What actually breaks in automations, and what it costs to fix.
Written from real jobs.
Your n8n workflow didn't break. It stopped being told.
Webhook triggers fail silently — how to tell three causes apart in ten minutes.
The n8n error that hides its own fix from Cloud users
When a Code node times out waiting for a task runner, n8n names the setting that would fix it - but only if you are self-hosted. On Cloud that sentence is removed before you see the message. Here is what the runner is, which timeout you hit, and how to stop depending on it.
Retrying a failed n8n run is not retrying the operation
The retry menu resumes from the node that failed, which is why it skips the check you wrote to stop a second payment going out. In these threads the duplicates mostly come from a fresh run, not from that button. Here is what each retry mechanism does.
WooCommerce 11.1 brings back the admin menus you disabled
The snippet shops use to hide the Marketing menu and the WooCommerce home screen still runs, and for those menus it no longer does anything. WooCommerce calls the technique unsupported. The code shows it half-works, and half-working is harder to notice than a clean break.
Google is moving your Search campaigns to AI Max this month
The migration runs through September, in place, on a day Google does not announce in advance. Search term matching arrives switched on, which changes what your budget buys. Google publishes a query that reads the migration date off your campaigns, and it has a flaw worth knowing.
Somebody tidied the spreadsheet, and the automation kept running
Zapier keys on the row number, Make asks you to type one in, and n8n matches on a value in a column. So sorting a sheet, deleting rows or hiding a column breaks a different thing on each platform, and the failure is usually silence rather than an error. Here is what each one tracks.
Make connects to Google, then the file picker returns 403
The connection is created, the authorisation succeeded, and the spreadsheet dropdown returns "Method doesn't allow unregistered callers". Five community threads carry that error over nine months and none has an answer from Make. The one-line test below tells you what is actually broken.
Nobody touched it and it broke: n8n Cloud upgrades itself overnight
A Cloud workspace on the fastest update cadence takes each new release the moment it ships, on average one a day. One of them broke every JavaScript Code node in late August, and the people it hit had not chosen to be on that release at all.
Zapier Functions stopped on 1 September: what to do with the code
Zapier retired Functions and left two different statements about who it affected. Moving code to Code by Zapier means re-architecting it, because secrets, packages, runtimes and the SDK all work differently. Here is how to tell whether a Zap is quietly dead.
Your AI assistant went quiet: OpenAI switched the Assistants API off
The shutdown landed on 26 August 2026, a year after it was announced. The people it hurt are not the ones who missed the announcement, but the ones who did not know their chatbot was built on it. Zapier, Make and n8n each handled it differently, and only one of them named the date.
Google Calendar made your recurring series private, retroactively
On 7 July Google changed how visibility works on recurring events. Existing series were rewritten to the strictest setting present anywhere in them, and per-instance API writes that loosen visibility are now accepted and ignored. Neither of the two pages announcing it tells the whole story.
Autosave failed, database is locked: the n8n errors that are not in your workflow
Four self-hosted failures from one month of the n8n forum, and none of them was caused by the workflow the message pointed at. A save rejected before it reached n8n, a database file locked with nothing writing to it, a duplicate that cannot be deleted, and a queue that keeps reconnecting.
Your Make scenario ran green and the data is wrong
A space you cannot see, a plus sign that changes meaning when you type it, a comma inside a customer name. Four documented ways a Make scenario finishes without a single red module and hands the next app a value you never meant to send.
Google Ads wants a passkey before it mints a new refresh token, and a new passkey is not ready today
Existing tokens keep working, so nothing breaks until the day you reconnect - and that is the day you find out the passkey the flow now demands may take days to become usable. What Google documents, which numbers disagree, and who can still avoid the flow entirely.
One inbox for comments across Instagram, Facebook and TikTok
The APIs exist and they are cheaper than the tools. What stops you is not money and not capability - it is that both platforms want a company, a verified domain and a review queue before they hand over the interesting half.
Your Zap tested green and fails on the first real record
Zapier documents the test as a representative sample, not a rehearsal. Three gaps between the test and the live run are written down in its own help centre, and each one produces a different failure on the first customer.
Overselling since August? WooCommerce 11.0 changed failed orders
Since 4 August, an order that moves to Failed puts its items back on the shelf. Where Failed really means a payment that fell through, that is a long-overdue fix. Where it means anything else, the change quietly resells stock that is not there - and the only trace is a note inside orders nobody opens.
Meta says your webhook is verified. So why do no messages arrive?
Verification answers one question - is your address reachable. Delivery answers a different one - what have you subscribed to. The green tick and the Test button prove the first, and business owners reasonably read them as proof of the second.
Make removes cycles on 1 October: what actually changes
Make says your scenarios keep running and that is true. Some of them will also produce two digests where they used to produce one, and the notification goes to whoever owns the account rather than to whoever watches the output.
Your automation ran twice and the customer was charged twice
The platform is usually not the one repeating itself. Duplicates come from the source system, which documents that it may send the same event more than once, and from the replay button that re-runs every step on purpose.
How to untangle a live automation without breaking it
The scenario grew past the point where anyone wants to touch it, and it is running the business. Here is the order of operations that keeps it running while you take it apart.
Your site is clean and the warning is still there
Removing the malware is the short part. Getting the site back into search runs on Google's review queue, and the two mistakes that turn days into a month are both made by the owner.
Your 300-page Elementor site got slow: what is worth fixing
Most of the speed you are missing is sitting switched off in Elementor's own settings. Here is what turns on safely, what breaks when you turn it on, and where a rebuild becomes the cheaper answer.
Your site is down, the host says it's fine: first 20 minutes
What the host actually checked before saying "everything looks fine on our end", and how to spend the next twenty minutes ruling out the rest.
n8n saved your change and production is running the old one
Autosave makes the editor feel finished. Production runs the published version instead, and one class of change publishes itself while the rest waits - which is why the same workflow can be both fixed and broken.
The run did not fail. It disappeared - and no node turned red
Automations that die under volume leave no error on any step, which sends owners hunting for a bug that does not exist. And when a setting really is being ignored, the difference between a misconfiguration and a defect comes down to one test.
Your bot told a customer something untrue: what to check
Two courts have now ruled that the bot is you. Here are the six places a working AI agent starts getting things wrong, and how to tell whether your fix helped.
Your Shopify metafield filter may be returning the wrong products right now
Filter by a metafield that is not set up for filtering and Shopify ignores the filter, returning everything as if it matched. From API version 2026-10 that becomes an error - which is the first time most integrations will find out.
n8n 3.0 drops npm installs. That is the smaller half of the problem
Moving your instance into Docker is a scheduled evening of work. The same release deletes nodes your existing workflows are built on, and that half cannot be checked after you upgrade - only before.
Make deletes the Google Chrome app on 31 August, and there is no replacement
This is not a migration notice. The app is being removed from every organisation, Make offers nothing to move to, and the scenarios that use it stop on the day - so the work is finding them and deciding what happens instead.
Your Make scenario timed out, and Make called it a warning
The email says warning, the schedule stays on, the error counter never moves and nothing is stored for retry. Plus the limit that actually applies is not the one printed on your plan.
It worked for months, you changed nothing, and today it fails
When an automation breaks without anyone touching it, the change usually happened at a third company whose release notes are addressed to developers. Even the schedule that governs it can contradict itself on the vendor's own page.
Your marketplace isn't in A2X or Link My Books. Now what?
The official answer is one sentence - export a CSV and enter it separately. Here is what that sentence leaves out, and how to book a payout so your margin survives it.
Reconnect this account - and the reconnect button is greyed out
Automation platforms tell you a connection needs reauthorizing, then show you a button you cannot press. The rules that actually killed the connection belong to the service on the other end, and they are published only in developer documentation.
How to watch automations across several client accounts
Every alert these platforms send is a by-product of a run. Silence is the absence of runs, so it is the one failure none of them reports - and the fix is to count expected runs, not errors.
One quote says $80, the next says $8,000: pricing an automation fix
Why quotes for the same broken automation differ by 10x, what each price band actually buys, and how to tell a real scope from a guess.
WordPress 7.1 has landed. Your Orders screen may not have survived
WordPress 7.1 rearranged the markup of post list tables on 19 August. WooCommerce shipped the fix nine days earlier, in 11.0.1. The breaking half can install itself overnight; the fixing half only installs if you turned that on, per plugin.
WordPress auto-updates: how many hours you are exposed
One 2026 core patch was under attack 90 minutes after release, while plugin fixes wait hours before your site is even told - including when you press Update Now yourself.
Your Zap ran all night: how to stop a Zapier loop eating your tasks
A Zapier infinite loop burns your quota in green, successful runs. How to spot one, shut it down in the right order, and keep it from coming back.