The automation looks fine, because nobody sees what it drops.
A flow that fails loudly gets fixed the same day. The expensive ones fail quietly: a step errors, the item goes nowhere, the run log says finished, and the leads that never arrived are invisible. This reads one workflow and its recent runs, reports what it loses, and ranks the repairs by how many faults each one removes.
SimulationThe workflow and its runs are written for this demonstration, and the audit is real code running in your browser rather than a model. Nothing is connected to an automation platform. The production build reads the workflow and its execution log through the platform's API and audits the real thing.
New form submissions are enriched, added to the pipeline sheet, and announced in Slack.
6 of 20 runs failed, so the success rate is 70 per cent, and 28 per cent of the items that went in never came out. There are 5 faults to repair, 2 of them serious.
Lead capture to Sheets
Every step, in the order items travel. A step is marked when runs failed on it, and a step with a fault is outlined even when no run has failed there yet.
when the lead is new
Step detail
fetchEnrich the company
HTTP request, Clearbit
- Timeout
- 10s
- Retry
- none
- Header
- Authorization: ****
- Items in and out
- 1 in, 1 out
- On failure
- nothing, the item is dropped
- CriticalA credential is written into the step configuration
- CriticalA failing step drops the item silently
- HighA step that can fail transiently has no retry
Repair plan
4 repairs- 1Give the failing step an error branch, then a retry
Route failures to a step that records the item and tells someone, and add two retries with backoff on the call itself. This is the repair that stops the silent loss.
- 2Move the credential out of the step and into a secret
Store the token in a credential store or an environment variable, reference it by name, and rotate the exposed one.
- 3Set a dedup key so a resubmission is not a second lead
Use the record id or the email as the key, and let the dedup step drop a repeat instead of passing it on.
- 4Add a workflow level error handler
One handler that runs when the flow fails, records the run and tells a person. It is the safety net under every step.
Every fault, with what proves it
- CriticalA credential is written into the step configuration
Enrich the company sets Header to a literal secret (Authorization: ****), visible to anyone who can open this workflow.
Everyone with access to the automation has the key, and rotating it means editing the flow.
- CriticalA failing step drops the item silently
Enrich the company was the failing step in 6 of 20 runs, and it has no error branch.
Every one of those runs lost its items without a record. Nobody is told, so nobody follows up.
- HighNothing stops the same record being handled twice
New form submission has no dedup key, and Remove duplicates is configured with an empty key.
A resubmitted form becomes a second lead in the sheet and a second email to the same person.
- HighThe workflow has no error path at allwhole workflow
None of the 7 steps has an error branch.
Any failure ends the run where it happened. There is no retry, no alert and no record of the item.
- HighA step that can fail transiently has no retry
Enrich the company failed in 6 runs and its retry setting is "none".
Most failures of this kind are a timeout or a throttle. A retry would have saved the run.
How the audit works
Eight checks over the flow and its execution log. Every fault carries the evidence that produced it, and the repairs are grouped by root cause, so the plan stays short.
What it looks for
- A step that fails repeatedly with nowhere for the item to go.
- A step that can fail transiently with no retry set.
- A credential written into a step, named but never printed.
- Nothing stopping the same record being handled twice.
- An action sending at full speed with no ceiling.
- A schedule running in a timezone the business does not.
- A step nothing reaches, and a flow with no error path at all.
What it does not claim
- It reads the log it is given. A failure the platform never recorded is a failure this cannot see.
- It does not know your business rules. Nothing stopping a duplicate is a fault only because a duplicated lead costs a second email.
- It does not fix anything. It ranks the repairs by how many faults each removes and leaves the change to a person.
- The score summarises the weights. It is not a certification.