Beyond the Prompt: Solving the 'Execution Friction' That Kills AI Momentum

1. The "Trial Phase" Trap: Why AI Momentum Stalls

While interest in Generative AI has reached an institutional fever pitch, a sobering reality is settling into the enterprise: curiosity does not equal capability. According to McKinsey, Generative AI has the potential to contribute up to $4.4 trillion annually to the global economy. Yet, for most organisations, this value remains locked behind a "Trial Phase" trap.

The transition from general-purpose models to "business-ready" AI is failing because the hardest question in the building remains unanswered: "How do we actually use this in production?" Curiosity is being stifled by a lack of a clear pathway from a simple prompt to a reliable system. For the Enterprise Architect, the challenge is no longer about proving that AI can talk; it is about proving that AI can work within the rigid constraints of a stable, high-stakes operation.

2. The Governance Ghost: Addressing the Fear of Disruption

Psychological and structural barriers in the boardroom often manifest as "The Governance Ghost," a lingering fear that AI will disrupt "what already works." As noted in recent strategic directives, "Boardrooms question whether new systems should be trialled, how to manage governance and risk, and how to justify cost versus value."

To move past this, architects must quantify these fears. According to the Athenaworks impact scale (1 to 5), the top barriers to AI adoption are not technical curiosity, but structural risk:

• Data Privacy & Security (Impact Level: 5): The necessity of zero-trust models and GDPR/CCPA compliance.
• Compute Infrastructure (Impact Level: 4): The strain of scaling unoptimised models.
• Compliance & Regulations (Impact Level: 4): The requirement for auditable decision-making.

Without a robust governance framework, AI is a liability. Only by replacing unpredictability with precision can we transform AI from a risky trial into a high-performance corporate asset.

3. The Friction Points: Why General-Purpose Models Fail Enterprises

"Execution Friction" is rarely a failure of the prompt; it is a limitation of a generic architecture attempting to solve specialised problems. When shifting from pilots to production, three technical friction points inevitably stall momentum:

• Accuracy & Reliability: In high-stakes domains, hallucination rates range from 3% to 27%. Even frontier models like GPT-5, which has reduced errors significantly, maintain a 6.2% hallucination rate. In a "no-fail" environment, "convincing but wrong" is unacceptable.
• Scalability & Cost: Massive general-purpose models carry immense computational overhead. While API costs scale linearly with your success, the lack of cost-predictability creates a ceiling for enterprise-wide deployment.
• Domain-Specific Expertise: Generic models lack the depth for Legal (contract analysis) or Finance (audit workflows). This is often exacerbated by "terminology confusion," where a model misinterprets proprietary or industry-specific jargon, leading to operational errors.

4. Solving the Friction: The Enterprise AI Optimisation Toolkit

To bypass these bottlenecks, architects must transition from prompt engineering to systems engineering. The following toolkit provides the technical precision required for governance and risk management.

5. The Rise of Agentic AI and Domain-Specific SLMs

The next phase of momentum is the shift from chatbots to Agentic AI. While a standard Large Language Model (LLM) generates an answer, an AI agent generates an outcome. These systems autonomously plan, use external tools, and execute multi-step tasks.

However, the "Execution Friction" of bloated general-purpose models makes them poor candidates for these autonomous roles. The solution lies in Small Language Models (SLMs). Models such as NVIDIA Nemotron Nano 2 (9B parameters) and Phi-4-Mini are the "sharp tools" required for specialised digital factory workers. The Nemotron Nano 2, for instance, provides 6x higher throughput than comparable generalist models, allowing agents to react with the speed and efficiency required for real-time workflows.

6. Scaling at a Human Pace: Strategy vs. Wishes

A "Wish" is a plan that ignores the cost of people and existing friction; a "Strategy" accounts for them. The shift to SLMs is not just an architectural preference; it is the economic prerequisite for moving from the "per-request" cost models of Big AI to the fixed-cost stability of the enterprise digital factory.

The math of scaling reveals the necessity of this shift. Running a high-volume application through a GPT-4o API at $0.00275 per request totals $165,000 annually for 5 million requests. Conversely, a self-hosted, optimised SLM on a reserved instance costs approximately $727 per month.

The Break-even point is ~264K requests/month.

To scale successfully, enterprises must avoid "starting with platforms" and instead map a single, language-heavy workflow. Success is measured by outcomes (time saved and risk lowered), not the polish of a demo.

7. Conclusion: From Prompts to Governance

AI momentum is restored only when technical precision replaces curiosity. The shift from "tool selection" to "workflow design" is the hallmark of a mature AI strategy.

Strategic Checklist for Executives:

• Define Ownership: Assign accountability for AI performance, including clear review and escalation paths for handling edge cases.
• Map Language-Heavy Processes: Identify bottlenecks in reviews, approvals, and documentation where AI reduces friction.
• Prioritise Workflow Design over Tool Selection: Solve the process friction first; choose the model (SLM vs. LLM) second.
• Secure Data Pipelines: Ensure internal knowledge bases are connected through governed, secure pathways to maintain data sovereignty.

Momentum returns when you stop asking what the AI can do and start engineering what the AI must do.

How this applies in practice

Execution friction is most visible in businesses where AI trials have stalled because the tools were never connected to real workflows. A common example is a team using AI to draft documents or summarise reports, but still relying on manual processes for approvals, handoffs, and status tracking. The fix is not more prompting , it is workflow automation that embeds AI into the operational layer. When combined with business dashboards for visibility and proper data integration across existing tools, AI stops being a novelty and starts reducing manual workflows in measurable ways. Read more about the foundational challenge in our piece on the AI implementation gap, or get in touch to discuss your situation.