An Australian consumer finance marketplace, seed stage, six people, one of them technical. Webflow front end, Make for automation, Zoho CRM holding the leads. A paid acquisition push budgeted to roughly triple inbound volume.
- Surface audited
- 9 automation scenarios, 4 CRM modules, 3 integration points
- Effort
- 61 hours, scope to handoff
- Findings
- 11 total, 3 critical and 5 high
- Deadline
- Campaign live 12 September
The situation
Enquiries went from a Webflow form through a chain of automation scenarios into a CRM lead record and a broker assignment. They wanted an architecture audit and a production readiness review before the campaign went live on 12 September.
Where the operations budget goes
The constraint is the automation operations budget, and the surprise is where it goes. Lead capture costs 14 operations per lead, which at 380 leads is 5,320 a month against an allowance of 10,000. That alone leaves room.
But two scenarios poll on a fifteen minute schedule, and a scheduled run consumes an operation whether or not it finds anything, so those two spend 2,900 operations a month checking for work that arrives by form anyway. Add the rest and they sit at 88% of the allowance at current volume, which puts the ceiling at 464 leads and leaves headroom of 84. When the allowance runs out, scenarios stop, and the form keeps accepting submissions and showing a success message.
We found this in the execution history rather than in the scenario design, where it is not visible. The same history is how we got the failure rate, and it comes with a caveat: incomplete execution storage was switched off on six scenarios, so 1.8% is what we could measure on the three that had it on, and it is a floor rather than a figure.
Replacing the polling with webhooks and moving enrichment into a batch run takes two days and is mostly deletion. Doing it after the campaign starts costs three to four times that, because the work stops being a config change and becomes a reconciliation, matching whatever was written during a stopped window against live broker assignments by hand.
What we found
Ordered by severity. The criticals set the sequence below.
- Critical
Operations budget exhausts at 464 leads a month
They were running 380, with a campaign booked to roughly triple inbound volume.
- Critical
Two polling scenarios consume 2,900 operations a month and produce nothing
Both poll on a fifteen minute schedule for work that arrives by form anyway. A scheduled run costs an operation whether or not it finds anything.
- Critical
Error routes absent on six of nine scenarios
Failures write nothing and say nothing. Silent loss, invisible in the scenario design.
- High
Lead creation inserts rather than upserts
Resubmitted forms duplicate, and duplicates reach broker assignment.
Constraints and trade-offs
They pushed back on moving off the automation platform entirely, which was the right call with three weeks on the clock, so we restructured inside it.
They declined a second CRM organisation for staging on licence cost, also reasonable at their size.
We lost most of a day to the CRM administrator being on leave with no one else holding the workflow rules.
Sequencing
Error routes first (half a day)
Stopped silent loss the same afternoon.
Replace polling with webhooks (two days)
Before any campaign spend went live. Mostly deletion.
Deduplication (deferred past launch)
Their call. Re-checked at day 14 and cleaned three duplicate pairs by hand.
Staging environment (deferred to Q1)
Licence cost, reasonable at their size.
Outcome
Operations per lead came down from 14 to 5 and the polling scenarios went away, which moved the ceiling from 464 to somewhere north of 1,800. The campaign launched on 12 September as planned. We re-checked volume and error rate at day 14 and found 3 duplicate lead pairs from the deferred dedupe work, cleaned by hand.
| Metric | Before | After |
|---|---|---|
| Operations per captured lead | 14 | 5 |
| Monthly lead ceiling before scenarios stop | 464 | 1,800+ |
| Campaign launch | At risk | 12 Sep, on plan |
Build with
confidence.
AI builds the product. Vibecop makes sure it won’t break in production, fail under scale, or expose your users to risk. One audit. Fewer expensive surprises.