Integrate 2026 Day 2 Highlights

Integration in an AI Era

Dan Toomey, hailing from the Gold Coast in Australia and a seasoned speaker at Integrate for a decade, presented a well-structured talk divided into three acts. He explored the evolution of APIs transitioning from human users to AI agents and discussed why integration is now the backbone of all digital operations. The three acts focused on the shift in API consumers, the role of integration as the operating layer, and practical design suggestions for creating AI-friendly APIs and integrations.

Act One: Transitioning from Humans to AI Agents

Dan kicked off with a thought-provoking scenario: most API documentation is designed for human consumption via web browsers, yet it’s uncertain whether humans will be the primary users in two years. For the last two decades, APIs were crafted and accessed by developers, making human-readable documentation essential. However, with the rise of AI, APIs are now being accessed dynamically in real-time, leading to the need for machine-discoverable capabilities. AI agents bypass traditional documentation, relying instead on specs, descriptions, and historical success rates. The takeaway? If your API is not easily discoverable, understandable, or dependable, it risks being ignored altogether, leaving much of your functionality wasted.

What AI Agents Seek

Dan outlined key characteristics that agents require, providing examples for both effective and ineffective implementations:

• Concise semantic clarity – Effective descriptions state what the API returns along with data structures, predictable behaviours, and error codes. Ineffective ones mimic marketing language, making validation overly complex.
• Defined request/response formats – Clear, consistent, and strongly-typed data formats are vital, eliminating the guesswork involved in interpreting examples.
• Accessible authentication – Utilising machine-to-machine OAuth with client credentials, providing clear token lifespans, and deterministic error messages are crucial. Complex methods requiring human interaction should be avoided.
• Consistent error management – A structured method for error reporting across endpoints, properly using HTTP status codes, and clearly signalling whether an agent should attempt a retry.
• Measured historical success – Agents disregard APIs that show frequent failures.

Act Two: Integration as the Operational Layer

AI systems lack direct interaction capabilities; they rely entirely on APIs to acquire data and initiate actions. Hence, integration becomes the operational layer, translating AI intentions into real-world results. AI’s role has expanded beyond merely assisting in application development to autonomously executing workflows that involve multiple API calls. The implications are significant: if your integrations are fragile, your AI will also be vulnerable, as it heavily depends on those integration platforms.

The Cascade Example

Dan presented a real-world example involving a retail “smart replenishment” agent that relies on three integrations: an internal inventory API, a machine learning sales forecast API, and an external supplier ordering API. When a developer undocumentedly renamed a crucial field, the agent misinterpreted it, leading to a costly order of 8,000 units due to a simple integration error. This illustrates how a singular oversight can spiral into substantial financial repercussions.

Key Considerations: Access, Protocols, and Context

Dan identified three vital considerations when dealing with AI agents. First, the value of an agent correlates with its access – organisations that place their APIs first have a notable advantage over those required to adapt legacy systems. Second, he touched on the protocols landscape, noting that while the MCP helps discover tools, it isn’t without its flaws, including decreased reliability in larger catalogues and security vulnerabilities. Third, agents operate within limited context windows; each description consumes budget. Take care to avoid “context saturation,” where lengthy descriptions impede reasoning capability.

Act Three: Designing APIs for AI Agents

Dan concluded by outlining practical design principles for creating agent-friendly APIs:

• Clearly defined, machine-readable descriptions – Specify what each endpoint accomplishes, when to utilize it, and potential side effects.
• Guiding error management for agents – Implement well-defined HTTP status codes and structured error formats with necessary retry flags.
• Principle of least privilege – Each callable tool represents a security risk. Design with assumptions that the caller is autonomous.
• Mechanical rate control – Enforce stringent rate limits to prevent abuse; agents must not be solely guided by goodwill.
• Compose rather than isolate – Structure APIs into small, predictable units that facilitate seamless integration.

Implementation on Azure

Dan illustrated how these principles can be applied using Azure services: leveraging API Management for robust governance and exposure of agent-ready APIs; using Logic Apps for workflow automation due to its vast array of connectors; and employing Azure OpenAI for generating client-side code. The core message was that while APIs have traditionally served as the plumbing of applications, they now form the very essence of the product in an AI-centric world. Organisations need to cater to both human and agent consumers of APIs, as overlooking one of these audiences could lead to a rapid decline in relevance.

Audience Interaction

During the Q&A, it was noted that two-thirds of attendees already document their APIs, with more planning to start soon. Dan acknowledged the complexity of debugging non-deterministic systems and the importance of strong governance throughout the integration process. He reflected on the need for teams to secure their systems properly, especially considering how autonomous agents may not exercise sound judgment on their own.

Why Azure Logic Apps Excel for Integration Tasks

Bill Chesnut, an API evangelist at Six Pivot with a wealth of experience in both BizTalk and Azure, underscored the superiority of Logic Apps for integration tasks, primarily highlighting its support capabilities. Facing a cold, he passionately argued that the visual design approach he advocates isn’t merely a preference; it’s essential for maintainable production integrations.

Understanding Integration Workloads

Bill defined an integration workload as one wherein control lies primarily on one side. It involves processing messages that arrive from third parties, such as clients, suppliers, or SaaS applications. Since organisations do not control the data sent by others, unexpected changes can occur. Additionally, reliance on third-party systems can introduce failures due to maintenance periods or timeouts. His main concern highlighted that testing data is often unreliable because what testers submit rarely matches the quality of actual production data.

Supportability: The Key Argument

Bill emphasised that many developers overlook the ease of support when handing over integrations. This oversight has become more apparent as teams embrace AI, leading to increasingly complex solutions that are often difficult to maintain. He shared an illustrative incident where a support issue escalated because of an unanticipated edge case in live data, highlighting the necessity for high-quality support solutions.

Importance of Restartability

Restarting failed processes is another theme of supportability. Against the backdrop of EDI, Bill illustrated the importance of being able to reprocess without resending confirmation messages that could confuse partners. Logic Apps allow for seamless action restarts, whereas function apps often require starting from scratch, which can lead to complications.

Choosing the Right Options: Logic Apps vs. Function Apps

Bill positioned Logic Apps Standard against code-oriented options like function apps, focusing on message-based integrations. He praised Logic Apps for its extensive range of connectors, which facilitate rapid development. However, the most significant asset is its visual run history, as it reveals exactly what happened throughout the integration process—this transparency is invaluable for debugging.

Insights on Testing

Bill offered a provocative opinion on testing: while unit tests are now predominantly AI-generated, they are often ineffective for integration workloads because they focus on static data rather than potential live data changes. Integration tests require a complete view of systems involved, which can be difficult. Bill acknowledged AI’s role in enhancing data generation but reiterated the ongoing importance of data quality.

Demo and Conclusion

Bill concluded by demonstrating the differences between Logic Apps and function apps. After resolving a brief visual hiccup, he illustrated how Logic Apps process shift data and emphasized its easy-to-follow run history compared to the less informative outputs of function apps. He suggested that implementing clear error-handling measures significantly aids support teams, allowing them to track issues efficiently and resolve them without excessive developer intervention. His final takeaway stressed that developers often overlook the support aspect of integrations; Logic Apps enhance support capabilities, thus allowing developers to focus on innovation.

Audience Queries

In response to questions, Bill noted that the visual nature of Logic Apps aids in capturing complex workflows clearly, making them accessible. He expressed enthusiasm for the new .NET SDK and discussed the potential for AI to streamline task management but cautioned against full reliance on AI without adequate checks in place. He advised splitting complex workflows into manageable components to maintain clarity and efficiency.

Lessons Learned in Agentic Integration: Modernising MuleSoft and Accelerating New Developments

Dave Phelps, a cloud solution architect at Microsoft, alongside Gil Perry from Quokka, shared practical insights into using agentic AI to facilitate enterprise integration projects, especially those involving transitions from MuleSoft. They stressed the importance of balancing the speed of integration development with thorough validation to avoid downstream bottlenecks.

Reasons Behind Integration Initiatives

Gil painted a vivid picture, explaining that companies do not seek modernization out of boredom but rather due to genuine pressures such as rising operation costs and the need to consolidate platforms. Firms often face the dual pressure of needing to transition whilst keeping operations ongoing, complicating their modernization efforts.

Reality vs. Theory in Integrations

He highlighted the discrepancies between the ideal textbook APIs versus the messy real-world integrations characterised by gaps in documentation and knowledge. The focus should be on improving understanding, rather than just rushing into migration.

From Prompts to a Maturity Framework

While AI has the potential to increase workloads drastically, it also poses challenges in ensuring that quality validation is maintained alongside increased development speed. Through refining their processes, the team found that the true challenge lies not in generating code but in comprehending and fulfilling business needs effectively.

The Paradigm Shift: Focusing on Understanding

Shifted focus on requirements over raw code generation led to the creation of vital documents and specifications that facilitated better project understanding. This laid a strong foundation for defining integration requirements devoid of platform-specific hindrances, easing the transition between various integration tools.

Specialised Agents Approach

Introducing specialised agents that handle specific tasks brought about greater consistency in integration processes, with discovery agents scanning the ecosystem and detailing connection dependencies. The improved understanding of integration architecture allowed teams to organise and prioritise migrations effectively.

Agent Usage in Delivery Processes

A tangible evolution occurred as agents began influencing the delivery workflow, moving towards automated systems that could operate with less human supervision without sacrificing governance. This resulted in more efficient processes that aligned well with the continuous improvements within the teams.

Results and Demonstrations

The culmination of these approaches led not only to faster understanding and improved documentation but also a fundamental shift towards a requirements-driven delivery platform. This enabled swift transitions from legacy integration to new developments, showcasing the extensive benefits of planning integration requirements thoroughly.

Responses to Queries

Audience interaction revealed a substantial engagement with AI in generating requirements, indicating a positive uptake of innovative tools in integration development. Suggestions on refining integrations and managing complexities focused on capturing essential operational aspects, emphasising the importance of collaborative reviews early in the process.

Common Challenges in Migrating from BizTalk to Azure Integration Services

At Integrate 2026, Sandro Pereira, a veteran in enterprise integration, candidly discussed the frequent pitfalls faced during transitions from BizTalk to Azure Integration Services.

Sandro’s initial assertion was that simple lift-and-shift migrations often fail; although business requirements can carry over, completely different technology stacks necessitate thoughtful redesign. Neglecting this results in wasted resources and unmet expectations regarding the performance of new systems.

She cautioned against excessive reliance on AI migration tools, mentioning that these tools can only reflect the quality of existing BizTalk solutions, which may not always be pristine. Furthermore, awareness of the new cost dynamics in cloud environments is essential for ensuring sustainable budgets.

What Went Awry? Logging and Exception Management in AIS

During his presentation, Daniel Probert from Afitness illustrated the crucial need for robust exception handling and logging. He recounted the story of a support incident involving a massive batch failure, highlighting how poor logging can unnecessarily prolong resolution times.

Daniel argued that every integration is bound to break, raising the critical question of whether teams are prepared to troubleshoot effectively. Good logging should clarify which processes failed, the reasons behind the failures, their impact, and direct points of intervention.

He advocated for investing in quality error management as a vital architectural standard, emphasising the long-term cost savings and efficiency gains it ensures.

Who Holds Architectural Responsibility in an AI-Driven Coding World?

At the conference, Mattias Lögdberg of DevOps Solutions pondered over the pressing issues regarding architectural accountability in the wake of rapid AI-generated code production. He illustrated this with the story of the Winchester Mystery House, where an architect’s absence led to chaotic construction decisions.

His central message was that the challenge lies in preserving an understanding of architecture amidst accelerating implementation speeds. He introduced a continuous verification governance framework using their Helium tool to maintain compliance and visibility.

Unlock Significant Savings through Migrating from MuleSoft to Azure

Andrew Rivers from 345 Technology stressed the financial advantages of transferring from MuleSoft to Azure’s Integration Services. He illustrated how a MuleSoft infrastructure could lead to exorbitant licensing costs while demonstrating how Azure offers a more cost-effective, consumption-based pricing model, enabling substantial savings.

AI, particularly through tools like GitHub Copilot, has significantly enhanced both migration speed and accuracy. Teams can automate redundant processes that would otherwise take months, streamlining the transition from legacy systems.

The Final Developer in Integration?

Rob Hofman from Delaware discussed AI’s transformative potential throughout the integration development lifecycle, emphasising that AI enhances, rather than replaces, the role of developers.

Utilising Delaware’s Smart Link framework, Rob explained how AI is systematically integrated across phases, significantly improving collaboration and output quality. His key insight was that a structured approach to AI integration yields better, more consistent results than uncoordinated experimentation.

Engineering Resiliency in Azure Logic Apps

Andrew Wilson from Black Marble provided an in-depth look at the need for resilience in Azure Logic Apps. He argued that while building workflows is simple, ensuring they are production-ready requires intentional design to account for failures.

He shared three fundamental principles: anticipate failure, constrain the impact of failures, and design recovery paths. A successful workflow diagram represents the ideal path but should also reflect how the system reacts to common disruptions.

Developing Azure Integration Services with GitHub Copilot

Cameron McKay of MMP Digital showcased the potential of integrating GitHub Copilot across the entire Azure Integration Services lifecycle, stressing that understanding the failure modes is critical for achieving reliable outcomes.

The session addressed the impending shift to usage-based billing for GitHub Copilot, urging teams to set appropriate budget controls while predicting a dynamic future of AI-assisted development structures.

Turbo360 Updates for Azure Integration Services Customers

Turbo360’s CTO Michael Stephenson shared recent enhancements aimed at benefiting Azure Integration Services users. A significant focus was on the BizTalk-to-Azure accelerator program, offering seamless migration support.

A key update highlighted in his session was the introduction of an enhanced cost management module, allowing teams to gain deeper insights into integration costs and supporting strategic financial decisions around integration design.

Controlling AI Token Costs with Azure API Management

During his session, Microsoft Azure expert Alex Vieira tackled the complexities of managing AI token expenses as enterprises scale AI adoption. He outlined the various challenges faced and demonstrated how Azure API Management can mitigate these issues by mediating API traffic effectively.

He underscored the significance of integrating AI cost management practices into overall cloud financial operations to ensure controlled and predictable expenses.

Secure Networking in Azure Messaging

Christina Compy from Microsoft outlined the necessary steps for securing services like Service Bus and Event Hubs in Azure. She articulated the architectural distinction between inbound and outbound services, emphasising the significance of tailored network security measures.

Announcing several important developments, including IPv6 support and enhanced service-level agreements, Christina underscored the importance of careful planning in network security within Azure environments.

Event-Driven Applications with Azure Functions and MCP

Thiago Almeida, Principal PM for Azure Functions, revealed developments in connector namespace that will allow Azure Functions and other services to access the expansive library of managed connectors without needing to rely solely on Logic Apps.

The integration streamlines how developers interact with SaaS systems and prepares Azure Functions for a future of enhanced event-driven workflows through AI.

API Center: Your Comprehensive Catalog for AI and APIs

Sreekanth Thirthala Venkata Manne introduced the Azure API Center, which aims to centralise AI and API assets, addressing potential chaos as AI proliferates in enterprise settings.

With an emphasis on quality and governance, the API Center offers automated syncing of registration from repositories, ensuring that all assets remain current and compliant while enhancing discoverability for development teams.

New Horizons in Azure Event Grid: Scalable and Ready for IoT

Seth Shanmugam showcased Azure Event Grid’s evolution as a robust messaging backbone, integrating features essential for IoT readiness and high reliability in event-driven applications.

Highlighting new capabilities, he affirmed the platform’s preparedness for diverse application needs, from secure filtering to compliance with industry standards.

Creating AI-Native Platforms with Azure Container Apps Sandboxes

The Azure Container Apps team unveiled the new sandboxes designed for AI-native workloads, addressing several common runtime issues, thereby setting a stage for seamless platform development.

These innovations streamline resource management and ensure rapid deployment, supporting enterprises in accelerating AI project rollouts.

Introducing Logic Apps Automation for Streamlined Business Processes

Divya Swarnkar, Product Manager for Azure Logic Apps, concluded the conference with the announcement of Logic Apps Automation—a new skill tier aimed at democratizing access to process automation capabilities.

This innovation allows less technically adept staff to build and deploy workflows efficiently, retaining the governance and controls essential for enterprise-grade applications.