The AI Automation Stack

A portable, self-hosted operational backend for growing businesses. The technical proof of our premium consulting capability — and the architecture we deploy when a Smart Website is no longer enough.

Most Growing Businesses Are Paying To Glue Tools Together

A typical small business ends up paying $500 to $2,000 a month for a stack of disconnected SaaS tools that don't talk to each other. The operational work between them — lead follow-up, reporting, intake, document generation, inbox triage — still gets done by hand. The alternative is not "go without automation." It's to run a small, focused, owned operational backend — built on the same patterns used by enterprise teams — that handles those workflows on infrastructure the business controls. The AI Automation Stack is the architecture we deploy when a Smart Website + AI Workers package is no longer enough.

Outcomes A Client Actually Buys

The architecture below exists to make these outcomes real. The technology is the mechanism; these are the results.

Less Manual Work

The intern stops re-typing form submissions into the CRM. The owner stops downloading invoices and forwarding them to the bookkeeper.

Smarter Reporting

Power BI dashboards refresh on schedule with alerts when they fail. KPIs surface in a daily briefing email instead of a spreadsheet nobody opens.

Email Cleanup

Incoming mail is classified, labeled, and summarized into a once-a-day digest. The inbox stops being a job. The system never deletes anything.

Workflow Automation

Lead intake, onboarding, invoice handling, approval routing — every cross-tool flow that used to be done by hand.

Document Generation

Proposals, SOWs, weekly client briefings drafted from templates and structured data, surfaced for human review and sign-off.

Operational Visibility

Knowing what the business is doing without opening five tabs. Dashboards built for executives, not analysts.

Client Intake

A public form routes new prospects into the CRM, the calendar, and a project tracker without anyone copying anything.

Owned Data

None of the above sends client data to a third-party AI vendor by default. The language model runs locally.

A Small Set Of Containers, Designed To Travel

Small enough to fit on a $200 device. Capable enough to handle the workflows of a 25-person business. Identical from one deployment target to the next.

📥

n8n

Workflow automation engine. Every cross-tool flow lives here — form intake, scheduled reports, API stitching, document generation.

🧠

Local Language Model

Classification, summarization, drafting — running on the client's own infrastructure. Data never leaves the device.

🔒

Reverse Proxy + TLS

Caddy handles HTTPS automatically. No expired certificates, no manual renewals.

🌐

Secure Public Access

Cloudflare Tunnel exposes selected services without opening firewall ports or requiring a static IP. Substitutable with Tailscale or a VPN.

📊

Container Management

Portainer for a web UI; the client's team can see what's running, restart services, and read logs without command-line work.

🔗

AI Integration Layer

A small protocol server lets AI assistants (Claude and others) call tools running on the stack — extending it into agentic territory.

Portable By Design

The compose file is the same on every deployment target. The deployment surface changes; the architecture does not. The client never has to commit to a single vendor, a single cloud provider, or a single hardware decision.

Mini PC

The sweet spot for a 5–25 person business with a back-office shelf. One-time hardware cost of $400–800.

Mac or Linux Workstation

An existing back-office machine with spare cycles. No new hardware purchase. Common in design and media businesses.

Windows PC with Docker

Common in offices that haven't standardized on Mac or Linux. Same container set, same deployment.

VPS

Hetzner, DigitalOcean, Vultr. $10–40/month. The right choice when there's no on-prem option.

Cloud VM

AWS, Azure, GCP. For clients with existing cloud accounts and procurement preferences.

Hybrid

Automation runtime on-prem, BI in the cloud, AI on a dedicated box. Mixed deployments are normal for businesses with multiple sites or compliance constraints.

Raspberry Pi (Proof-of-Concept)

The current prototype environment, ~$200 hardware. We run our own consultancy on this configuration to prove the stack is lightweight enough to run anywhere.

What This Replaces In SaaS Spend

For a business that wants the equivalent capability from SaaS, the apples-to-apples comparison looks roughly like this.

$100–300 Workflow automation (Zapier / Make pro tier)
$50–200 Hosted AI for classification (OpenAI API at modest volume)
$25 Managed reverse proxy + TLS
$25 Public tunnel (ngrok paid tier)
$20 Container management UI (hosted Portainer)
$220–570 Per month, replaced

That's $2,600 to $6,800 a year, before any of the data-ownership or vendor-lock concerns. The AI Automation Stack on a mini PC pays for itself within the first quarter. This is not an argument that SaaS is always wrong — it's the argument that for a 1- to 25-person business, owning the stack is often cheaper, more flexible, and produces better outcomes than gluing together half a dozen subscriptions.

What The Prototype Runs Today

What we run on our own stack — and what the equivalent client outcome looks like.

On Our Stack

Local LLM summarizes long emails into a daily brief.

Client Outcome

An operations lead stops spending the first hour of the day in their inbox.

On Our Stack

n8n watches Power BI dataset refreshes and alerts on failure.

Client Outcome

A BI manager finds out about broken pipelines before the CEO does.

On Our Stack

Cloudflare Tunnel exposes a public lead-intake form.

Client Outcome

A consulting firm receives qualified leads into their CRM without paying for a forms platform.

On Our Stack

n8n generates weekly client briefings from Power BI data via the local LLM.

Client Outcome

The account manager spends 15 minutes reviewing a draft instead of two hours writing one.

What This Looks Like For A Client

This is the premium tier of Digital Ops Systems work — sized for clients who have already worked with us at the Smart Website level, or whose operational needs go beyond a website from day one. The deployment target is the client's choice, or chosen together in the first session.

1

Week 1: Workshop

Identify the three to five workflows that, if automated, would return the most time per week. Decide on the deployment target.

2

Weeks 2–4: Build

Build the workflows in n8n on infrastructure deployed to the client's chosen environment. Each workflow ships with a one-page runbook.

3

Week 5: Handover

The client gets the compose files, the documentation, and a session on how to modify and extend the system.

4

Optional Retainer

Ongoing maintenance, new workflows, advisory. Engagement-priced by scope, not by hour.

What This Stack Is Not

Not An Enterprise Database Server

For workloads above modest throughput — more than a few hundred automation runs per hour, or multi-gigabyte AI inference — the same stack belongs on a mini PC, a VPS, or a real server. The portability is the point.

Local LLM Is Not A Frontier Model

The local model is good for classification, labeling, and short-form drafting. Where the work needs frontier-model quality, we architect the automation to call the cloud API while keeping the surrounding system local.

Self-Hosting Has An Operations Cost

Backups, updates, certificate rotation, monitoring — all real. We bake the runbooks into every engagement, but the client still has to follow them or pay someone to.

Not A Replacement For ERP, CRM, Or Accounting

This stack connects those systems, automates the work between them, and surfaces the data they hold. It is not a Salesforce competitor.

Ready To Talk About The Operational Systems Behind Your Business?

The Smart Website packages are the right starting point for most businesses. If you've outgrown packaged work and need the operational backend behind it, this is the conversation to have.