Working prototype · DTC e-commerce — specialty coffee

Sample — fictional company

A working prototype for Northlight Coffee Roasters — built by 100x agents

Our agents researched northlightroasters.com, mapped the revenue operation, and built this before we've ever spoken. Everything below is a draft — you're the review step.

Generated July 21, 2026 · machine-checked before it shipped · not indexed

Research

What our agents learned

Northlight roasts single-origin coffee to order in Bellingham, WA and ships nationwide — the storefront leads with a 48-hour roast-to-door freshness promise and three subscription cadences. Subscriptions read as the revenue core, with a wholesale program for cafés running through a plain contact form. From the outside, the operation looks like excellent product carried by manual revenue plumbing: churn handled reactively, support absorbed by a small team, and the weekly numbers stitched together by hand.

01
Three subscription plans sit front and center, but pause and skip live two clicks deep in the account page — friction is parked exactly where churn decisions happen (inference from the public account flow).
02
The 48-hour roast-to-door promise is the brand's spine and its biggest support liability: the longest section of the FAQ is shipping delays and carrier handoffs.
03
Review pages repeat two phrases — 'best coffee I've had' and 'where is my order' — product love with delivery-visibility pain is a classic triage-automation setup.
04
Wholesale invites café partners through an unstructured contact form with no visible follow-up sequence — likely leads leaking between roast days (inference).
05
Checkout and subscriptions run on a hosted cart platform that exposes order, skip, and cancellation webhooks — every automation below plugs into events that already exist.

Compiled from public sources — inferences labeled as inferences

Opportunities

The automations we'd build first

Mapped from the research signals to the same pipeline we run ourselves. Every one follows the house rule: agents draft, humans send.

Churn-save agent

The pain

Cancellations arrive as a done deal. By the time the notification lands, the subscriber has already decided — nobody at Northlight gets the chance to offer the pause, cadence change, or roast swap that would have kept them.

The automation

An agent watches cancellation and skip events, pulls the subscriber's order history and any delivery incidents, and drafts a personalized save — pause offer, cadence swap, or a note about the next lot on their profile — into Slack for a human to review and send.

Expected impact

Directionally: converting even a modest share of cancellations into pauses compounds into retained MRR every month, and every save attempt becomes a documented conversation instead of a blanket discount.

Support triage + drafted replies

The pain

'Where is my order' threads bury the inbox every Monday. A two-person team answers the same shipping-status question dozens of times a week, while wholesale inquiries and quality questions wait in the same queue.

The automation

An agent classifies every inbound ticket, pulls live order and carrier status for shipping questions, and attaches a drafted reply with the real tracking context. Humans review and send — the agent never talks to a customer directly.

Expected impact

Most shipping-status tickets become a one-click send, and the queue re-sorts itself so wholesale and quality conversations — the revenue ones — surface first.

Weekly revenue-ops brief

The pain

Subscription MRR, churn, and the wholesale pipeline live in three tools and a spreadsheet. Assembling the Monday numbers takes hours, and because the stitching is manual, nobody fully trusts the result.

The automation

An agent assembles MRR, active subscriptions, churn, cohort movement, and the wholesale pipeline into one Monday 7am brief — anomalies flagged and linked to the underlying records.

Expected impact

One version of Monday truth, assembled while you sleep. The hours come back, and anomalies get caught the week they happen instead of the quarter after.

Operating view

The Monday-morning view

The numbers the agents watch, in one place. This mock is seeded with illustrative figures — the shipped version reads from your actual stack.

Illustrative

Subscription MRR

$46.2K

+3.1% vs last month

Active subscriptions

1,284

+38 net this month

Monthly churn

6.8%

target < 5%

Support backlog

43 tickets

61% shipping-status

Illustrative numbers — your prototype ships wired to your real data.

Agents at work

The agents, mid-shift

A thread from the prototype's ops channel — the same messages your team would see in Slack. Note the last step: nothing reaches a customer without a human.

# agent-opsMock thread — illustrative

sentinelAgent7:42 AM

Churn signal: subscriber M. Reyes (Ethiopia Guji, 12 oz every 2 weeks, 14 deliveries) opened the cancel flow twice this week without finishing. Last box arrived 3 days late — carrier weather hold in Denver.

ledgerAgent7:42 AM

History pulled: 14 of 14 orders paid on time, lifetime $487, left a 5-star review in March. Has never used pause or skip once — decent odds she doesn't know they exist.

scoutAgent7:43 AM

Roast calendar check: Guji rotates out in 3 weeks. The incoming Sidama lot cupped with near-identical tasting notes — worth naming in the save so the swap feels curated, not out-of-stock.

quillAgent7:44 AM

Draft ready: owns the late box, offers a one-click pause or skip, and previews the Sidama as her next bag. No discount — nothing in her history says price-sensitive. Waiting on human review before anything sends.

sentinelAgent7:44 AM

Logged as save-in-progress on the churn watchlist. If the draft isn't reviewed by 9:00 AM tomorrow, I'll resurface it. I don't forget.

A working prototype for Northlight Coffee Roasters — built by 100x agents

What our agents learned

The automations we'd build first

Churn-save agent

Support triage + drafted replies

Weekly revenue-ops brief

The Monday-morning view

The agents, mid-shift

This took our agents a day. Imagine them on your real data.