Inherited Automations: How to Audit and Fix Zapier, Make & n8n

A customer writes in: they never got the confirmation email. You check the CRM. The deal is there. The inbox is empty. The Zap last ran three weeks ago, and nobody on the team is in the Zapier account.

Or the bill jumps. Last month Zapier was $50. This month OpenAI is $400 and Make used a month of operations in four days. The builder is on another Slack now.

Or you open Make and a scenario that invoices clients has been off for a month. No email. No Slack ping. The schedule just stopped.

Those three symptoms are the same problem: you inherited automations you do not own. They are failing, drifting, or sitting on someone else’s login. You need a takeover, not a new build.

Here is the sequence. Freeze what you have so history does not expire. Take back logins and billing. Inventory every live flow. Rank the ones that can hurt a customer or a card. Find why each dangerous one is failing. Fix, rebuild, or retire it. Turn on alerts that catch silence, not just red errors. Write a one-page runbook so the next person is not you at 11pm.

You can finish that in a week or two. Do not start by adding another Zap.

How to tell your automations are failing without anyone noticing

Most inherited stacks do not explode. They go quiet. The toggle is still green. The customer is the monitoring system. If you only look for error emails, you will miss most of it.

Three kinds of danger: wrong sends, silent failures, runaway cost

Wrong sends still “succeed”. The email went to the wrong list, the invoice used last month’s template, or the agent posted an internal note into a customer thread. Success in the log is not success in the business.

Silent failures stop without a scene. A Zap is off. A Make scenario is deactivated. A webhook died when someone rotated a secret. Nothing paged you.

Runaway cost looks like growth. A loop retries a failing API. An AI step rewrites the same document on every CRM update. The bill is the alert, and it arrives after the money is gone.

Score every live flow against those three. A flow that can do all three at once (customer email + payment + an AI call) is this week’s work.

Why “no error email” does not mean working

Zapier does not email you on the first failure. If autoreplay is on, it retries at 5 minutes, 30 minutes, 1 hour, 3 hours, and 6 hours. You get the error email on the last attempt, about 10 hours and 35 minutes after the original failure (What is replay). A Zap can be broken all day and your inbox still looks fine.

Make scenarios can pause or deactivate without a warning you will notice. Consecutive errors take the schedule offline. If nobody owns the Make inbox, the deactivation email goes to a freelancer who left in March.

n8n can show success while doing nothing. A node returns an empty list, an IF takes the false branch, an HTTP node gets a 200 the next step ignores. The execution is green. The spreadsheet did not update.

Absence of alerts is not evidence. Prove a flow works by finding a recent successful run that produced the side effect you expect. If you cannot find that run, treat the flow as failed.

Day-one check: pick the three flows that touch customers or money. Find the last run you would defend. If it is older than the flow’s cadence, it is already failing.

Step 1 — Freeze, don’t delete (first 48 hours)

The instinct is to clean up. Delete the “test” Zaps. Remove the freelancer. Turn everything off until you understand it. That is how you lose the only evidence you have.

Safe to pause vs never delete

Pause is reversible. Delete is not. You may pause a flow that is sending wrong data. You do not delete anything or remove people until you have exported what they built.

Zapier history is short. Retention is typically 29 to 69 days, with up to 60 days guaranteed, and 10,000 runs published for Zap history. Read both: Customize data retention and Zap history. Wait another month and you may have nothing left.

n8n Cloud: Starter keeps 7 days and 2,500 executions. Pro keeps 30 days and 25,000 (Cloud data management). Inherited Starter? Export this afternoon.

Make history you can export, but incomplete executions are off by default. Successful history is the highlight reel.

Safe in the first two days: pause a flow that is sending wrong emails or looping; leave everything else on; screenshot the live lists; do not rotate keys unless one is leaking.

Never: delete Zaps “to tidy up”; remove the freelancer before you export; rebuild in a new account; add an AI step to “fix” a flow you do not understand.

If a flow is hurting customers now, pause it and write why, when, and who might already have been hit.

Export now: Zapier member CSV, Make history, n8n JSON, assistant text

Export is the freeze. Do it before anyone loses access.

Zapier. Export the member list. The CSV link expires in 7 days (Remove members). Download it to a company folder. Screenshot recent Zap history while it still exists.

Make. Export each live scenario’s history to CSV. Note whether incomplete executions are on. Turn them on after you understand the scenario. Save the blueprint if you can.

n8n. Download each workflow as JSON. Screenshot credential names (not secrets).

Assistants, custom GPTs, and prompts. Copy the system prompt, tool list, and action URLs. If they live on a personal ChatGPT seat, you do not own the agent. The Assistants API shut down on 26 August 2026 (Migration). Export this week even if it still runs.

Put every export in a company drive, dated. This folder should not get messy.

Step 2 — Take back ownership, logins and billing

A takeover that leaves billing on a personal card is not a takeover. Work this in parallel with the freeze. You need access before history dies.

Zapier: transferring Zaps and connections when you remove a member

Private connections keep working for 30 days, or until they expire, whichever comes first. Shared connections stay. Zaps can be transferred. The member CSV link lasts 7 days (Remove members).

Do this in order:

1. A company user is admin or owner. If the only owner is the freelancer, transfer before you remove them.
2. Export the member CSV. Names, emails, roles, in writing.
3. Note private vs shared connections. Private Gmail on alex.freelancer@gmail.com is a ticking clock.
4. Recreate those connections on a company Google / Slack / HubSpot user.
5. Transfer Zap ownership to the company user.
6. Then remove the freelancer.

Reverse that order and you get 30 days of “it still works”, then a quiet death that looks like “turned off due to an app connection issue”.

Make: scenario ownership and the personal card problem

Make scenarios have owners. The owner is often the freelancer, on a personal email and a personal card. There is a Make community thread that should scare you enough: someone was still being charged two years after leaving a startup. Assume that is true of yours until you open the billing page.

Check the org owner, the card, who owns each scenario, and where deactivation emails go. Move billing and the org to the company. Transfer ownership. Then remove people. If Make will not transfer a scenario, duplicate it under the company user and pause the original until the duplicate proves itself.

No owner login? Treat the org as unowned. Pause anything that can send or charge.

n8n: copy workflows and credentials before you delete the user

Do not delete the user first. Copy their workflows and credentials while you still can (Manage users). Transfer every active workflow to a company user, recreate credentials in the store (do not paste tokens into Slack), confirm a test execution, then delete the old user.

If n8n is on the freelancer’s VPS, snapshot the machine, copy the JSON, and stand up an instance the company pays for.

OpenAI and other AI: org owner, keys, spend alerts

The Zap looks cheap. The OpenAI project on a personal account is not.

Check: company OpenAI organisation or a personal ChatGPT key; who can create and revoke API keys; spend alerts (Production best practices); which flows hold that key; Assistants, Chat Completions, or Responses.

Revoke keys only after a company project, a new key, and a list of every consumer. Custom GPTs: copy instructions, actions, and knowledge file names. Personal seat: move or retire.

The rule: company email, company card, shared connections

Every production automation runs on a company email, a company card, and a shared connection. Personal Gmail, personal Make billing, and a freelancer’s OpenAI key are the outage.

If you do not want Zapier Team yet, put automations@yourcompany.com on the tools. On an AI Solutions takeover, billing moves to their email.

Step 3 — Inventory everything that runs

A Zapier folder called “Old” is not an inventory. One row per live flow is, including the ones you think are off.

Zapier: the Zaps list and the Alerts page

Filter Zaps to on, then look at off. Off is “this used to do something”, not safe. Note name, owner, apps, on/off.

Then open the Alerts page. Issues from the last 7 days. Disconnected apps are Critical. It lists the Zap owner (Alerts page).

For each on Zap: last success, last error, held runs, private connection. Empty history means it never fires or retention ate the evidence.

Make: scenarios, history, incomplete executions

List every scenario, including “Archive”. Open history on the active ones.

Incomplete executions are off by default (Incomplete executions). A module failed, Make stored nothing, the schedule moved on. Turn the setting on once the scenario is safe to retry. Until then, “no incompletes” means you were not looking.

Note schedule, last execution, last status, error handler. A 15-minute scenario last run 9 days ago is already silent.

n8n: active workflows, executions, credentials

List active workflows, then inactive ones that still have a production webhook or cron. Confirm executions still exist. A credential named Joe's Google is a Step 2 leftover. Open the last execution. Green plus empty is a silent failure.

The hidden layer: custom GPTs, agents, Apps Script, CRM, cron, webhooks

The Zapier bill is not the stack. Add rows for custom GPTs with actions, OpenAI Assistants, Apps Script triggers, HubSpot or Salesforce mail rules, cron on a VPS, and webhooks in Stripe, Typeform, Calendly, or Shopify that point at your automation tools.

Follow the data: “what happens when a form is submitted”. A Calendly webhook into a Zap into an Apps Script is three rows.

Fill-in inventory table

Use these columns. Do not add fifteen more.

Last successful run tells you if it is alive. What breaks if it stops is how you rank. purpose: unknown beats a missing row.

A 12-Zap inherited account, walked through

Twelve Zaps. Freelancer left five weeks ago. You have owner access this morning.

Do not fix twelve things. Pause 2, 4, and maybe 12. Treat 3, 8, and 10 as the dangerous three. Put 7 on the cost list. Turn 11 on only after you know who it emails. Leave 1 and 5 until ownership moves.

Step 4 — Rank by risk: which 1–3 flows to fix first

Repair the whole folder and you will spend the week on a Slack emoji Zap while the invoice email stays dead. Score each row 1–5:

Add them. Fix the top one to three. Everything else is leave on, pause, or retire.

In the 12-Zap account: invoice paid first; double GPT drafts second; Mailchimp unsubscribes third. Weekly report can stay off. Pipedrive can stay paused. Typeform only needs a connection move.

Is it safe to turn off a workflow nobody can explain?

Yes, if you can name what it touches and you will watch for a week. No, if it stands between a payment and a receipt, or a form and a CRM, and you have not checked the last successful run.

The test is what happens to a customer or a ledger if this stops for seven days. Slack gets quieter: pause it. Invoices go unacknowledged: fix it, or replace it with a checklist the same day. Unknown plus money is this afternoon.

Write the decision on the row. Pause, 2026-09-22, watching Stripe is professional. Leave it on in case is how you inherit the same mess.

A scoring sheet for new work is a different job: the AI Automation Audit. This week is what already runs.

Step 5 — Find out why each one is failing

This is the core. You have one to three flows. You need a cause. Read the last twenty runs, not the last one. A pattern is the diagnosis.

Zapier: auto turn-off, held runs, connection deaths

Zapier will turn a Zap off for you. If a Zap errors 95% of the time it runs and has run more than 20 times in the past 7 days, it switches off (Zap is not running). That is how a flow dies while everyone assumes Zapier emailed.

Grace window: 24 hours on Team, 72 hours on Enterprise. Override: “Keep running if errors occur”. That keeps a noisy Zap online, and a broken Zap sending partial data. Use it only when you understand the error.

Held runs: a task limit can hold all actions, not just the new ones (Troubleshoot held runs). If the invoice Zap “stopped” on the 28th, check the task count before you rewrite the mapper.

“Turned off due to an app connection issue” is Step 2 coming due. Reconnect on a company account. Do not rebuild because the toggle is off.

Autoreplay hides the first failure. A 9am complaint and a 7:30pm “success” may be a replay. Read the original run.

Worked example, Zap 3. Fourteen errors in 11 days, then off. Gmail invalid_grant. It hit the 95% / 20-run rule. Reconnect Gmail on automations@, transfer ownership, turn it on, replay what history still has, send the rest by hand.

Make: deactivation, rollback, ordered processing, retries

Make deactivates a scenario’s schedule after consecutive errors. The threshold is “Errors before deactivation” (Scenario settings, Errors and warnings). A threshold of 1 means one 429 took production down.

Rollback is on by default. A later module fails, Make undoes earlier ones. Good for “create record then send email” if you understand it. Confusing when the CRM record vanished. Read the incomplete execution.

“Process data in order” stalls the queue. One bad bundle blocks everything behind it. If a fast scenario now sits for hours, check that setting first.

Connection and rate-limit problems can pause a scenario for 20 minutes. If the API is still unhappy, another pause. From the outside it looks slow.

Incomplete executions retry: 1, 10, 10, 30, 30, 30 minutes, then 3 hours + 3 hours, about 7 hours and 51 minutes (Automatic retry). The Retry handler default is 3 attempts over 15 minutes. Different clocks. No handler and incompletes off means the failure disappeared.

Make incomplete-executions recovery. A Stripe charge.succeeded receipt scenario deactivated 12 days ago. Incompletes were off.

1. Do not turn the scenario on yet.
2. Turn incomplete executions on. Set Errors before deactivation to a number you can live with, not 1, not 99.
3. Add an error handler to a company inbox or Slack. Confirm it is not the freelancer.
4. Run once on a single known Stripe event. Watch every module.
5. If it completes, turn the schedule on.
6. For the 12 missing days, Make will not replay what it never stored. Export Stripe events and process them once, or send receipts by hand.
7. Runbook: incompletes on, threshold N, who gets the deactivation email.

n8n: green runs, pruning, missing error workflows

The common inherited failure is success-but-nothing. A Set node mapped the wrong field. The IF node went left. The last node ran on an empty item. The execution is green. Open it and read items at each node. If Slack received [], the bug is upstream.

Second: missing history. Self-hosted n8n prunes execution data by default after 336 hours (14 days) and 10,000 executions (Execution data). Arrive on day 20 and the evidence may be gone. Cloud Starter is 7 days / 2,500. Export JSON now.

Third: no error workflow. n8n will not email you unless you built that (Error handling).

A minimal n8n error-workflow recipe. Create ops-error-router. Trigger: Error Trigger. Extract workflow name, execution URL, last node, error message; send Slack or email to the company channel. No business logic. On every production workflow, set Error Workflow to ops-error-router. On flaky API nodes, Retry On Fail. In settings, save failed executions.

Green plus missing side effect: the error workflow will not fire. Pair it with a heartbeat (Step 7).

Shared causes: Google OAuth, passwords, API changes, 429s, plan limits

Google OAuth apps in Testing issue refresh tokens that expire after 7 days (Google OAuth 2.0). An “internal” Cloud project that never went to production dies weekly. That is Google’s consent screen. Fix: a production OAuth app on a company Cloud project, or the vendor’s native Google connection.

Password policies do the same more slowly. Someone rotated HubSpot, forced 2FA, or disabled a contractor’s Google. Check the identity still exists in the target app.

API changes break mapped fields. Typeform renamed, Stripe event version bumped, HubSpot property archived. Fix the mapper.

429 is rate limiting. Make pauses ~20 minutes. Zapier errors and may replay. n8n retries if you told it to. If the flow bursts every night, delay or lower the schedule before you raise plan limits.

Task and operations limits hold or stop everything. Check the plan before you rewrite.

AI-specific failures: Assistants shutdown, GPT actions, invented params, tokens

Assistants API shut down on 26 August 2026 (Migration). If anything still answers on a leftover path, export the instructions this week.

Custom GPT actions fail like Zaps: URL moved, auth stale, spec mismatch. Open the GPT, list the actions, hit the URLs. If it “stopped being smart”, it may have stopped calling your tools.

Hallucinated parameters look like success. The model invents a deal_id, the CRM returns nothing, the Zap continues. Guard with a lookup that must return exactly one record. Zero results: stop.

Token cost is the runaway-cost bucket. A support-draft Zap that includes the whole ticket on every update will spend more on a Tuesday than the Zapier plan costs in a year. Usage by key and by day. No spend alerts means you are not in production (Production best practices).

Double-draft Zaps: pause v1, confirm v2’s prompt, cap tokens, move the key. An afternoon, not a migration.

Step 6 — Fix, rebuild or retire?

Once you know the cause, pick one of three. Most inherited flows need a fix. A few need to die.

Fix: reconnect, correct the map, turn on incomplete executions, add the error workflow, replay what you can, cover the rest by hand.

Rebuild: same job, cleaner path, company-owned connections, alerts from day one. Rebuild one flow when the existing one is unreadable.

Retire: pause, document why, delete after you have lived without it. A Pipedrive Zap for a CRM you do not use is not an asset.

When Zapier to n8n is worth it (and when it is not)

A full migration is a project, not the takeover. Spofylabs does not treat “move everything to n8n” as part of a 5–10 day fix unless it is a separate quote.

Worth a conversation when you already pay Zapier task overage on high-volume, stable flows, and someone can host or pay for n8n Cloud. Not worth it because a blog said n8n is cheaper.

If you are curious about building the next workflow on a different stack: How to Automate Your Business with AI. This week, pick fix, rebuild-one, or retire.

Write the outcome on the row. Fixed Gmail, replayed 6 runs, 3 invoices sent by hand. That is what the Care month will read.

Step 7 — Turn on alerts that catch silence, not just errors

Error emails catch red runs. They do not catch a Zap that never fired, a scenario that deactivated, or an n8n workflow that succeeded with empty data. You need failure alerts and heartbeats.

Zapier: notifications, Alerts page, autoreplay, the override

Turn on error notifications to a company inbox or Slack (Manage error notifications). Book a weekly look at the Alerts page. Disconnected apps are Critical. The page covers 7 days.

Leave autoreplay on when a late send is better than no send. Be careful when a late send duplicates a charge. The error email still waits for the last replay.

Do not use “Keep running if errors occur” as your alert strategy. Use that override only for a known, bounded error. If you do not know the error, the auto turn-off is doing you a favour.

If the inherited “Error notifier” Zap was off, see who it emails before you flip it on.

Make: handlers, deactivation threshold, deactivation mail

Every production scenario needs an error handler that notifies a human. Set Errors before deactivation to a number that matches the flow. Confirm deactivation emails go to the company owner. Incomplete executions on, handler in place. If you only enable incompletes and never tell anyone, you have a queue of retries nobody is looking at.

n8n: one shared error workflow, Retry On Fail, save failures

Use one shared error workflow. Point every production workflow at ops-error-router. Retry On Fail on outbound HTTP nodes. Save failed executions so prune does not erase a 2am failure. Upgrade or export if you are on Cloud Starter.

Heartbeats: catch “it did not run at all”

A heartbeat asks: did the flow run when it should have? A daily check if a daily flow is more than 36 hours stale; zero new CRM contacts from the form Zap; Stripe charges without a receipt tag; a Slack message the weekly digest always posts.

One heartbeat per dangerous flow. Invoice Zap: charges without receipts. Lead Zap: forms without contacts.

OpenAI spend alerts

Set them on the company project (Production best practices). A hard cap plus an email you would notice. Same inbox as Zapier and Make. If spend alerts go to the freelancer, you will learn about the bill when finance forwards it.

Step 8 — Write the one-page runbook

One page per dangerous flow. A 40-page Confluence space will not get read.

FLOW: [name as it appears in Zapier / Make / n8n]
PURPOSE: [one sentence, customer language]
TRIGGER: [app + event, or cron]
OWNER: [role, not a person who might leave]
COMPANY LOGIN: [password manager item]
CONNECTIONS: [which Google / Slack / HubSpot user]
CREDENTIALS LOCATION: [ops vault / n8n credential name / OpenAI project]
LAST PROVED SUCCESS: [date, link to a run]
KNOWN FAILURE MODES:
  - [e.g. Google OAuth Testing = 7-day token]
  - [e.g. Zapier 95% / 20 runs auto-off]
HOW TO REPLAY:
  - Zapier: [replay from history; what not to replay]
  - Make: [incomplete executions; Stripe backfill if history is gone]
  - n8n: [retry execution; when to re-send the webhook]
WHO TO CALL:
  - Ops: [name]
  - Vendor admin: [name]
  - Spofylabs Care / internal backup: [name]
DO NOT: [e.g. delete this Zap; hit replay on the charge step]

Pin it. Someone new should replay a failed invoice without calling you. Update it when a connection, threshold, or owner changes.

How long a takeover takes, and DIY vs hiring help

A focused takeover is 5–10 days. History, access, and the dangerous three take calendar time. Waiting on the freelancer for owner transfer can eat two of those days.

DIY if you have one to three simple Zaps, you already own the account, and the failure is a dead connection or a held task limit. Follow Steps 1–8.

Hire help when the owner login is not yours; more than a handful of flows touch customers or money; Make incompletes were off; n8n is on a personal server; or you cannot name what a live flow does.

What a fixed-scope Takeover includes at Spofylabs: inventory of every live flow; fix the 1–3 dangerous ones (wrong sends, silent failures, exploding cost); alerts on failure; a one-page runbook; ownership and billing moved to the company. $600–$1,200, 5–10 days. Not a full-stack rebuild, not a Zapier-to-n8n migration, not unlimited new scenarios.

Care is the month after: $1,000 / month, a weekly check, a human on the failures. New workflows are a separate quote.

One simple Zap you already own: DIY. Twelve Zaps on a personal card: a Takeover week is cheaper than another quiet month.

FAQ

Will turning the Zap back on replay missed tasks? No. Missed events stay missed unless you replay them from history, and history only goes back as far as retention (29–69 days / up to 60 days guaranteed, 10,000 runs). Autoreplay only retries a failed run. For gaps, export the source app and process those records on purpose.

Why does my Make scenario keep deactivating? Consecutive errors hit Errors before deactivation, then the schedule stops (Scenario settings). Check the repeating error, the connection, and whether “Process data in order” is stalled. Fix the error, set a threshold, point deactivation mail at a company inbox, turn incomplete executions on.

Why do Google connections expire every 7 days? The OAuth client is likely in Testing. Refresh tokens last 7 days (Google OAuth 2.0). Publish to production on a company Cloud project, or use the platform’s native Google connection.

Where did n8n execution history go? It was pruned. Cloud Starter: 7 days / 2,500. Pro: 30 days / 25,000. Self-hosted default: 336 hours and 10,000 executions (Execution data, Cloud data management). Export workflow JSON. Save failed executions.

Our OpenAI Assistant stopped working. What now? Check the key, the org, and spend. Then check whether you are on the Assistants API, which shut down on 26 August 2026 (Migration). If it still runs, export instructions this week.

Is it safe to turn off a workflow nobody understands? If it does not touch a customer or a ledger, yes: pause, watch a week, retire. If it does, no. Fix it or replace the job with a checklist the same day.

How long does an audit take? Inventory and ranking can be a day if you have admin access. A full takeover is 5–10 days. Waiting on owner access stretches it.

Do we need to migrate off Zapier to make this stable? No. Stability is ownership, history, alerts, and one or three fixed flows.

Take the stack back, then keep it boring

You do not need another automation. You need the ones you already paid for to belong to the company, fail loudly, and have a page that says what to do when they do.

If you want that done as a fixed week of work, Spofylabs Takeover is $600–$1,200 over 5–10 days: inventory, fix the dangerous flows, alerts, runbook, billing on your email. If you want someone on it after that, Care is $1,000 / month. Send the tool and what is failing. A short call is enough to tell whether it is a Takeover week or a flow to turn off.