Easy to Deploy, Hard to Operate: What AI Workloads Change
Table of Contents
You can build an app with an AI coding tool on Saturday and have it live by Sunday night. Click deploy, paste in an API key, and you’re done. I think that’s great, and I’m not going to tell you to stop using those platforms.
But a few weeks or months later, I keep seeing the same thing happen. Easy-deploy platforms make deploying easy. They don’t make operating easy.
By operating I mean the boring work after launch: keeping it running, knowing when it stops, getting it back when it falls over, and changing it without breaking it. With AI workloads, that’s usually where things go sideways.
Why AI apps hit the limits sooner
Most default plans are sized for ordinary web traffic, where requests are quick and small. AI work is heavier in a few ways:
- You’re sending images, PDFs and long documents to a model. Once they’re encoded, they take up a lot more memory than the file size suggests.
- A single LLM call can be slow, and an agent that chains several together can run for minutes.
- Agent memory and chat history keep growing.
- Once you’ve built more than a handful of workflows, they start running at the same time.
Any one of these is fine on its own. Put a few together on a small default plan and you run out of headroom, usually on a day you aren’t looking.
What it looks like when it breaks
I researched recent posts on hosting and automation forums from people whose setups had gone down. I’ve left names out. A few stood out.
One person had an AI agent workflow that kept crashing on a hosted platform’s starter tier. When they ran the same workflow on their own server, it didn’t crash.
“I’m hitting reproducible OOM crashes … at very small payloads (~1-2 MB binary).”
Someone else saw the same thing on a 512 MB app host. It only happened when the LLM nodes ran.
The one that stuck with me was a builder whose automation server ran out of memory and restarted. The database had been sitting on the container’s temporary disk, with no volume and no backup.
“the n8n instance came back completely empty — no workflows, no credentials”
That was a production system.
Updates cause their share of problems too. Several people had AI workflows break after their hosting platform updated, even though they hadn’t changed anything:
“This workflow was running perfectly on the previous version without any changes on our end.”
Then there’s the question of who notices. In one thread, someone asked how people find out when a client’s automation quietly stops. Most answers came down to “the client told me” or “found it by accident.” One reply:
“It had been green for months.”
A default plan isn’t an operations plan
In most of these cases the platforms did their job. They ran the code they were given. The default settings just weren’t anyone’s decision about that particular app.
These are the questions I’d want answered for any AI app or automation running in production:
- How much memory does the heaviest workflow use at its peak?
- Where does the data live, and does it survive a restart?
- What happens to queued work if the process dies halfway through?
- Who gets an alert when something stops?
- Who checks that everything still works after the platform updates?
In nearly every thread I researched, the person learned these answers after the outage. Nobody had been given the job of running the thing.
What I mean by professional deployment
I’m not talking about huge servers or an enterprise toolchain. I mean making a handful of decisions on purpose:
- Size resources for the peak. Find out what your heaviest AI workflow actually uses, then leave room above it.
- Keep slow work out of the request. Long LLM calls and big jobs go to a queue or a background worker. For a busy n8n setup, that usually means queue mode on Postgres instead of the default single process on SQLite.
- Put your data somewhere that survives a restart, and back it up. Then try a restore at least once.
- Set up monitoring that reaches a person. Watch uptime, errors, memory and LLM usage. Also check that work is actually happening, because a workflow that stops running doesn’t throw an error.
- Control changes. Pin versions, test updates before production gets them, and know how you’d roll back.
- Give the system an owner: one person or team who watches it, fixes it and signs off on changes.
The sixth one holds the rest together. Alerts that go to an inbox nobody reads don’t help much.
When a managed platform is the right call
In fairness, a good number of the problems I researched weren’t the platform’s fault at all.
- One person thought their host had a 30-second limit on LLM calls. The timeout was in their own server’s start command.
- An outage blamed on a usage limit turned out to be a stray character in the source code.
- The 512 MB case above was fixed by upgrading on the same platform. In their words: “Fixed by upgrading to the 25$/month instance with 2gb RAM.”
Self-hosting doesn’t get you out of this either. I saw plenty of budget VPS setups and self-managed upgrades go wrong too.
If your workload is light and an outage wouldn’t cost you much, a managed platform on its defaults is probably fine. If it does matter, a managed platform can still be the right choice, as long as someone has sized it, is watching it, and owns it.
Ask the question before the outage
At the end of a lot of these threads, usually right after something broke, the same question came up in some form: “Any precautions I can take for the future?” It’s a good question, and it’s cheaper to ask it beforehand.
With new clients, I usually start with a Production Readiness Review. I go through how your app or automation is deployed: how it’s sized, where the data lives, how you’d find out it was down, and what the AI parts really need. You get a written report with the fixes in priority order. If you want help after that, it usually goes one of two ways. You keep the code and I set up the pipeline and monitoring for you, or you hand over a stable system and I run it.