Agentic AI: Design for scale or prepare to fail

This page is part of A C-suite guide to capturing the potential value of AI.

As agentic AI moves from hype to hands-on enterprise reality, organisations are asking consultants to help them make the leap from isolated proof-of-concepts to enterprise-wide agentic AI systems that drive measurable business value. Scaling AI is deceivingly complex: it needs new processes, motivated people, secure data practices, and clear accountabilities all of which require a deep understanding of the organisational impact, in addition to the technology and source data.

What scale looks like and how it creates value

Scaling AI requires a fundamental shift in organisational maturity. It goes beyond deploying new models or agents or even expanding use cases. It is a transformation process that turns isolated pilots into solutions that are used at an enterprise-level in a reliable and secure way.

Scaling agentic AI drives value in two ways:

• ROI rises, while costs per unit of output falls. As more employees gain access to the solution and additional systems are automated, the initial gains seen in a pilot phase are multiplied and the investment in the solution development is spread further.
• Scaled systems involve the comprehensive integration of data, enabling organisations to tackle cross-domain processes. Agents are better able to manage multi-step, conditional workflows linked end-to-end to core business processes, driving competitive advantage and compounding value. 
  

Five common barriers to scaling agentic AI

Five interconnected barriers stand between organisations and scaled agentic AI deployment, as shown in figure 1.

^{Figure 1: five interconnected barriers preventing scaled agentic AI deployment}

While 74% of companies plan to deploy agentic AI, only 21% have a sufficiently mature governance model to do so: most businesses lack the ownership, guidelines and processes required to address the roadblocks shown in figure 1.

As AI moves from experimentation to production, the negative impact of under-preparation increases – cost controls lag, while the legal and reputational repercussions rise. Moreover, a ripple effect impacts all these pillars at once. For example, integrating a new AI agent on top of existing systems, an approach known as brownfield integration, requires a well-defined strategy that identifies the governance gaps and assesses data availability. In the absence of such a strategy, the AI guardrails may be inadequate, increasing risk. The organisation may also incur additional efforts and costs to operationalise the new solution due to unforeseen blockages.

Beyond the AI-related security risks (prompt injection, data poisoning, social engineering, model theft, etc.), which are compounded by a potential backlash against the technology, there is the deeper question of accountability: who is responsible when an AI agent causes harm, for example, by deleting a production database? As the technology evolves faster than regulations, organisations need to assess a complex compliance landscape and adapt their governance frameworks before scaling.

For each barrier, several elements must be considered: the suitability of the architecture, data and ops models, as well as the risks of vendor lock-in and integration incompatibility.

Although the barriers to scaling agentic AI are real and substantial, they are not insurmountable. By approaching these challenges early and strategically, organisations can overcome them. Deloitte helped a global pharmaceutical company, for example, to develop a dedicated AI agent platform to enhance go-to-market processes across business domains, based on a predefined scaling strategy.

Strategies to support agentic AI scaling

AI can be used to assist, augment and ultimately automate processes, as shown in figure 2. Organisations typically begin by using the technology to boost productivity (assist) and then for task optimisation (augment), before ultimately automating processes end-to-end. In each case, a distinct set of capabilities and scaling strategy needs to be in place at the start of the proof-of-concept phase.

^{Figure 2. Strategies to scale agentic AI}

Assist: improve individual productivity

Some AI solutions are designed to directly enhance employees’ skills: think of AI chat-based tools, such as Microsoft Copilot or Deloitte Headstart. They require employees’ trust and a safe, enterprise-wide environment (e.g., devoid of judgment and quotas for AI-usage) in which staff can use these tools.

As shown in figure 2, enabling enterprise-wide use of these solutions requires a rethink of frameworks, guardrails and ways of working to mitigate risks, ensure compliance with internal and external regulations and support employees through the transition.

Once these tools are scaled, employees become more productive, while maintaining control and trust. The organisation builds a foundation of AI literacy and confidence that can support scaling of more complex solutions.

Augment: optimise tasks with human oversight

AI can be integrated into well-defined tasks and roles, automating specific workflows with clear measurable outcomes, while maintaining human oversight. For example, Claude Code and GitHub Copilot can be used to automate portions of the coding and review processes to accelerate the delivery of software.

Scaling these solutions across an organisation requires significant adjustments of the enterprise architecture frameworks (including extensive monitoring) and data governance. Existing frameworks need to be adjusted as early as possible to prevent a scaling process hitting a dead-end, due to misaligned setup.

Once at scale, augmentation solutions make specific business processes more efficient. The organisation gains operational efficiency, while maintaining accountability.

Automate: agents handle multi-step, cross-domain workflows

AI agents can implement complex, multi-step workflows that span domains and require real-time decision-making. For example, Deloitte helped a global healthcare company to redesign multiple processes to be driven by agentic AI, supporting scaled automation of the procurement lifecycle.

To scale, agentic systems need to coordinate with each other to address complex business processes. That requires a redesign of enterprise architecture, frameworks and processes around data management, systems orchestration and guardrails, agent communication, and monitoring that should be identified and prioritised during the proof-of-concept phase. The goal is to define operational boundaries (e.g., which tools the agent can use, the transactions that are allowed), ensure observability and auditability of the agent (through telemetry, logs, alerts), while maximising ease of development and reusability (e.g., technical and governance building blocks).

Once implemented, agents can oversee routine decisions and workflows autonomously, escalating only when human judgment is required. The organisation gains speed, efficiency, and the ability to tackle complex, cross-domain problems that were previously impossible to automate.

Next steps: design governance, data architecture, and operating models

Scaling agentic AI is a strategic journey that requires intentional design, organisational alignment, and systemic thinking. Making clear choices on where to use and scale AI is a strategic conversation that goes beyond technology.

What matters is starting with intention: design for scale from day one, think about governance, data architecture, and operational models before you deploy, keep costs and sustainability in mind, and treat agentic AI as an enterprise transformation initiative.