Evidence at a glance
The Reality of AI Agents: From Code Review to Customer Operations
In November 2026, Merge, an enterprise software integration platform, released a report on its internal adoption of OpenAI’s GPT-6 Astra, ChatGPT, and Codex models. Unlike standard implementations where AI remains restricted to Q&A interfaces, Merge built an architecture dubbed 'Merge for Workforce' that connects models directly into CRM systems, document stores, and engineering pipelines. This shift turns OpenAI’s models from mere benchmarking participants into the underlying engines for daily operations, including product development and client relationship management.
Merge’s engineering team now utilizes Codex for code reviews and configuration interface evaluations, while the marketing team leverages GPT-6 Astra for content curation and campaign planning. This is not merely an efficiency play; it is an attempt to use a unified integration fabric to solve data fragmentation across SaaS platforms, evolving AI agents from discrete task runners into an orchestration hub capable of authorized operations within business systems.
The Quality Gap: Proprietary Connectors vs. Public MCPs
Merge highlights a critical operational metric: tool call success rate. The company reports a 99.9% success rate with its proprietary connectors, compared to roughly 60% for standard public MCP (Model Context Protocol) servers. This discrepancy underscores the primary engineering barrier in current AI agent deployments: it is not the model’s reasoning capability that fails, but the fragility of the middleware. When an agent cannot reliably call a business function, the workflow collapses.
To bridge this gap, Merge implemented three core layers: Merge Unified for standardized customer data, Merge Agent Handler for secure access interfaces, and Merge Gateway for model routing and cost controls. For engineering leaders, the architectural lesson here is to treat 'reliable connectivity' as a hardened platform component rather than delegating API handling to individual agents. This architecture ensures that business systems like Salesforce or Gong remain stable endpoints, regardless of which model is currently driving the orchestration logic.
Automation Boundaries: Human-in-the-Loop as Operational Resilience
The case of Merge’s executive dinner planning illustrates the clear boundaries of current automation. While AI handles the heavy lifting—extracting leads from CRM, managing invite responses, and maintaining attendee trackers—the process still requires human judgment. The planning time for these dinners dropped from dozens of hours to just 30 minutes, but CEO Shensi Ding continues to review all AI-generated output, refining tone and final guest lists based on relationship context that the models cannot possess.
This 'human-in-the-loop' model highlights the objective limits of automation. AI is an exceptionally efficient execution layer, but it lacks the nuance of human relationship management. Organizations that fail to account for this 'context deficit' risk poor communication or strategic errors. For teams looking to integrate AI into their business workflows, success depends not just on automating tasks, but on clearly defining where AI stops and human final decision-making begins, which remains the ultimate determinant of a project's return on investment.
Engineering Judgment: Governing AI as a Production Factor
The broader significance of the Merge case lies in its approach to internal AI governance. Through 'Merge for Workforce,' employees are encouraged to choose the specific AI tools that best fit their workflows, fostering healthy competition and internal debate. This turns model selection into a managed production factor rather than a one-size-fits-all IT mandate. It forces companies to treat tool access, model routing, and permission management with the same rigor as traditional API gateway infrastructure.
For engineering leadership, the takeaway is clear: enterprise AI infrastructure is no longer just about subscribing to a model provider; it is about building an extensible, auditable platform that enforces granular access controls. AI only evolves from a demonstration toy into a business growth engine—as evidenced by Merge doubling its Q3 event throughput—when it is seamlessly programmable into existing toolchains. As a next step, teams should prioritize building robust monitoring for their connectors before scaling agentic workflows, ensuring system stability under the high frequency of AI-driven tool calls.