Meridian Home Co — Operational Audit

• 75–100 hours/month on comment processing alone. That's 45–60% of Sofia's working time on a workflow with a compounding volume problem.
• Inbox volume is growing with campaign success. The spring home decor campaign drove the inbox from ~90/day to 241/day in a single week. Seasonal peaks will push this further.
• The workflow scales linearly with engagement growth. The better she does at social content (which she's clearly good at), the worse this problem gets.
• Sofia grew Pinterest saves from 4.2K to 11.8K this quarter with a content strategy she built herself. She runs informal A/B tests on caption styles. She proposes seasonal themes ahead of schedule. The inbox grind is eating all of that capacity.

This isn't just about saving time. A properly configured pipeline would produce better replies than the current process:

Sofia currently has no system prompt. Every Gemini conversation starts fresh. No brand voice. No return policy. No CS routing rules. No platform context. She's getting generic AI output and then manually correcting 100% of it.

43% of her edits are adding information a system prompt would include automatically — things the business already knows:

• Support email: [email protected]
• Return window: "30-day hassle-free returns"
• Free shipping threshold: "Free shipping on orders over $75"
• Product care links and collection page URLs

A system prompt containing brand voice, return policy, free shipping threshold, CS routing rules, and platform metadata (which Hootsuite provides via API) would include all of this automatically. The first drafts would be better than what Gemini gives Sofia cold — because they'd have context Gemini doesn't.

Consistency improves at scale. A pipeline doesn't get tired at comment #118. It doesn't forget the return window on the 82nd reply. It doesn't shorten replies as the day wears on.

The validation is fast: We start with a proof run — 20 recent comments, brand-voice replies generated, you review them. 5 minutes. We adjust and send 20 more. Three rounds of this — 15 minutes total — and the system is calibrated. After that, the pipeline runs and Sofia reviews exceptions, not the whole inbox.

Tier 1 Manual Fix — Today, Zero Code

You are the social media voice of Meridian Home Co, a DTC home goods brand.
Reply to customer comments on our social media posts.

Voice: Warm, genuine, inviting. Short replies (1–3 sentences max). Never corporate or generic.

Returns: "We offer 30-day hassle-free returns — reach out to [email protected]"
Shipping: "Free shipping on orders over $75. Standard delivery is 4–7 business days."
Support: "Our team at [email protected] is happy to help!"
Damaged items: Acknowledge, apologize sincerely, always route to support email.
Negative comments: Acknowledge, empathize, offer to help, route to CS.

When the comment is simple (emoji, "love this", "gorgeous"), reply with 2–4 words.
When it's a question, answer directly and briefly.
When it's a complaint, empathize and route to CS email.Copy

Time saved: ~25–45 min/day. Cost: $0. Sofia can do this today.

Tier 2 Dashboard Tool — Prototype Built

We built a batch comment processing tool (meridian-comment-engine.html, 41KB). It lets Sofia:

• Paste 50 comments at once, or import CSV from Hootsuite export
• System prompt pre-loaded with Meridian's brand voice (extracted from her 180+ actual replies)
• Generates all 50 replies in parallel via API
• Review interface: approve, edit, or reject each reply
• Copy approved replies back to Hootsuite

Time saved: 1.5–2.5 hours/day. Cost: $2–8/month in API calls. Limitation: Still requires manual copy-paste between tool and Hootsuite — better process, same nails.

Tier 3 Full Automation Pipeline

Architecture: Hootsuite Inbox API polls for new comments automatically. A two-model pipeline processes them:

1. Classifier model (fast/cheap — Gemini Flash, $0.075/1M tokens): Categorizes comment by type (question, praise, complaint, spam, CS-route) and platform. Tags with metadata including post context.
2. Reply model (with trained brand-voice system prompt): Generates replies using Meridian's voice — extracted from Sofia's 180+ actual captured replies — plus return policy, free shipping threshold, and platform-specific norms (Instagram vs TikTok vs Pinterest vs Facebook).
3. Confidence scorer: High-confidence replies (simple questions, praise, standard routing) auto-post. Low-confidence ones (complaints, damaged items, edge cases) queue for Sofia's review. Estimated exception rate: 10–15% of items.

Sofia goes from processing 118 items/day to reviewing 12–18 exceptions. The pipeline improves over time — every edit she makes to an exception becomes training signal.

Time saved: 3–4.5 hours/day. AI compute: $2–10/month. We can have this running within a week. Confidence: 91%. We have her brand voice, return/shipping policy, routing rules, reply patterns, and 180+ examples of her actual responses — all captured from the data.

API Documentation & Implementation References

• Hootsuite Inbox API — inbox stream endpoints, comment retrieval, reply posting
• Gemini API — Flash model for classification ($0.075/1M input tokens)
• Claude API — quality checking and brand voice generation
• OpenAI API — GPT-4o alternative for reply generation

Three distinct workflows, all manual, all still active as of March 17, 2026:

Workflow 1: Etsy Live Daily Sales Scoreboard (~16 hrs/month)

Every day, Maya runs a ShopifyQL query for Etsy channel sales, reads the numbers from the Shopify UI, types them into Google Sheets, then uses Calculator for margin math. The exact loop from March 17, captured in real time:

09:48:22  [11s]  SHOPIFY analytics → Etsy yesterday sales
09:48:33  [ 4s]  SHEETS (types values in)
09:49:01  [ 9s]  SHOPIFY analytics (re-check)
09:49:10  [38s]  SHEETS (entering hourly breakdown)
09:51:02  [14s]  CALC  ← manual arithmetic on numbers she just copied
09:51:31  [ 5s]  SHOPIFY analytics (re-query)
09:51:36  [22s]  SHEETSCopy

This exact pattern repeats every working day for 3 months. 1,412 Calculator invocations in 30 days alone.

Workflow 2: Multi-Platform Sales Reporting (~5 hrs/month)

For Monday/Wednesday/Friday reports sent to leadership, she pulls sales data across 3 Shopify storefronts (meridian-us, meridian-ca, meridian-wholesale) and enters it into reporting sheets. ShopifyQL query captured from her browser URL parameters verbatim.

Workflow 3: Social Media Metrics (~17 hrs/month)

Daily visits to Instagram (21.4h total), TikTok (18.2h), Pinterest (14.6h), Facebook (9.1h) — manually reading follower counts, saves, and engagement numbers, then typing them into Google Sheets. This is also automatable via platform APIs.

Tier 1 Manual Fix — Today

Replace all Calculator usage with Google Sheets formulas. Margin %: =gross_profit/net_sales. MTD pacing: =total_to_date/days_elapsed*days_in_month. Running platform totals: =SUM(range). If the sheet does the math, Calculator goes to zero. Saves ~6 hrs/month immediately.

Tier 2 Dashboard Tool — Prototype Built

meridian-ops-dashboard.html (79KB) — a 5-tab application covering Maya's workflows: Etsy scoreboard, multi-platform sales reporting, social media tracking, product management, and analytics. Still requires manual data entry but organizes the process and eliminates all arithmetic.

Tier 3 Full Automation — Scheduled Script

A Google Apps Script deployed as a time-triggered function inside the Google Sheets environment:

• Runs daily at 08:45 UTC — before Maya's shift starts
• Calls Shopify Admin API across all 3 storefronts using the exact ShopifyQL queries we captured from her browser
• Writes results directly into her existing spreadsheets (we have the sheet ID from the data)
• Formulas auto-calculate margins, deltas, and pacing — Calculator goes to zero permanently
• Cost: $0 (Apps Script is free within Google Workspace limits)

Maya goes from 38 hours/month of manual data transcription to 0. Her time redirects to interpretation, anomaly detection, and content pipeline support — the judgment work that can't be automated.

API Documentation & Implementation References

• Shopify Admin REST API — products, orders, inventory endpoints
• Shopify Admin GraphQL / ShopifyQL Reference — the exact query language Maya's browser queries use
• Google Apps Script Time-Based Triggers — automated scheduled execution
• Apps Script UrlFetchApp — Shopify API calls from within Sheets
• Captured queries from Maya's browser: WHERE sales_channel = 'Etsy' DURING yesterday, multi-store sales by channel, discount performance by code

Andre runs a three-application manual pricing pipeline every day: Chrome (eBay/Etsy listing data) → Windows Calculator (margin math) → Notepad (clipboard buffer) → back to Chrome. Verbatim from his keystroke sequences:

89.00 / 0.44   → listing price for eBay standard (44% net margin after 13.25%+2.9%+$0.30)
89.00 / 7      → per-unit shipping component estimate
312.50 / 28    → monthly revenue / days = daily run rate check
24 / 0.51      → reverse margin at 51% net for lower-fee eBay category
147.80 / 18    → price per unit from wholesale lot purchaseCopy

He uses Notepad as a human clipboard because Windows Calculator loses intermediate values on app switch. This three-app pipeline ran on 19 of 19 tracked working days without exception. 97 confirmed calculations in keystroke data (estimated 12–22/day based on Calculator window events).

Every single observed calculation is a division or subtraction that could be a spreadsheet formula. 100% replaceable. The formula is: listing_price = item_cost / target_margin_decimal.

Tier 1 Already Built — Pricing Calculator

andre-pricing-calculator.html (26KB). Four platform tabs (eBay, Etsy, Amazon, DTC) — enter cost, get instant listing price with full fee breakdown for each marketplace. Visual margin bar. Adjustable target margins. Works offline in any browser.

This replaces Calculator + Notepad immediately. Andre can bookmark it today. No API keys, no setup.

Tier 2 Also Built — Full Pricing Engine

meridian-pricing-engine.html (54KB). Full product catalog with multi-marketplace comparison, bulk pricing mode for new catalog listings, and sale price modeling built from Andre's actual fee structure data. Covers the entire product catalog workflow, not just quick checks.

Tier 3 Full Pipeline — Zero Daily Touch

Scheduled script pulls product costs from Shopify Admin API, applies the marketplace fee structures we decoded from Andre's keystrokes, and outputs a pricing sheet with recommended listing prices for all 4 platforms. Can optionally push prices directly to marketplace APIs. Andre spot-checks weekly instead of calculating daily.

• Fee structures we decoded: eBay 13.25% FVF + 2.9% + $0.30 | Etsy 6.5% transaction + $0.20 listing + payment processing | Amazon 15% referral + FBA | Shopify DTC 2.9% + $0.30
• Margin targets by platform: 0.44 (eBay standard), 0.51 (eBay lower-fee), 0.62 (Etsy)

We can build this and have it back this week.

API Documentation & Implementation References

• Shopify Products API — pull product costs and variants
• eBay Inventory API — list/update eBay listings with pricing
• Etsy Open API v3 — listings, pricing, inventory management
• Amazon SP-API Pricing — competitive pricing and fee estimates

James processes upsell outreach lists by manually copying customer data from a Google Sheet (CS Master) into Recurly, one row at a time. The Feb 12 session shows a 19-minute window with 12+ consecutive copy-paste pairs in a single minute-bin at 14:00 — copying customer name, email, plan tier, then entering promo code HOME2026, and repeating.

What We Built

james-upsell-batch-processor.html (39KB). CSV import of customer list → template builder with variable chips (first_name, email, plan_tier, renewal_date) → generates all messages at once → one-click copy per customer. Eliminates the Alt+Tab loop between Sheets and Recurly.

Full Pipeline

Recurly has a bulk subscription API. A scheduled script reads the CS Master Sheet, applies the upsell template (with variable substitution), and pushes messages directly through Recurly's API. Full batch runs in 2 minutes of oversight instead of 15–35 minutes of manual work. Implementation: days, not weeks.

API Documentation & Implementation References

• Recurly API Docs — subscriptions, accounts, bulk operations
• Google Sheets API — read CS Master sheet data programmatically
• Zendesk Tickets API — alternative for sending outreach messages via ticket creation

Rosa logs cancellation reasons as free-text every time she processes a cancellation or win-back call. 74 logged reasons in 30 days — 3.5 per working day. This data is gold for understanding churn — but it's trapped in unstructured text in her personal workbook. Nobody else can see it. It doesn't aggregate. It disappears.

Our analysis extracted and categorized every cancellation reason from her keystroke data:

Percentages sum above 100% because Rosa often logs multiple reasons per case. These represent % of cases mentioning each reason.

Critical Insight: Payment Lapsed Is Recoverable Revenue, Not Real Churn

8–12% of the cases Rosa is treating as cancellations are payment failures — the customer never chose to cancel. Their card expired or their bank account changed. These should route to a payment recovery flow (Recurly has this built in), not a win-back campaign. Rosa is currently spending the same win-back effort on them as on customers who consciously decided to leave. That's wasted capacity and misread churn numbers.

What We Built

rosa-cancellation-tracker.html (23KB). Structured logging with dropdown taxonomy (11 canonical reason codes derived from Rosa's data), daily stats, win-back rate tracking, PAYMENT_LAPSED vs real churn split, CSV export for BI. Replaces free-text with structured data that flows into a dashboard. Implementation: 2 hours.

The daily activity log system exists but has degraded into compliance theater:

• James's shift summaries are copy-pasted identically every single day: "Handle incoming tickets and follow up on escalations." Keystroke data caught the literal Ctrl+A, Ctrl+C, Ctrl+V sequence — he's copying his previous summary and re-submitting it.
• Sofia's shift summaries are identical across 50+ consecutive days: "Process Hootsuite inbox and schedule content for all platforms."
• David has zero shift summary data for the entire 3-month period.
• The "Issues Encountered" field was filled in substantively only for the first 10 days (Jan 3–14), then went to "None" for the next 3 months without exception.
• The escalation flag was never triggered a single time across 412 entries over 3 months. Not once. Nobody on this team has asked for help through the system.

Whether that means nothing is wrong, or means nobody trusts the system, is worth understanding.

• James: Reported "3" (on a 1–5 scale) in 48 of 50 team logs. Two exceptions: "2" on back-to-back days with no corresponding note.
• Maya: Reported "3" in 52 of 53 team logs.
• Sofia: Consistently 2.5–3.0, never above 3.5 on any monthly average.
• Andre: Uniformly "2" across all 3 months, every entry.

These are not mood signals — they're default form values being submitted without engagement. The feeling tracker has no analytical value in its current state. If it's worth keeping, it needs a redesign that makes honest answers feel safe.

Andre's daily notes — which he writes genuinely, unlike the others — contain candid disclosures that never triggered an escalation:

• "Feeling run down this week, not sleeping well" (Jan 2026)
• "Power outages in the area slowing everything down" (Feb 2026)
• "Had a bad migraine all day, managed to push through" (Feb 2026)
• "Been sick since Monday but staying on top of listings" (Mar 2026)
• "Third day with a fever but not falling behind" (Mar 2026)

Session data on those days shows full 7–9 hour working sessions. He never takes sick days — he works through everything. The escalation flag stays off every time. This is someone who won't surface problems upward — and who runs a 3-channel e-commerce operation alone. That combination is a sustainability and retention risk.

Comparing January 2026 activity data vs March 2026: Maya went from weekly sales reporting and spreadsheet maintenance to also doing Canva graphic creation, 80+ Shopify product descriptions, sold-out item management, Klaviyo email segment updates, Pinterest board curation, product photography asset organization, and managing influencer tracking sheets. Her Shopify activity grew 2.2× (621 events in Jan → 1,886 in Mar).

All of this logs as "Operations Support" at the same task label, same rate. If you're getting graphic design, copywriting, influencer coordination, and CMS management under "Operations Support" — either the role needs to be re-scoped, or someone doesn't know how much function has drifted into her lane.

From a content review session captured on Feb 19:

"the spring linen series is outperforming everything... last month saves were 4.2K
now we're at 11.8K... I've been reposting our older 'room reveal' style content
with slightly different hooks and they're picking up again — same products, fresh
framing. Pinterest especially loves these."Copy

She maintains a content performance tracker, runs informal A/B tests on caption styles (question-hook vs statement-hook), analyzes platform data across channels, and proposes seasonal themes ahead of schedule. She creates graphics in Canva (18.2h total) and does occasional video editing in CapCut. Pinterest saves grew from 4.2K to 11.8K this quarter on the back of her seasonal content strategy.

This is not a "social media scheduler." This is someone analyzing performance data and running experiments. The comment inbox grind is taking 45–60% of her working time. That's the cost.

James uses ChatGPT on 16 of every 21 working days for two specific, task-relevant workflows — and he wasn't told to. He figured out on his own that AI could help with shipping status drafts and apology responses. He named the conversations ("Shipping Status Draft", "Apology Response") and returns to them repeatedly. This is someone who, given better tools, would use them immediately and well.

The problem isn't initiative — it's that both tasks he's using ChatGPT for are already solved within Zendesk, without leaving the app. He just doesn't know those features exist. That's a 2-hour fix, not a management problem.

Maya Santos and Andre Torres both carry the "Operations Support" label. They do fundamentally different work:

• Andre: Runs the multi-marketplace e-commerce operation. eBay, Etsy, Amazon, Shopify DTC listings. Adobe Photoshop for product images. Pricing. Listing copy. Keyword research. Inventory coordination across 3 channels.
• Maya: Runs internal reporting and media intelligence. Weekly summaries for leadership. Social media tracking across 4 platforms. Shopify data entry. Graphic creation in Canva. Newsletter segment updates in Klaviyo. Influencer tracking sheets.

These are not interchangeable. If Andre leaves, marketplace listings stall. If Maya leaves, leadership loses its data visibility layer. The shared label creates a false sense of redundancy where none exists.

David Chen (Customer Service, 3.5-year tenure, 432 hours logged) has: 312 total app events in 30 days (vs James's 14,218), 2 keystroke events, 4 idle records, and zero shift summary reports for the entire 3-month period. The Go2 tracking agent appears to be partially deployed on his machine.

We can confirm he's working — Zendesk, Shopify, Slack, and carrier tracking tools are visible in the one full day of URL data we captured. But we cannot assess his throughput, quality, or efficiency. We can't compare him to James. He's a black box.

Additionally: 14% of his work log entries (38 of 271) have been manually edited, mostly rounding to exact :00 minute boundaries. Not confirmed as an issue — could be habit, could be something worth looking at alongside the tracking gap.