Twenty-Nine Automations and a CRM That Updates Itself
A direct-to-consumer supplement brand with 90,000 customers, 11,000 active subscriptions, about 900 affiliates, a wholesale channel, and an outsourced call center · Apr 2026 to present
- Pipedrive
- Shopify
- n8n
- Klaviyo
- +1
The Challenge
"Everything is carried in manually. The system doesn't update itself." The CRM was a silo. Outcomes were tracked in a spreadsheet. Callers "were just getting voicemails all day." Leadership's numbers were "probably about a third" of reality. A prior contractor's automation had created thousands of duplicates, and the ops lead was wary: "it will clear all labels in the system if it's done by an automated way."
Wholesale orders were photographed on paper and re-keyed twice.
Our Approach
- A read-only API audit of the CRM, the store, and the subscription platform before proposing anything: 0 webhooks, 1,682 duplicate clusters, 63 rotting custom fields
- A canonical field spec and a tag allowlist agreed with the ops lead before the first write
- Every workflow activated one at a time behind a sign-off gate
- Every bulk change as dry-run first, then commit with a per-run rollback log
The Solution
- 29 workflows: orders, subscriptions, refunds, tracking, delivery-confirmed call tasks, affiliate revenue, and reviews and webinar events
- A 42,000-record historical import and about 5,400 duplicate merges, every run with a rollback log
- Affiliate attribution written as unit-tested code and replayed over the full customer history
- Two workflows handed over by the ops lead, re-derived against the live API and dry-run before a single write; the texting one was never imported and nothing was sent
- Derived CRM properties pushed nightly into the email platform for segmentation
- An operations dashboard built from read-only snapshots of five systems
When We Said No to AI
The CEO wanted AI callers. We evaluated it, priced it, and recommended deferring: the customer base is older and values a human voice, and the call data showed the fix was call timing, not call volume. A ringless-voicemail campaign with an AI voice clone was built and proven, then shelved when the clone quality failed the client's bar. That was the right call, and it's why the automation engine got built instead.
Results
- CRM coverage grew from 13,097 to 96,659 people.
- 29 workflows running 150 to 190 times a day with zero-error sweeps.
- Unresolved affiliate attributions cut from 7,870 to 2,362.
- Connect rate of 22 percent at 4 p.m. versus 7.5 percent at 7 a.m., which changed the calling schedule.
- Affiliate attribution replayed over 98,364 customers and verified on 52 of 52 recent affiliate orders.
No AI in the production system. Here's why that was right: every event had a deterministic destination. We used AI to run the audits, generate and patch workflows, and write the reconciliation scripts.
All 29 workflows, the attribution rule, the dedup passes, the allowlist, and a safe-to-send gate built for outbound texting, not yet deployed.
The mechanisms behind it
Facing a similar challenge?
Let's talk about what's going on with your systems. 30 minutes. No pitch deck. We'll tell you if AI is the wrong tool.
We'll respond within 24 hours.