Merchandising Visitation (Rep & Region) Notifications

The Problem Today

Every Redline Regional Manager (RM) writes the same kind of message to their regional Slack channel 2-3 times a week. It is a coverage update for the Merchandising Specialists (MS) on their team: a greeting, a weather note, "The Numbers," today's hitlist, and a standing reminder to mark check-in status in Mobiwork. The real example from the source doc looks like this:

None of this is hard. It is just tedious, and it is the same shape every time. The RM has to log into the billing view, copy last week's numbers, look up the day's routes, glance at the weather, and then write it all up in their own voice. A "good 15-30 minutes" per message, per the RM's own estimate.

What it actually costs

This is a small per-instance cost that adds up because it repeats forever, across every region, multiple times a week. The math at the low and high ends:

For a team of, say, 8 RMs that is on the order of 600 RM-hours a year spent retyping numbers and weather. The exact figure scales linearly with headcount — give us the real RM count and we will plug it in — but the point holds at any N: this is recurring manager time spent on assembly, not judgment. The judgment part (does the message read right, anything to flag the team) is a 30-second review, not a 25-minute write-up.

How It Works

A scheduled job assembles a draft for each region on the RM's normal cadence and posts it to a private review thread (or DMs the RM). The RM reads it, edits if needed, and clicks Approve & Post. The bot then publishes the finished message to the team channel as the RM. Human stays in the loop by default; auto-post is an opt-in per region once an RM trusts it.

The four data pulls

Nothing here is invented by the model. The numbers are fetched from systems of record; the LLM only does the writing.

• Numbers — last week's billing total, VINs serviced, and weekly goal attainment pulled from Predian for that region.
• Hitlist — today's stores / routes for the region, pulled from the route source (Mobiwork today; route engine later — see MS Routes).
• Weather — today's forecast for the region's metro via a weather API, used to add conditional guidance.
• Standing content — the Mobiwork status reminder and @help tag, which are constant and come from the template, not the model.

The LLM's job — and only its job

An LLM (Bedrock) takes the fetched facts plus a per-region template and writes the message in that RM's voice. It does not source numbers — those are injected verbatim. The model handles three things humans find tedious but it does well:

• Tone matching. We seed each RM's template with 5-10 of their past messages so the draft sounds like them, not like a robot. Warm-and-chatty stays warm-and-chatty; terse-and-bulleted stays terse.
• Weather-conditional language. Rain in the forecast → "plan covered areas where you can." Heat advisory → "hydrate, take your breaks." Clear → a short upbeat line. The rule set is small and reviewable.
• Goal framing. Phrasing attainment honestly ("$3k of the $20k weekly goal — let's push") without the model ever doing the arithmetic itself; we compute the percentage and hand it over.

Data flow

EventBridge schedule (per region, RM's cadence)
        │
        ▼
   draft-builder Lambda
        ├── Predian API ........ last-week $, VINs serviced, goal %
        ├── Route source ....... today's hitlist for region
        ├── Weather API ........ today's forecast → condition flags
        └── Template store ..... RM voice profile + standing blocks
        │
        ▼
   Bedrock (LLM) — compose message in RM voice from facts + template
        │
        ▼
   Slack: post DRAFT to RM review thread
        │  with buttons: [ Approve & Post ]  [ Edit ]  [ Skip today ]
        ▼
   RM clicks Approve  ──►  Slack interactive callback Lambda
        │
        ▼
   Publish final message to #region-team channel (as / for the RM)
   log: drafted_at, approved_by, edits_made, posted_at

The "Edit" button opens a Slack modal pre-filled with the draft text so the RM can tweak inline and post without leaving Slack. "Skip today" suppresses the post (slow day, off-cadence) and is logged so we learn the real cadence per region rather than guessing it.

Architecture & Stack

Deliberately boring and cheap. A scheduled Lambda that fans out per region, a second Lambda for the Slack interaction callback, and a small template store. No always-on compute, no database to speak of beyond a log and the per-region template config.

Build Plan

This is a fast build because each piece is small and the failure mode is benign (a bad draft is caught by the RM, not posted blindly). Most of the calendar time is wiring credentials and getting the voice right, not code.

Data & Access Needed

Everything below already exists at Redline — this build connects to it, it does not create new systems of record. The gating items are read access to Predian and a clean way to read today's routes.

Risks & Open Questions

Cost to Own

This is the rare build where run cost is genuinely negligible and the maintenance story is mostly "leave it alone."

The real maintenance

• Voice drift. When a new RM joins or an existing one changes style, someone updates a template row with fresh samples. Minutes, no deploy.
• Source contract changes. If Predian's API or the route export changes shape, the corresponding fetch needs an update. This is the most likely break and is the same dependency every Redline automation shares.
• Weather phrasing rules. A short, owned list. Add a case (e.g. snow) when someone asks; otherwise static.