We've been here before, sort of
At Codit, we’ve built integration solutions on Azure for a long time. Logic Apps, Service Bus, API Management: the toolbox is mature, the patterns are well-established, and the platform keeps getting better.
But there was always a category of problem we couldn’t solve cleanly. Not because the connectivity wasn’t there, but because the data arriving through those connections was unstructured.
Azure AI Services has been able to extract text from scanned documents for years. So the problem was never the extraction, it was what came after. Raw OCR output is noisy, inconsistent, and completely different from one partner to the next. You’d end up either routing it to a human for manual handling, or building a brittle parser that broke the moment someone changed the layout of their template.
On the other side Logic Apps has had an Office 365 connector since practically the beginning. Trigger on incoming mail, grab the attachment, extract the subject: no problem. But if the actual unstructured content of the email body or the attached PDF was what mattered, you were stuck. The data was there. The connection was there. The understanding wasn’t.
That’s the gap AI fills. Not by replacing the integration layer, but by adding comprehension to it.
We could detect that unstructured data arrived. We couldn't understand what was in it. ”
AI as a component, not a replacement
This is the framing that matters most for integrations, and it’s one we want to be deliberate about.
In this context, think of AI as a preprocessor to your canonical integration model. It sits in front of your existing workflows, makes sense of whatever arrives in whatever shape, and hands off clean structured data to the pipeline that already knows what to do with it. The orchestration, the reliability guarantees, the connector ecosystem, the retry logic, the monitoring: none of that changes. What changes is that you can now tap into integration scenarios that were always just out of reach.
What was missing was the ability to turn what came through those connections into something a workflow could act on reliably. AI provides that, as a component inside your existing architecture, not as a complete replacement for it.
An email arrives, the AI preprocesses it into structured data, the Logic App takes over from there. The architecture looks familiar. What you can do with it is new.
What this unlocks in practice
The clearest example, and one we’ll go deep on in the next post, is order intake.
In B2B integration, purchase orders rarely arrive through clean EDI channels. Customers send emails. They attach PDFs. Sometimes the order is just written in the message body. Every customer has a different format. Building and maintaining parsers for each one is expensive and fragile.
With AI in the pipeline, the format stops mattering. A Logic App monitors an inbox, picks up incoming emails, passes the content and attachments to an AI model, and gets back structured JSON: order number, customer reference, line items, quantities, delivery address. That data flows into your ERP or order management system exactly as if the customer had used a proper EDI format. Most of them never will, and now that’s fine.
The same pattern applies across a wide range of scenarios: incoming contracts where you need to extract key dates and parties, support requests where you need to classify and route based on content, shipping documents with customs data that needs to be parsed and validated. In each case, AI doesn’t replace the integration workflow: it enables the workflow to handle inputs it previously couldn’t.
There’s a second dimension to this beyond document and email processing. Logic Apps now includes an Agent Loop framework that allows you to build conversational agents directly within your integration layer. Rather than just processing inbound data, the agent can reason across steps, decide which actions to take, and interact with backend systems to get things done. That opens up a different class of integration scenario entirely, and it’s something we’ll dig into properly later in this series.
What comes next
This is part one of a series on AI-augmented integration with Azure Integration Services.
In the next posts we’ll explore where this pattern actually pays off: the use cases that become newly viable when AI is in the pipeline. Order intake is one, but it’s not the only one, and understanding the broader landscape helps when you’re deciding where to start.
But we’ll also get practical. We’ll take the order intake scenario (emails, PDFs, plain-text orders, customers who’ve never heard of EDI) and build it end to end. What does the Logic App look like? Where exactly does the AI call happen? How do you structure the prompt? What do you do when the model isn’t confident in its output?
And then we go further into agents: what it actually means for an AI agent to be an active participant in a B2B integration flow, not just a processing step but something that can reason across multiple systems, handle exceptions, and collaborate with other agents. That’s where the genuinely new territory starts.
If you’ve been doing integration work for a while, some of what’s coming in this series will feel familiar. The platform is the same. The problems are the same. What changes is how much of it you can now automate, and that turns out to be quite a lot more than before.
Talk to the author
Contact Michel
He's our Data & App Integration Domain Lead