Every other week, someone on Twitter declares that AI agents will replace Zapier, Make, and every other integration platform. OpenAI launches Operator. Anthropic shows computer use. Google demos Mariner. The takes start flying: "iPaaS is dead," "why would anyone use Zapier when AI can just do it," "automation platforms are the next Yahoo."
We've been building integrations for over 12 years. Left Hook was Zapier's first integration developer partner. We've watched every hype cycle come and go in this space. And we're here to tell you: those takes are mostly wrong. Not entirely wrong (there's a kernel of truth in there), but wrong enough that if you're making business decisions based on them, you're going to waste time and money.
Here's the more honest version of the story.
The Hype Cycle Is Doing What It Always Does
Let's do a quick walk through recent history. ChatGPT launches in late 2022, and within a week, people are declaring that Google Search is dead. GPT-4 launches, and suddenly developers are obsolete. AI agents appear, and now iPaaS platforms are toast. Coding assistants get good, and computer science degrees are worthless.
You see the pattern. A genuinely impressive technology appears, it gets demoed doing something cool, and then the internet extrapolates that demo to its most extreme possible conclusion. Every. Single. Time.
We saw this exact cycle with no-code tools ("developers are done"), with chatbots ("call centers are done"), with RPA ("manual data entry is done"). The technology was real in each case. Chatbots did change customer support. No-code did make some things faster. RPA did automate some data entry. But the "replaces everything" narrative? Never happened. Not once.
And here's the thing most people miss: those technologies didn't fail. They found their place. They became part of the stack. Chatbots handle tier-one support while humans handle the hard stuff. No-code tools let non-developers build internal apps while developers build the complex systems. RPA fills gaps between legacy systems that will never get proper APIs.
AI agents will follow the same pattern. They'll find their lane. But that lane isn't "replace the entire integration infrastructure that took a decade to build."
What AI Agents Can Actually Do Today
We want to be fair to the technology here, because it is genuinely impressive. If you haven't watched a demo of Claude's computer use or OpenAI's Operator, go look. It's remarkable. An AI agent can open a browser, navigate to a website, fill out forms, click buttons, read screens, and execute multi-step instructions that you give it in plain English.
That's cool. We mean it. It's a genuine technical achievement.
Now here's what the demos don't show you:
They're slow. A task that takes a human 30 seconds takes an agent 2-3 minutes. That's not a temporary limitation that'll get fixed next quarter. Screen-reading and UI interaction are fundamentally slower than API calls. When Zapier moves data from HubSpot to Asana, it's talking directly to both systems through their APIs. No browser involved. No pixels to interpret. Just data moving at machine speed. An AI agent doing the same thing is literally watching a screen load, finding the right text field, and typing into it like a very patient person with perfect memory.
They're expensive. Every action an agent takes burns tokens. A complex workflow that touches five systems might cost a few dollars per run in API calls. At scale (say, hundreds of runs a day), that adds up to a monthly bill that would make your accountant wince. Meanwhile, the equivalent Zapier workflow costs fractions of a cent per execution.
They're error-prone in ways that matter. An agent can click the wrong button. It can misread a field. It can lose context mid-workflow and start doing something unexpected. And here's the critical difference: when a Zapier workflow fails, it fails the same way every time and tells you exactly why. When an AI agent fails, it might fail differently each time, and the error might not be obvious until you notice that 50 records have the wrong data in them.
They're impressive demos, not production-ready automation. There's a massive gap between "look what it can do" and "I trust this to run 500 times a day with my client data." Most of the people declaring iPaaS dead have never had to debug a workflow at 2 AM because a critical business process broke. Production reliability is boring. It's also non-negotiable.
(We know this sounds dismissive. We don't mean it to be. The technology is early. It will get better. But "it will get better" is not the same as "it's ready now," and the people conflating those two things are doing you a disservice.)
What iPaaS Tools Actually Do Well
Let's talk about what Zapier, Make, n8n, and the rest of the iPaaS world have spent 10+ years building, because we think most people underestimate this.
Reliable, deterministic execution. When you build a workflow that says "when a deal closes in HubSpot, create a project in Asana," it does that. Every time. The same way. At 2 PM and at 2 AM. On the 5th record and the 5,000th record. That consistency is the whole point.
Error handling that actually works. What happens when the API is down? When a required field is empty? When the data is in a format the destination system doesn't expect? iPaaS tools have spent years building retry logic, error paths, fallback behaviors, and notification systems for exactly these scenarios. This infrastructure is invisible when things work (which is most of the time) and absolutely critical when they don't.
Logging and auditability. Every execution is logged. Every data transformation is recorded. Every error is captured with context. When your CFO asks "why did this record end up in the wrong state," you can pull up the execution log and trace exactly what happened, step by step. Try doing that with an AI agent that browsed through three websites and clicked a bunch of stuff.
API-level integration. This is the big one. iPaaS tools connect to systems through their APIs, which are stable, documented, and designed for machine-to-machine communication. AI agents connect through UIs, which change constantly (a redesigned button or a moved menu item can break everything), are designed for humans (not machines), and are orders of magnitude slower.
Managed authentication and credential storage. OAuth flows, API key management, token refresh, permission scoping. All handled. All secure. All maintained. This is plumbing that nobody thinks about until it breaks, and getting it right is genuinely hard.
This infrastructure took a decade to build. It solves real, unglamorous problems that AI agents haven't even started addressing. And honestly? We don't think AI agents will address most of them. Because the solution to "I need to reliably move data between two systems" isn't "have an AI watch a screen and click buttons." The solution is a well-maintained API integration. That's not going to change.
The Real Relationship: AI Inside the Platform
Here's what's actually happening (as opposed to what Twitter thinks is happening): AI isn't replacing iPaaS platforms. It's being absorbed into them.
Zapier already has AI steps you can add to workflows. You can drop an AI action into the middle of a Zap that classifies an email, extracts structured data from a PDF, drafts a response, or summarizes a document. Make has AI modules. n8n has AI nodes. Every major integration platform is incorporating AI as a capability within their existing orchestration layer.
And this makes total sense when you think about it. The platform handles the plumbing: trigger detection, data routing, error handling, logging, authentication. AI handles the fuzzy parts: understanding unstructured text, making judgment calls, working with data that doesn't fit neatly into fields.
Here's what this looks like in practice. One of our clients receives intake forms via email. The emails have inconsistent formatting (some are structured, some are just paragraphs of text, some have attachments). The old workflow required a human to read each email, extract the relevant data, and manually enter it into their CRM. The new workflow? Zapier catches the email, an AI step extracts the structured data regardless of format, and the rest of the Zap pushes it into HubSpot with proper field mapping. The AI handles the fuzzy part (reading messy emails). The platform handles the reliable part (routing data to the right place).
That's the pattern. AI augments the workflow. The platform orchestrates it. Neither replaces the other.
Where AI Will Genuinely Change Things
We don't want to be total curmudgeons here. AI is changing the automation space in some meaningful ways. We just think the changes are more subtle and practical than the hot takes suggest.
Natural language workflow building. This is already real. Zapier's co-pilot lets you describe a workflow in plain English and it builds the Zap for you. "When someone fills out my Typeform, add them to my HubSpot contacts and send them a welcome email." Done. That's a significant reduction in the learning curve, and it genuinely makes automation more accessible to people who found the traditional builder intimidating. We're fans.
Better data transformation. If you've ever wrestled with Zapier's Formatter steps to clean up a date format or split a full name into first and last, you know how tedious that can be. AI steps can handle these transformations conversationally. "Take this date and format it as MM/DD/YYYY." "Extract the company name from this email signature." That's real time savings.
Handling unstructured data. PDFs, emails, images, documents with inconsistent formatting. This is where AI genuinely shines and where traditional automation always struggled. If your workflow involves any step where a human currently reads something and makes a judgment about what it contains, AI can probably handle that step now.
Smarter error recovery. This one's early, but we think it's coming. Instead of a workflow just failing and sending you an alert, an AI step could attempt to diagnose and resolve common issues. "The field was empty, but based on the other data, it's probably X." Not here yet in a production-ready way, but We'd bet on it within a couple of years.
The Actual Threat to iPaaS (It's Not AI)
Here's a take you won't hear from many people in the automation space, because it's uncomfortable: the biggest threat to iPaaS platforms isn't AI agents. It's native automation built into every SaaS product.
HubSpot workflows are really good now. Salesforce Flow is powerful (if complex). Notion has automations. Slack has Workflow Builder. Monday has automations. Basically every SaaS product is building its own automation layer.
When HubSpot can trigger its own internal workflows (send an email, update a deal, create a task, notify a rep) without any third-party tool involved, the need for Zapier shrinks for those specific use cases. The more each platform can do internally, the fewer cross-platform workflows you need.
This is the convergence problem we wrote about in Are All Apps Becoming the Same?. As platforms expand their feature sets, they're also expanding their internal automation capabilities. That erodes the iPaaS value proposition more directly than any AI agent does.
The counterargument (and it's a good one) is that native automations only work within a single platform. The moment you need to connect HubSpot to QuickBooks to Asana to Slack, you're back to needing an orchestration layer. And most real businesses operate across dozens of tools. So iPaaS isn't going anywhere. But the edges are getting nibbled.
The Pragmatic Take
If you're reading this trying to figure out what to actually do, here's our advice. (If you want the broader framework, the SMB Automation Playbook covers the full decision-making process.)
Don't wait. "I'll start automating once AI figures everything out" is a losing strategy. The tools available today work. They're mature. They're reliable. And the skills you build using today's platforms (understanding data flows, API concepts, workflow logic, error handling) will transfer directly to whatever comes next. You're not wasting your time. You're building foundational knowledge.
Do experiment with AI-powered features. Add an AI step to a workflow and see what it can do. Use Zapier's co-pilot to build a Zap from a description. Try using AI to handle a data transformation that was previously annoying. These features are here now and they're genuinely useful.
Don't bet your business on AI agents for critical workflows. Not yet. Maybe in two years. Maybe in five. But right now, if a process is mission-critical (invoicing, client onboarding, compliance reporting), you want deterministic, reliable automation. Not an agent that might click the wrong button 1% of the time. Because 1% of 10,000 runs is 100 mistakes, and some of those mistakes might be very expensive.
Watch the space, but filter hard. The signal-to-noise ratio in AI and automation discourse is terrible. (We've got a whole post on who to actually listen to if you want help filtering.) Most of the people making the loudest claims have never maintained a production workflow. Listen to the builders, not the commentators.
The boring truth is that speed without judgment is expensive. AI agents are fast at doing things. They're not yet good at knowing which things to do, when to do them, and what to do when something goes wrong. iPaaS platforms are built on judgment (yours, codified into workflow logic) and reliability. That combination isn't going anywhere soon.
The future isn't AI replacing your integration platform. It's AI making your integration platform smarter. And honestly? That future is already here. It's just not as dramatic as the hot takes need it to be.
This post is part of The SMB Automation Playbook, a series on practical automation for small and mid-size businesses.