How to spot AI agency app-stitching (and why it breaks in 90 days).

If you own a small business in New Bedford, Dartmouth, Fairhaven, or anywhere on the South Coast, you've probably been pitched an AI system in the last six months. The pitch went something like this:

"We'll build you a custom AI receptionist that captures every after-hours call, books appointments automatically, and integrates with your existing tools. We can have it live in two weeks. The investment is $4,500, plus a small monthly maintenance fee."

Sounds reasonable. The number is high but not insane. The agency is local-ish. They have a website with stock photos and testimonials. You ask a few questions, they have answers. You say yes.

Two weeks later, a system is "live." It demos beautifully when they show it to you. Three months later, it's broken in subtle ways. Booking patients at 3am on Christmas. Transferring calls to the wrong extension. Sending duplicate text messages. Occasionally just not answering. The agency's responses to your support tickets get slower. Eventually they stop responding entirely. You're now paying $200 to $400 a month for a system that's actively damaging your business.

This pattern has a name. It's called app-stitching, and it's currently the dominant business model for what calls itself the "AI agency" industry serving small businesses. This piece explains what it is, why it always fails, and how to spot it before you pay.

What app-stitching actually is.

Behind almost every "AI agency" pitch is the same technical architecture. The agency is not building software. They're connecting other people's no-code platforms with thin glue.

A typical "AI receptionist" sold by an app-stitching agency looks like this under the hood:

Each box is a third-party SaaS product the agency has stitched together with point-and-click integrations. Nothing is custom. Nothing is owned. The agency's actual deliverable is the configuration: a series of "if this, then that" rules they set up across these platforms.

This isn't necessarily bad for prototyping or testing an idea. It's catastrophically bad for systems that have to actually run your business. Here's why.

Why it always breaks.

App-stitched systems fail predictably for four reasons. Once you understand these, you'll spot the failure mode of every agency-built system you encounter.

1. Every seam is a failure point.

In the diagram above, there are six connection points between platforms. Each one of those is an HTTP call between two services that can fail, time out, return unexpected data, or change behavior when one of the platforms updates its API. The system as a whole is only as reliable as the worst of those six seams. In practice, the failure rate of stitched systems is high enough that something breaks every 2-4 weeks even in normal operation.

2. The platforms aren't designed to talk to each other.

Make.com, Zapier, OpenAI, Twilio, and Calendly are independent companies. They update their APIs on their own schedules. When one of them rolls out a change that affects how data is formatted, the agency's stitched workflow breaks. The agency then has to notice it broke, find which seam, and fix it. Most agencies don't monitor this. You find out when a customer complains.

This isn't theoretical. In August 2025, Make.com switched its entire pricing model from "operations" to "credits" and bumped overage packs by roughly 25%. In January 2025, Zapier changed which steps count toward your task limit and turned on automatic pay-per-task billing for overages by default. Both changes silently re-priced or repaired thousands of agency-built workflows overnight. If your agency hasn't talked to you about either, that's a tell.

3. The cost compounds.

Each platform charges its own fee. Make.com charges per credit (formerly per "operation"). Twilio charges per minute and per SMS. OpenAI charges per token. Zapier charges per task, and as of early 2025, automatically bills 1.25x the cost of a task above your plan limit, capped at three times your plan size. As your business grows, the costs grow with it, often non-linearly. A workflow that costs $80 a month at low volume can cost $400 to $600 a month at higher volume, and you have no way to optimize because you don't own the code.

4. There's no graceful failure.

Real software handles edge cases. When a customer says something the system doesn't expect, real software either responds intelligently or hands off cleanly to a human. App-stitched systems just... do whatever happens to happen next. Sometimes that's nothing. Sometimes it's something embarrassing.

The five red flags.

If you're being pitched by an AI agency right now, here are the specific things to listen for. Any one of these is concerning. Two or more, walk away.

What "actually coded" means (and why it matters for you).

The opposite of app-stitching is custom code. Real software written specifically for your business, deployed to infrastructure you own, talking to APIs only where it's necessary, bypassing the no-code middlemen entirely.

This sounds like a technical distinction. It's not. The difference shows up in your business in five practical ways:

• It actually works. Custom code handles edge cases because the developer wrote them in. Stitched workflows just execute whatever the diagram says.
• It costs less to operate. Custom code calls APIs directly, which means you pay only for the actual usage (typically 60-80% less than going through middlemen).
• It's actually yours. When the work is done, the code is in your repo. The accounts are in your name. The AI consultant can disappear and you're fine.
• It adapts. When you need to change how something works, a developer changes the code. With stitched workflows, you're at the mercy of whether the platform supports the change you want.
• It scales. Custom code doesn't get more expensive when your volume goes up. Subscription-based stitched workflows do, sometimes dramatically.

The 90-day failure cycle.

Here's the specific pattern that plays out at hundreds of small businesses across Massachusetts every quarter. If you've been through this you'll recognize it. If you haven't yet, watch for it.

Days 0-14: Deployment. Agency walks you through the demo. Everything works. You're impressed. You pay the second 50%.

Days 14-30: Honeymoon. The system handles routine cases well. You see real value. You tell other business owners about it.

Days 30-60: First cracks. A customer says something weird and the system mishandles it. The agency fixes it (slowly). Another seam breaks. You start wondering if this is normal.

Days 60-90: Support deteriorates. Tickets take days instead of hours. The agency starts using phrases like "platform limitation" and "outside of original scope." Charges for fixes appear.

Days 90+: The ghost. Support quietly stops. The agency is busy onboarding their next client. Your system continues running but breaks in increasingly weird ways. You either pay them more to "extend support" or you start looking for someone else to fix it.

Days 180+: Replacement. You either turn the system off or hire someone to rebuild it properly. The "$4,500 build plus $200/month support" cost has, by this point, become roughly $8,000 to $12,000 of total spend with nothing to show for it.

What to do if this just happened to you.

If you've recognized your own situation in this piece, the playbook from here is straightforward:

1. Don't keep paying the monthly fee while you figure out next steps. If the system is broken, paying to keep it broken isn't doing anything except funding the agency's next client acquisition.
2. Document everything you can. Screenshots of the workflow, exports of any data, contact info for any platforms in use. You'll need these whether you rebuild or just walk away.
3. Get a second opinion before committing to a rebuild. Sometimes 80% of what the agency built is salvageable. Sometimes it's not. A neutral developer can tell you in 15 minutes.
4. When you do rebuild, change the questions you're asking. Don't ask "can you build me an AI receptionist?" Ask "what would you actually build, what would it cost to run, and what happens if you stop existing?"

The honest alternative.

None of this is to say AI automation doesn't work for small businesses. It works extremely well. The issue is the gap between what AI agencies advertise and what they deliver, and the predictable failure pattern of the gap.

Real custom-built AI automation looks like this: a one-time flat fee, no required monthly retainer, code that lives in accounts you own, and a system that keeps working whether or not the developer is still around. The cost is typically less than the first six months of subscription-based equivalents, and the system gets cheaper to operate over time instead of more expensive.

That's not magic. That's just how software works when it's built correctly.

Justin