Using the Agent Framework

Recently, we had the opportunity to present to Airtable’s AI User Group on a topic that has become increasingly important as organizations race to adopt AI: how do we move beyond AI as a feature and begin operationalizing AI as part of a workflow?

The reality is that most AI initiatives do not fail because of the model. They fail because of the system around the model.

Organizations often start by adding an AI field, generating some text, or connecting a chatbot and declaring success. Months later, adoption stalls, outcomes are inconsistent, and leaders begin asking where the promised value went.

The problem is not the AI.

The problem is that AI was added to a process instead of being designed into one.

From Prompts to Systems

One of the biggest shifts organizations need to make is moving from thinking about AI as a single prompt to thinking about AI as a system.

When humans perform work, they do not operate from a single instruction. They use context, history, policies, judgment, relationships, and organizational knowledge to make decisions.

For AI to create meaningful business value, it needs access to those same layers of information.

That is the foundation of the AGENT Framework.

The AGENT Framework

The AGENT Framework provides a practical approach for identifying, designing, and operationalizing AI-powered workflows.

AAssess

Before adding AI, identify the workflow.

Not every process deserves AI investment. Organizations should evaluate workflows on business impact, complexity, AI readiness, and organizational priority.

In the presentation, we evaluated four potential workflows. While several had high business impact, Support Triage emerged as the strongest candidate because it combined high business value, medium complexity, strong operational data, and existing SOPs and knowledge.

In other words, it had enough structure to benefit from AI while still requiring meaningful decision making.

GGround

AI is only as effective as the context it receives.

One of the most common complaints about AI is that it hallucinates or provides inaccurate responses. In many cases, the model is simply attempting to answer questions without enough information.

Grounding provides that missing context.

For the support triage example, grounding drew on three layers of context: the customer, their history, and the organization’s own guidance.

The goal is not simply to answer a ticket.

The goal is to answer the ticket with the same contextual awareness a human support professional would bring to the conversation.

EExecute

Execution is where AI becomes operational.

Instead of a single AI prompt, multiple specialized agents work together, each with one clear job.

This creates orchestration rather than isolation.

NNavigate

Human oversight is not a limitation.

It is a feature.

One of the most important lessons from enterprise AI implementations is that operational trust determines adoption.

Teams need systems that are observable, explainable, recoverable, and maintainable.

The AI should provide recommendations.

Humans should retain control over critical decisions.

In the demo workflow, AI generated recommendations, action plans, and draft responses. Human reviewers could approve, modify, or reject those recommendations before actions were taken.

TTune

AI systems are never finished.

Organizations should continuously evaluate how the workflow is performing, against measures the business already cares about.

The goal is not to deploy AI and walk away.

The goal is continuous improvement.

Just as organizations improve business processes over time, AI-enabled workflows require ongoing tuning and optimization.

The Live Demo

To bring the framework to life, I built a support triage solution in Airtable.

The workflow connected:

When a customer submitted a ticket, AI agents worked together to:

• Assess sentiment
• Recommend priority
• Gather customer context
• Evaluate existing feature requests
• Generate an action plan
• Draft a customer response
• Recommend creation of a new feature request when appropriate

The human reviewer remained in control while benefiting from a much richer set of information than would traditionally be available during triage.

Rather than simply answering tickets faster, the workflow helped the organization make better decisions.

Building the Demo with MCP

One of the most exciting parts of the project was using Airtable’s MCP integration with OpenAI.

Rather than starting from a blank base, I used conversation-driven design to help generate the initial schema, tables, and relationships.

The experience felt less like asking AI to “build a base” and more like paired programming.

The AI handled much of the repetitive setup work, while I focused on workflow design, data relationships, governance decisions, and user experience.

This significantly accelerated time-to-value while keeping humans firmly in control of architecture and business logic.

Final Thoughts

The future of AI in organizations is not about replacing humans.

It is about helping humans make better decisions by combining data, context, automation, and governance.

Organizations that treat AI as a feature will continue to struggle with adoption.

Organizations that treat AI as a workflow capability will create systems that scale.

That shift is where real value begins.