pilots.
.jpg?width=525&height=295&name=15%20Dec%20(2).jpg)
Agents do not simply automate existing processes. They require rethinking how work flows through the organisation. McKinsey's data shows high-performing organisations are three times more likely to fundamentally redesign workflows when deploying AI, rather than layering agents onto legacy processes. Among high performers, 55% report redesigning workflows compared to 20% of other organisations.
This redesign work proves difficult. Existing processes evolved over years to accommodate systems, policies, and human decision points that may no longer be necessary. Identifying which steps agents can handle, which require human oversight, and how to manage exceptions demands detailed process mapping and stakeholder alignment across functions.
Agentic systems act autonomously, which introduces questions about accountability, audit trails, and intervention protocols. Organisations need frameworks defining when agent outputs require validation, how to handle errors, and who bears responsibility for autonomous decisions. Building these frameworks requires coordination between legal, compliance, risk, and operational teams.
.jpg?width=494&height=278&name=15%20Dec%20(4).jpg)
Most mid-sized organisations operate hybrid environments mixing cloud-native applications with legacy on-premises systems. Agents need access to data and systems across this landscape. Integration complexity multiplies when agents must interact with systems lacking modern APIs, when data quality varies across sources, or when latency requirements conflict with security protocols.
Agent operations consume compute resources continuously rather than in scheduled batches. Token costs for large language models, while declining, remain meaningful at scale. Organisations experimenting with agents in controlled pilots often encounter unexpected cost structures when expanding deployment. Without clear understanding of usage patterns and resource consumption, budget planning becomes unreliable.
.jpg?width=538&height=303&name=15%20Dec%20(5).jpg)
McKinsey's research identifies a small subset of organisations, approximately 6%, that qualify as AI high performers. These organisations attribute more than 5% of EBIT to AI initiatives and report significant value capture. Their approaches differ from typical implementers in specific, measurable ways.
High performers set transformation objectives. While 80% of all organisations cite efficiency as an AI objective, high performers are 3.6 times more likely to target transformational change rather than incremental improvement. They pursue growth and innovation outcomes, not just cost reduction.
High performers redesign workflows systematically. As noted earlier, 55% of high performers fundamentally rework processes during AI deployment compared to 20% of others. They do not ask "where can we add AI?" but rather "how should this process work if designed around AI capabilities?"
High performers commit leadership attention. These organisations are three times more likely to report that senior leaders demonstrate active ownership of AI initiatives. Leaders model AI use themselves, participate in deployment decisions, and allocate significant budget. Over one-third commit more than 20% of digital budgets to AI, compared to smaller allocations from other organisations.
High performers establish clear governance. They define processes for when model outputs require human validation, track key performance indicators for AI solutions, and embed AI into business processes with appropriate oversight. All of these practices correlate positively with value capture.
High performers scale deliberately. Rather than attempting enterprise-wide deployment immediately, they expand from bounded domains to adjacent areas methodically. About three-quarters have scaled or are scaling AI, compared to one-third of others.
The pattern suggests that scaling agentic AI is less a technology problem than an organisational capability problem. The firms succeeding are those treating AI as infrastructure requiring systematic integration, not as a tool to bolt onto existing operations.
For mid-sized firms operating in the UAE, several factors influence how quickly agentic AI moves from experimentation to production.
1. Sector concentration matters.UAE organisations in financial services, healthcare, and technology sectors show higher agent adoption rates, mirroring global patterns. These industries have clearer regulatory frameworks for automation, established data governance practices, and operational processes already partially digitised.
2. Data readiness varies significantly.Organisations with clean master data, documented processes, and integrated systems find agent deployment more straightforward. Those operating fragmented technology stacks face longer implementation timelines as they address foundational data and integration issues before agents can function reliably.
3. Talent availability constrains some implementations.While UAE attracts technical talent, mid-market firms often compete with larger enterprises for AI specialists capable of designing agent workflows, establishing governance frameworks, and managing production deployments. This talent constraint affects how quickly organisations can move beyond vendor-led pilots to scaled internal capability.
4. Compliance requirements shape deployment patterns.Lastly, the UAE's Personal Data Protection Law and sector-specific regulations from bodies like the Central Bank establish boundaries for autonomous decision-making. Organisations must ensure agent actions remain auditable, explainable, and compliant with data handling requirements. This compliance overlay adds complexity but also provides clarity about acceptable use cases.
The McKinsey data suggests several principles for organisations considering agentic AI deployment.
Start with process clarity, not technology selection. Before evaluating agent platforms, map current workflows in detail. Identify decision points, exception handling, data dependencies, and hand-offs. Understand where variation occurs and why. Agents function best in processes with clear logic, even when that logic involves complex reasoning.
Define governance before scaling. Establish protocols for agent oversight during pilot phases. Determine audit requirements, error handling procedures, and human intervention triggers. Build these governance mechanisms into early implementations so they scale naturally rather than becoming obstacles later.
Treat workflow redesign as the primary work. Technology implementation is straightforward compared to organisational change. Allocate time and resources to rethinking how work should flow. Involve operational teams early. Their process knowledge proves more valuable than technical specifications when designing agent-enabled workflows.
Plan for continuous resource consumption. Unlike batch processing or scheduled automation, agents operate continuously. Model cost structures based on actual usage patterns from pilots. Build monitoring to track token consumption, API calls, and compute resources. Use this data to project costs at scale before committing to broad deployment.
Focus on bounded domains first. High performers scale methodically from contained use cases to broader applications. Choose initial domains with clear boundaries, good data, and measurable outcomes. Success in bounded contexts builds capability and confidence for expansion.
The divide between organisations experimenting with agentic AI and those scaling it will likely widen before it narrows.
High performers compound advantages quarter over quarter. As they deploy more agents, redesign more workflows, and build more capability, the gap between them and competitors expands. The window for catching up narrows as agent deployment becomes embedded in operational muscle memory rather than remaining a project-based initiative.
For mid-market firms, this creates urgency without requiring panic. The technology remains accessible. Vendor platforms continue maturing. Implementation patterns are becoming clearer as more organisations publish case studies and share lessons. The constraint is not access to technology but willingness to undertake the organisational work of workflow redesign and governance establishment.
The firms that recognise agentic AI as an infrastructure shift rather than a tool adoption will be the ones that scale.
Infrastructure requires systematic integration, governance frameworks, and operational redesign. Tools get added to existing processes. The difference matters.
McKinsey & Company, "The state of AI in 2025: Agents, innovation, and transformation", November 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai)
McKinsey & Company, "Seizing the agentic AI advantage", June 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/seizing-the-agentic-ai-advantage)
Why real-time data integration is crucial, replacing batch processing to meet modern business demands, enhance decision-making, and ensure compliance.
Improving Customer Service for SMEs: The AI Advantage
Part 1 of 5: Closing the Data Gap