You Sold the Retainers. Now Deliver Them: A Local AI Client-Ops Machine

Month three of the productized-service dream. Twelve clients signed, everyone's paying, the blueprint worked. And now every Sunday night is the same: twelve dashboards, twelve reports, twelve slightly-different versions of the same analysis, written by hand at 11 p.m. The blueprint told you how to sell a $1,000/month service. Nobody drew the part where one person delivers it — week after week, client after client, without quality quietly rotting as the roster grows.

That missing drawing is client operations, and it is the actual bottleneck between "this works" and "this scales." Here is the good news for a local-AI operator: the entire delivery loop — data in, report out, approval, send, archive — is exactly the kind of pipeline Ollama and n8n eat for breakfast. The money math said 100 clients. This post is about the machine that makes 100 clients survivable for a team of one.

The Delivery Math Nobody Puts in the Blueprint

Before the machine, look at the arithmetic that kills productized services. Assume each client needs a weekly deliverable — a report, a digest, a content pack:

Nothing about the selling breaks at 30 clients. The breaking point is output production. A productized service stays productized only if every client genuinely gets the same quality — and no human hand-writes 100 personalized reports a week.

The Shape of the Machine

Five stations, one loop:

1. Client profile — one structured file per client: goals, tone, metrics sources, quirks.
2. Scheduled pull — the data that feeds the deliverable, fetched, not requested.
3. Local draft — Ollama writes version one from the profile plus fresh data.
4. Human gate — you review, edit, approve. Ten minutes, not forty-five.
5. Send + archive — delivery goes out, everything logged for next week's diff.

Station 1: The Client Profile

Every report a model writes is only as personal as the profile it writes from. One JSON file per client:

{
  "client": "bayside-dental",
  "service": "local-search-content retainer",
  "tone": "plain, warm, no jargon; owner reads on phone Sunday evening",
  "metrics": {
    "source": "sheet: bayside-metrics",
    "kpis": ["calls", "direction requests", "top keyword positions"]
  },
  "standing_notes": [
    "Owner cares most about calls from the airport area",
    "Never mention competitor across the street by name",
    "Front desk person named Nok reads reports first"
  ],
  "send": { "channel": "email", "day": "sunday", "time": "17:00" }
}

The standing_notes field does the heavy lifting. That is the stuff a good account manager carries in their head at five clients and loses at fifty. Written down, it makes every future report smarter — including the ones a model writes.

Station 2 and 3: The Pull and the Draft

The weekly loop in n8n, with the drafting prompt pointed at your local model:

workflow: weekly-client-reports
schedule: "0 8 * * 6"
nodes:
  - name: load-profiles
    type: read-folder
    path: "/peak/clients/*/profile.json"
  - name: pull-metrics
    type: http-request
    per_client: true
    url: "{{profile.metrics.source}}"
  - name: draft-report
    type: http-request
    method: POST
    url: "http://localhost:11434/api/generate"
    body:
      model: "llama3.2"
      stream: false
      prompt: |
        You are writing the weekly report for {{profile.client}}.
        Tone: {{profile.tone}}. Standing notes: {{profile.standing_notes}}.
        This week's numbers: {{metrics}}.
        Write: what moved, why it likely moved, one action for next week.
        No filler, no apologies, no jargon.
  - name: queue-review
    type: telegram
    message: "Draft ready for {{profile.client}} — reply APPROVE or EDIT."
  - name: send-if-approved
    type: email
    condition: "approved"
    to: "{{profile.send.channel_address}}"

The review message matters more than it looks. It arrives on your phone with the draft; you read it on Saturday morning coffee, reply APPROVE, and move on. You just did 45 minutes of account-management work in four — and you caught the week the model hallucinated a trend, which is the whole reason the gate exists.

Station 4: What Stays Human (On Purpose)

The pattern: the machine produces candidates, the human makes them commitments. Every automated delivery still has your name on it — the gate is where you put it there honestly.

Onboarding Is Data Capture

New client should mean intake, not improvisation. A form (n8n Form trigger) writes the profile skeleton, and the first report the machine drafts doubles as the onboarding test:

Intake form -> profile.json (draft) -> first automated report
-> you edit it heavily -> diff the model's draft vs your edit
-> corrections go back into standing_notes

That diff is quiet magic: every correction you make in the first month becomes standing notes, so the model stops making that mistake permanently. Onboarding effort compounds into delivery quality instead of evaporating.

The Churn Watch

The same loop that delivers can listen. Signals worth flagging automatically:

• Usage or engagement down two weeks running
• Report unopened three weeks straight
• Reply tone turning short (the model can classify sentiment locally)

A flag is not a churn verdict — it is a reason to call a client before they compose the cancellation email. Retainer businesses rarely die of bad delivery; they die of quiet drift nobody noticed. A $0 local model watching the signals 52 weeks a year notices.

The flags themselves are cheap to build: a weekly comparison node in the same n8n workflow, a classification prompt against last month's replies, a threshold ("open rate below 20% for three weeks"). No dashboards, no new vendor — the data is already flowing through the loop that writes the reports. Churn watching is delivery infrastructure wearing a different hat.

Scale Checkpoints

• 10 clients: run it manually once a month to feel the loop; build the profile habit now.
• 30 clients: the machine drafts everything; you gate everything. Sundays are two hours.
• 60 clients: gate by exception — auto-send categories where the model is consistently boring-and-correct; human-review the rest.
• 100 clients: the blueprint's math arrives — and survives — because delivery stopped scaling with your calendar weeks ago.

Make It Practical This Month

• Write profiles for your existing clients tonight — the standing notes alone sharpen your own work.
• Automate ONE client's weekly report end to end, including the approval message.
• Keep an edits log: your corrections are the training data for next month's prompts.
• Add the churn flags when client three goes quiet and you find out from their cancel email instead.

The blueprint sold the hundred-client dream. The ops machine is what keeps it from becoming a hundred-client panic — same software as every other workflow in this stack, pointed at the least glamorous and most churn-preventing process you own.

Frequently Asked Questions

Isn't a model-written report impersonal? Only if the profile is thin. The report's personalization lives in the standing notes, the tone definition, and the metrics you chose to track — all human inputs the model assembles weekly. A well-profiled model draft edited by you beats a hand-written report you rushed at 11 p.m., every time.

What if the model gets a number wrong? That is what the approval gate is for. The machine pulls metrics (deterministic) and drafts prose (probabilistic); the gate is where the two meet a human before anything ships. As a backstop, assert in the prompt that numbers may only come from the provided metrics — then spot-check one figure per report until trust is earned.

How long before this pays back the setup time? The single-client version is an afternoon: one profile, one cron, one prompt, one approval message. At 45 minutes saved per client per week, a five-client roster pays it back inside two weeks. The machine's real value compounds later — it is the difference between capping at the clients you can hand-write for and not capping.

Does this replace the account manager? It replaces the account manager's production work: gathering, drafting, assembling, sending. It does not replace judgment calls, relationships, or the strategic conversation where retainers get renewed. If anything, it protects those — by clearing the Sundays that were eating them.