AI Changed the Way We Write Integrations. Forever

Here is a sentence. "When a student drops below half-time in Banner, update their Brightspace enrollment, notify financial aid, and pull their access to the payment portal." For most of my career, that sentence was not a sentence. It was a project. A statement of work, a data-mapping document, a consultant, a test environment, and a semester of your calendar. I spent fourteen years on the buyer's side of the higher-ed technology table, and I signed for that project more times than I want to admit.

Today that sentence is the program. So let me give you the direct answer the title owes you. AI changed the way we write integrations in three ways that will not reverse: who writes them, because every major integration platform now turns a plain-language description into a working workflow; what they cost to attempt, because the biggest software companies now say a quarter to a third of their new code is machine-written; and what an integration even is, because a shared protocol is quietly replacing the pile of bespoke connectors we all grew up maintaining. What AI did not change is the half of the job that was always the hard half. Deciding what should happen. Naming who owns it. Reading the log.

A week ago I told you your PowerShell scripts were debt, and I ended on a claim I owed you the math for: the rewrite is finally cheaper than the workaround. This is the math. Bring your skepticism, because the same technology that makes the rewrite cheap can also mint unowned automation faster than any tired identity engineer ever could.

The Whole Integration Industry Rebuilt Itself Around a Sentence

This did not start this year. SnapLogic shipped the first generative integration copilot while most of us were still writing chatbot acceptable-use policies. Boomi trained its builder on what it describes as more than two hundred million de-identified integration design patterns, then shipped a studio of named agents that design, document, and repair integrations. MuleSoft now turns your existing APIs into actions an agent can take from a natural-language request. Microsoft put plain-language agent flows into Copilot Studio and wired agents into the Logic Apps connector catalog enterprises already run. Every vendor in the category made the same bet: the person describing the integration no longer has to be the person who can code it.

Then there is the code itself. Sundar Pichai told investors more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers. Satya Nadella put Microsoft's number at twenty to thirty percent. The DORA research program found ninety percent of developers now use AI daily, a median of two hours a day. You can argue with any one vendor's marketing. You cannot argue with the direction. Writing glue code by hand, the most expensive habit in enterprise software, is ending the way most eras end. Not with a memo. With a quieter keyboard.

Notice what actually died, though. It was never the thinking. It was the typing. The mapping document did not go away; it became the prompt, and it still has to be right.

The Protocol Under the Prompt Is Moving Too

The second change is stranger and matters more for where this goes. For decades, integration meant hiring a protocol droid: something fluent in six million forms of communication, custom-built to translate one system to another. Every pair of systems needed its own, which is why your campus owns hundreds of translators nobody wants to inventory.

The AI industry just attacked that problem at the standards layer. The Model Context Protocol, MCP, gives any AI agent one common way to reach tools and data, and in barely two years it went from one vendor's open-source project to a Linux Foundation effort backed by companies that spend the rest of the week competing with each other. A sibling protocol, Agent2Agent, does the same for agents talking to agents, with over a hundred and fifty member organizations behind it. The MCP project claims tens of millions of SDK downloads a month and about ten thousand active servers; self-reported, yes. But the governance tells you what the industry believes. Nobody donates a protocol to a foundation unless they expect everyone to stand on it.

Higher ed is already in this story, mostly without noticing. Instructure says plainly that its Canvas agent is built on open standards like the Model Context Protocol, and that partners will integrate their own agents the same way. On GitHub you can find a campus IT shop's early MCP server for Ellucian Banner, wrapping Ethos and Banner's own APIs so an agent can speak student-information-system natively. Meanwhile I went looking for the sector's own coverage and found a silence: as of this writing, EDUCAUSE has published nothing substantive on MCP. The learning platform your students live in has adopted a protocol your governance program has never discussed.

The counterweights are real. Postman's latest State of the API research found two-thirds of developers know about MCP while one in ten uses it regularly; half of organizations run AI agents, while a quarter design their APIs with agents in mind. The protocol's own maintainers publish a security best-practices document that reads like a confession: confused deputies, stolen tokens, poisoned tools. Asana took its MCP server offline for nearly two weeks after a flaw let customers see other organizations' data. And underneath it all, an MCP server calls the same APIs your systems always exposed. The protocol does not replace your integration layer. It replaces the bespoke client code on top of it, and it hands the client's job to an agent. I have a prediction on record that cross-application agents will beat the AI features buried inside single applications. It is still on the clock, but the plumbing for it is being installed while we watch.

The Fastest Way Ever Invented to Mint Unowned Automation

Now the argument against my own headline, at full strength, because you should not trust a vendor who skips it.

The researchers who actually measure AI-written code keep finding the same shape. GitClear analyzed hundreds of millions of changed lines and found code duplication up eighty-one percent from its pre-AI baseline, while refactoring, the habit that keeps code maintainable, collapsed toward noise. Veracode ran AI-generated code from more than a hundred models through security tests and forty-five percent of it failed. The Cloud Security Alliance flagged research tracing a monthly tripling of CVEs attributed to AI-generated code, and called the count a floor. A randomized trial from METR found experienced developers were nineteen percent slower with AI assistance while believing they were twenty percent faster. Stack Overflow's survey of nearly fifty thousand developers found trust in AI output falling even as usage climbs, the top frustration being answers that are almost right. One enterprise survey put the hazard in a pair of numbers: ninety-two percent of technology leaders expressed confidence in AI-generated code while eighty-one percent reported an increase in production issues tied to it.

Read that pair again. Confidence up, failures up, at the same time, in the same buildings. That is the METR result wearing a suit: the feeling of speed and the fact of quality moving in opposite directions, and almost nobody assigned to notice.

This belongs in your integration strategy, not just your developer guidelines. An integration platform is precisely where generated code touches your systems of record with standing credentials. If a campus lets a hundred plain-language workflows into production with no owner, no review, and no log, it has not modernized anything. It has rebuilt the PowerShell problem I wrote about last week, at machine speed. The debt was never the language. It was automation nobody answered for. AI does not retire that debt. AI compounds it faster, unless the governance arrives with the generator.

On Campus, the Rewrite Finally Costs Less Than the Workaround

So why do I still stand behind last week's closing claim? Because on campus, both sides of the ledger moved at once.

The workaround got more expensive. The sector's own analysts count a typical institution spending about one point seven million dollars a year maintaining its systems, across dozens of product categories. And the SaaS migrations sweeping through Banner and Colleague country are removing the workaround's oxygen: when the database moves to the vendor's cloud, the direct-database scripts and nightly jobs your integrations secretly depend on stop being an option. Ellucian's own answer, an integration marketplace with ready-to-deploy connections, tells you the platform vendors see the same wall.

The rewrite got cheaper. Here is my stake, stated plainly so you can interrogate it. Our iPaaS Agentic Design Studio is exactly the pattern this column describes: you describe what you need in plain language, and an agent that knows Ellucian Ethos APIs natively writes the workflow, maps the data, and monitors it, with a self-healing layer that detects failing data flows and auto-remediates the common cases. It ships with a hundred pre-built workflows verified for SaaS, on a catalog of more than five hundred connectors. A recent customer announcement of ours quotes a university IT director calling the biggest win that his team is no longer tied to maintaining scripts. That is our announcement about our customer; after the column I wrote on interrogating vendor numbers, treat it accordingly. Ask us for the workflow logs. Credibility is the brand, and it matters most on the weeks the product is the subject.

And the honest limits, in print, as always. No independent study yet measures what AI actually saves on integration builds; the flattering numbers in this market are vendor-commissioned, ours included. An AI builder cannot decide what should happen when a student withdraws, cannot choose your grace periods, and cannot own the workflow it wrote. Someone on your campus still answers for every flow that touches a student record. No product, ours included, can name that person for you.

The discipline from last week survives the future almost unchanged. Inventory every integration, including the ones an agent wrote. Name an owner for each. Vault the credentials and keep the logs. And when you migrate a load-bearing flow, migrate it to something that reports its own outcomes, whether or not the person who described it could have coded it. The sentence became the program. That is the forever part, and I am not mourning it; I watched too many good people burn nights typing translators nobody thanked them for. But a sentence with no owner is just technical debt with better grammar. Next week the agents meet the auditors: access reviews and certification in a world where AI agents take action on your campus.

Take the Next Step

Primary: See the pattern in production: the iPaaS Agentic Design Studio , turns a plain-language description into an Ethos-native, monitored workflow, backed by the QuickLaunch connectors catalog. Governed by first-class identities for A.I. Agents that you create and managed like a human in your IT department. Create AI Agent Identities . Current capability detail lives in the release notes

Secondary: Follow the weekly column on the the QuickLaunch blog. Next week: preparing for audits in a world where AI agents are taking action on campus.

FAQ

How has AI changed the way integrations are written?

AI changed integration development in three ways: authorship, cost, and interface. Natural-language builders in platforms like Boomi, MuleSoft, Microsoft Copilot Studio, and QuickLaunch's iPaaS Agentic Design Studio generate working workflows from a plain-English description; major software companies report that a quarter to a third of their new code is now AI-generated; and open standards like the Model Context Protocol give AI agents a common way to reach tools and data instead of requiring bespoke connector code for every pair of systems.

What is the Model Context Protocol (MCP)?

The Model Context Protocol is an open standard, originally released by Anthropic in November 2024 and now governed under the Linux Foundation's Agentic AI Foundation, that defines one common way for AI agents and applications to connect to tools, data sources, and services. Instead of writing a custom integration for every agent-and-system pair, a system exposes one MCP server that any compliant agent can use. Education platforms have begun adopting it: Instructure states that its Canvas AI agent is built on open standards including MCP.

Is AI-generated code safe to put into production?

Not without human review and governance. Veracode's testing of more than 100 AI models found 45 percent of AI-generated code samples failed security tests, GitClear's repository research found code duplication up 81 percent with refactoring collapsing, and a CloudBees survey found 81 percent of enterprise technology leaders reporting increased production issues tied to AI-generated code. AI-written integrations and workflows should get the same treatment as any automation: a named owner, code review, vaulted credentials, logging, and monitoring that reports failures.

What was the Microsoft Entra Connect Sync September 30, 2026 deadline?

Microsoft required all Entra Connect Sync deployments to be on version 2.5.79.0 or higher by September 30, 2026; below that minimum, Microsoft's documentation states that all synchronization services fail until the engine is upgraded. Institutions running directory sync should verify their installed version rather than assume a working sync means compliance, and note that Microsoft's rolling retirement policy means each version retires 12 months after its successor ships, so the minimum version itself retires weeks after the deadline.

References

• SnapLogic Delivers the World's Only Generative Integration Solution — SnapLogic
• Introducing Boomi AI — Boomi
• Boomi Advances AI-Driven Automation with Major Announcements at Boomi World 2025 — Boomi
• MuleSoft unveils agent orchestration capabilities — Salesforce Newsroom
• Introducing agent flows: transforming automation with AI-first workflows — Microsoft Copilot Blog
• Announcing General Availability of Agent Loop in Azure Logic Apps — Microsoft Community Hub
• Google CEO says more than 25 percent of company's new code written by AI — The Hill
• Satya Nadella says as much as 30% of Microsoft code is written by AI — CNBC
• DORA 2025: State of AI-assisted Software Development — DORA (Google Cloud)
• Introducing the Model Context Protocol — Anthropic
• MCP Joins the Agentic AI Foundation — Model Context Protocol Blog
• A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms — The Linux Foundation
• Instructure Delivers on Safe, Simple AI Promise with IgniteAI and Major Ecosystem Updates — Instructure
• banner-mcp-server: MCP server for Ellucian Banner — GitHub (chatt-state)
• Postman State of the API 2026: key findings — Levo.ai
• Security Best Practices — Model Context Protocol Specification
• Asana warns MCP AI feature exposed customer data to other orgs — BleepingComputer
• The AI Code Quality Maintainability Gap — GitClear
• 2025 GenAI Code Security Report — Veracode
• Vibe Coding's Security Debt: The AI-Generated CVE Surge — Cloud Security Alliance
• Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR
• 2025 Developer Survey: AI — Stack Overflow
• 81% of enterprise technology leaders report production failures from AI-generated code — CloudBees
• Average Software Budget in Higher Education — LISTedTECH
• Banner SaaS Modernization: Rebuilding Legacy Student Logic — ABCloudz
• Ellucian Marketplace Accelerates Campus Innovation with Ready-to-Deploy SaaS Integrations — PR Newswire (Ellucian)
• Amberton University Automates Google Workspace Provisioning with QuickLaunch — QuickLaunch
• Microsoft Entra Connect Sync: Version release history — Microsoft Learn
• QuickLaunch Release Notes — QuickLaunch

*Raymond Todd Blackwood is the President of QuickLaunch and writes about identity, agentic AI, and the messy reality of higher-ed IT. #ItsExistential*