LANEFRAME · ZAPIER
Zapier build detail
3 of 3 builds complete
You pay per completed step, and each automation has a hard size limit. It forced the hardest trade-offs of the four.
It charged less than its own help pages said
A completed step costs 1 unit and a skipped one costs nothing. On one live run, Zapier charged less than its own documentation predicted, confirmed against the account’s real usage counter.
A size limit changed what got built
Each Zap is capped at about 32 steps. To fit, the build was split into smaller linked parts, and two planned features (text messages and company lookups) had to be dropped. n8n and Make kept both.
No way to export or back it up
On this plan there’s no file you can export to copy or restore a build, unlike n8n and Make. The live Zaps are the only record.
Z1
Lead Sorter & Demo Booker
Takes in leads, removes duplicates, scores them and sends each one to the right next step. It’s the same job as the n8n and Make versions, split into three linked parts to fit under Zapier’s size limit.
WHAT I THREW AT IT
Zapier’s bill didn’t match its own documentation
Zapier’s help pages said this run should cost 2 units; the account’s own counter showed it cost 1. A real gap between what the docs say and what actually happens, disclosed here.
The size limit forced two features out
To stay under the limit, the company-lookup and text-message steps were removed completely. On n8n and Make they were kept and fail gracefully. This build is smaller than planned because of Zapier’s size limit.
Worth knowing: Only 2 of the 5 standard break tests could be run. The other 3 tested features that were removed because of the size limit, so those features don’t exist in this build.
+ Show technical detail− Hide technical detailfor builders and technical reviewers
A webhook receives the lead, two Formatter steps normalize it, a guard-halted-logging Sub-Zap call and filter catch malformed intake, two HubSpot lookups guard against duplicates, and a three-way Paths router sends Founder/SDR/self-serve leads to their own CRM, Slack, and Sheets actions — with a second Sub-Zap handling existing-lead re-engagement and dormancy scoring.
Zapier’s bill didn’t match its own documentation
Zapier's help docs predict a 2-action Sub-Zap child bills 2 tasks for those actions alone. A real, live guard-halted run — confirmed against the account's own task counter moving by exactly +1, not assumed from a green checkmark — billed only 1. A genuine, disclosed gap between the platform's own documentation and its observed behavior, never resolved against Zapier support.
The size limit forced two features out
A permanently-failing enrichment call and an undeliverable Twilio SMS step were both removed entirely to free up step slots, rather than kept in a degraded, gracefully-skipped form the way the same failures were handled on n8n and Make. The finished build is structurally smaller than its original specification because of a platform constraint, not a design choice.
Methodology note: Only 2 of the 5 standard break probes were run live; the other 3 (downstream timeout, enrichment failure, out-of-hours delay) are disclosed as structurally not applicable — the features they'd test were the two dropped above, not features that were built and passed.
Z2
Lead Sorter: Email Version
The same lead sorter, but taking leads from incoming emails instead of a web form, because on Zapier each way in needs its own build.
WHAT I THREW AT IT
A normal-looking email was stopped by the safety check
A real referral email was read wrongly: two details came out wrong and the company website was missing. The safety check stopped it, as designed, so no bad data reached the customer database.
The second test found the real cause
Rewording the email fixed two of the three wrong details, but not the website. The cause was a setup mistake inside one step. It’s a real bug, not a one-off.
Worth knowing: This build has never taken a lead all the way through on a live test. Both attempts were stopped by the safety check, and it won’t until that setup mistake is fixed.
+ Show technical detail− Hide technical detailfor builders and technical reviewers
An Email Parser mailbox extracts lead fields from inbound referral emails, feeds the same fit-scoring and dedupe logic as Z1's router, and reuses Z1's own re-engagement Sub-Zap rather than duplicating it — a single 30-step Zap with no dedicated child Sub-Zaps of its own.
A normal-looking email was stopped by the safety check
A genuinely new referral email — not a synthetic malformed test — mis-extracted two fields and produced no domain at all, correctly triggering Path B's guard rather than routing to CRM logic with bad data. Zero HubSpot lookups fired, exactly as the guard is meant to do.
The second test found the real cause
Rewording the test email to match the trained template fixed two of three broken fields but not the domain extractor, which was found to be reading a stale, broken field reference — a real bug in the step's own configuration, confirmed by inspecting its live Data In (a literal "1" instead of an email), not a one-off fluke.
Methodology note: This build has never reached its own Founder/SDR/self-serve success path on a live run — both live-fire attempts hit the guard-halt before getting there, and the root-caused domain-extractor bug means a third attempt wouldn't be expected to succeed without first fixing that step's field mapping.
Z3
AI Support-Ticket Assistant
An AI assistant that sorts and answers support tickets from real documents and passes anything it can’t back up with a quote to a person. It’s the same job as the n8n and Make versions, but Zapier has no built-in approval step or confidence score.
WHAT I THREW AT IT
The hand-off rule failed half the time it mattered
It answered all 4 straightforward questions correctly. But on the 8 questions it should have passed to a person, it got 4 wrong, once even denying it could hand off at all.
An urgent alert never went out, and nothing flagged it
Given a ticket about an outage, the AI said it would post an urgent Slack alert, then never did. Its own log confirms it, and nothing warned a person that the alert was missed.
Worth knowing: When a connected app genuinely failed, the AI flagged it clearly and gave a manual workaround. It handles that kind of failure well. The judgment calls above are where it doesn’t.
+ Show technical detail− Hide technical detailfor builders and technical reviewers
A support ticket is classified, checked against two attached knowledge documents, and either answered with a quoted passage or escalated. Nothing reaches a customer without human approval — enforced entirely by a system-prompt rule and the deliberate choice of a draft-only Gmail action, since Zapier's Agents builder has no dedicated approval-gate toggle at all.
The hand-off rule failed half the time it mattered
8 of 12 probes passed overall, but the coverage rule — "if you can't quote a passage, escalate" — failed on 4 of the 8 uncovered and near-miss questions, including one case where the agent flatly denied having any escalation process when asked to invoke it. Zero failures occurred on the 4 covered questions; the gap is entirely in judgment calls, not lookups.
An urgent alert never went out, and nothing flagged it
Fed a ticket containing the system prompt's own trigger phrase for an outage, the agent narrated an intention to post a Slack alert and then never did — confirmed by the run's own tool-usage log, which shows no Slack call at all. The ticket was still otherwise handled correctly, but the one safety action its own rules were written to guarantee simply didn't fire, and no summary or log flagged that gap to a human.
Methodology note: A genuine Linear tool failure (a required Team field with no valid default) surfaced correctly — the agent flagged it plainly and produced a manual fallback a human could act on — which is the one failure class this build handles reliably. The prompt-level judgment failures above are the one it doesn't.
Want your next automation tested like this before your customers touch it?
Fix My BottleneckSEE THE SAME JOB BUILT ON ANOTHER TOOL