INSIGHTS
AI & Automation Insights for Operators
Practical guides on AI assistants, workflow automation, lead intelligence, and field operations. Written for business owners and operators who want results, not hype.
Why your automation is stuck waiting for approval
An automation stalls at an approval step because three separate clocks run on every approval: the workflow's wait timeout, the lifetime of the callback behind the button you sent, and the reviewer's own response time. Setting only the wait timeout leaves you with an approve button that still looks live after the run behind it is gone, so the durable fix is to store the approval as a decision record the workflow reads, not as a pause it holds open.
Why automation projects fail: three wrong assumptions
Automation projects usually fail because the model of human behavior encoded in the workflow is wrong, not because an integration broke. The three most expensive wrong assumptions are that a field edit is a decision, that a customer reply is a command, and that your inventory of what already exists is accurate. All three pass testing and only surface under real use.
Why your CRM shows three of the same customer
Repeated customer records usually come from a schema that stores one person per customer, not from weak matching. When a household contains several people with different phone numbers and emails, every channel creates its own record, and tightening the matcher makes it worse because aggressive matching is what merges the wrong people permanently. The fix is a contacts layer beneath the customer record, plus an ordered resolution rule that says which record wins.
API Deprecations Break Automations Two Ways
Vendor API changes break automations in two distinct shapes. Managed connectors like Zapier drift, because the platform migrates you through non-breaking changes automatically and stops precisely at the breaking one. Pinned integrations like n8n nodes and version-locked HTTP calls never change at all, then fail completely on the vendor's retirement date. Each shape needs a different control, and most teams have only built the one that catches the loud failure.
How Much AI Output Should a Human Review?
Set AI review as a fixed count per period, not a percentage of volume, because the confidence a sample buys depends on how many items you check rather than what fraction of the total they represent. Checking 59 consecutive outputs and finding zero errors bounds the true error rate below 5 percent at 95 percent confidence, whether you process 500 items a month or 50,000.
Deleted records don't sync, then they come back
A sync misses deletions because polling asks which records changed since a timestamp, and a deleted record is not returned by any query, so the delete is an absence rather than a change. Even after you detect it, propagating a hard delete removes the destination row that was holding your deduplication key, so the next update from the source finds no match and re-creates the record. The durable fix is the tombstone pattern databases use: sync an active flag and a deleted-at timestamp, and never remove the row.
Automation running twice? It is overlapping itself
A scheduled automation processes the same record twice when the next run starts before the previous one has written back its result, which happens whenever run duration approaches the schedule interval. Platform settings that promise one run at a time are throttles rather than locks: n8n's concurrency limit applies to the whole instance, Make's sequential processing halts on any unresolved incomplete execution, and Zapier documents that its queue delay does not guarantee steps never run simultaneously. The reliable fix is to claim each record with a status write before doing the work, so an overlapping run finds nothing left to take.
Opt-Outs Don't Sync Across Your Tools
Opt-outs stop only the tool that received them, because every messaging platform, CRM, and email system keeps its own consent flag and none of them tell the others. Federal rules give you ten business days from the moment a request is made in any reasonable manner, including phrasings your vendor's keyword parser never sees, so the fix is one consent record of record that fans out to every sender rather than a set of per-tool flags that drift.
Large files break automations. Pass links, not bytes
Automations fail on large files because every step that materializes the file holds its own copy, in memory and then in the retained execution record, and base64 encoding inflates each copy by about a third. The fix is to upload the file to storage once and pass a URL or object key between steps so exactly one step ever holds the bytes. Airtable's own API shows the size of the difference: the same attachment field accepts 5 MB when you send the file bytes directly and 5 GB when you send a link.
Matching customer records without a shared ID
An automation should link a record across two systems only on exact normalized keys: a stored cross-reference ID, a lowercased email, or a phone in E.164 format. Everything that merely looks similar goes to a human review queue. The reason is asymmetry, not accuracy: HubSpot states that records cannot be unmerged, so a false merge is permanent, while a false split costs one manual merge.
Webhooks arriving out of order? Fix the write, not the sort
Webhooks arrive out of order because sources like Stripe and Shopify document that they do not guarantee delivery order, and because automation platforms run webhook-triggered executions concurrently, so two runs for the same record can finish in the wrong sequence. Sorting by timestamp only addresses the first cause. The reliable fix is to treat the webhook as a signal, re-fetch the record from the source API, and guard the write with a per-record watermark so a late run cannot overwrite newer data.
Your automation only synced the first 100 records
An automation that syncs only the first batch of records is usually missing pagination, but the failure that survives correct pagination is offset paging over data that is being written to, which both skips and duplicates rows. Because a partial sync exits cleanly and reports success, the only reliable detector is comparing the source count against the destination count on every run.
Want a system like the ones we write about?
We build practical AI and automation systems for businesses that need better workflows, connected tools, and cleaner operations.
Request a Workflow Review