0 readers reading
How to Build a One-Person AI Back Office for Your Small Business in 2026

How to Build a One-Person AI Back Office for Your Small Business in 2026

Build a one-person AI back office for your small business using Claude, a Postgres database, MCP connectors, and scheduled routines. Learn the stack, costs, and first five automations.

Sham

Sham

AI Engineer & Founder, The Tech Archive

11 min read
0 views

Verdict: In 2026, a solo founder or small team can run most back-office work — inbox triage, content planning, finance tracking, and daily reporting — through a single AI dashboard that uses Claude as the reasoning layer, a PostgreSQL database as the "brain," and Model Context Protocol (MCP) connectors to talk to your existing tools. The setup is not plug-and-play, but it is within reach for anyone comfortable with Claude Code and a $25/month database. The real payoff is not replacing yourself; it is turning two hours of daily admin into a 10-minute review with approve-or-reject buttons.

Last verified: 2026-06-18 · Best for: Solo founders, creators, agencies · Core stack: Claude + Postgres + n8n/scheduled routines · Starting cost: ~$45–$80/month · Build time: 1–3 weekends

What this article covers

  • What a one-person AI back office actually is (and what it is not)
  • The minimum viable stack and real 2026 costs
  • A step-by-step build path you can follow this weekend
  • The five automations that unlock the most time savings
  • Governance rules that keep the system safe
  • FAQ and sources

What is a one-person AI back office?

A one-person AI back office is a private dashboard plus a set of scheduled Claude routines that aggregate the information you currently check across five to ten different apps: email, Slack, bank account, Stripe, calendar, analytics, social accounts, and project tools. Instead of logging into each app every morning, you open one screen that shows:

  • Items that need your approval today
  • Tasks the system completed overnight
  • Alerts that need a human decision
  • A short written summary of cash, content, and communication status

The idea is not to let AI send things without permission. It is to let AI prepare everything so you can decide in minutes, not hours.

What it is not: a no-code template you install in five minutes and forget. Systems like this require ongoing tuning because every business has different data, payment flows, and approval rules. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, mostly because of unclear ROI and missing governance. The goal of this guide is to help you avoid that failure mode by starting small, measuring time saved, and keeping a human in the loop.

Why this works now

Three things changed in the last 18 months:

  1. Claude can reason across long context. Sonnet 4.6 and Opus 4.8 support a 1 million-token context window and can hold enough business instructions, memory, and conversation history to act like a coordinator rather than a chatbot (Anthropic pricing docs).
  2. MCP makes integration cheap. The Model Context Protocol, open-sourced by Anthropic in November 2024, gives Claude a standardized way to read from your database, files, Slack, GitHub, Google Workspace, and more without writing a custom connector for every tool.
  3. Managed Postgres is cheap. Supabase Pro starts at $25/month and gives you an 8 GB PostgreSQL database, daily backups, and 100,000 monthly active users — enough to serve as the central brain of a one-person operation (Supabase pricing).

The result: you no longer need a $10,000 enterprise automation suite to connect your tools. You need one model, one database, one scheduler, and clear rules.

The minimum viable stack

Layer Tool Cost (2026) What it does
Reasoning Claude Pro or API $20–$100/month Reads data, drafts responses, decides what to flag
Database Supabase Pro $25/month Stores memory, tasks, finance, content, analytics
Automation engine n8n self-hosted or Cloud Pro $0–$50/month Runs scheduled routines, webhooks, and handoffs
Code/IDE Claude Code Included in Pro/Max Builds skills, routines, and the dashboard
Connectors MCP servers Mostly free Talk to Slack, Gmail, Notion, bank APIs, Stripe
Dashboard Custom Next.js/Vue app Hosting only (~$5–$20) Morning review, approve/reject, voice reply

Total realistic starting cost: $45–$80/month if you self-host n8n on a small VPS, or $70–$120/month if you use n8n Cloud Pro.

For comparison, a single virtual assistant or junior ops hire typically costs $500–$2,000/month. The AI back office pays for itself if it saves even one day per week of founder time.

A 7-step build path

1. Write down the jobs you repeat every week

Before touching code, list the tasks that eat your time. Common examples:

  • Sorting inbound email and DMs
  • Updating a cash or P&L spreadsheet
  • Planning content for the week
  • Following up on unpaid invoices
  • Checking ad spend and website traffic
  • Writing meeting notes or client updates

Pick the top three. Everything else is Phase 2.

2. Create a central database in Supabase

Start with a single Supabase project and a few tables:

  • inbox_items — raw messages, status, suggested reply, approved reply
  • tasks — to-dos created by routines or by you
  • transactions — income, expenses, tax reserve, bank account balance
  • content_plan — planned posts, scripts, thumbnails, publish dates
  • memory — facts the system learns about your preferences
  • routine_log — what each automation did and when

Keep it simple. You can add tables as routines grow.

3. Connect Claude to your database with an MCP server

Use the official Postgres MCP server or build a thin read-only MCP wrapper around your Supabase API. This lets Claude answer questions like:

  • "What was my top expense category last month?"
  • "Which client owes the most?"
  • "What content is scheduled for next week?"

For our other take on building a personal command center around a knowledge base, see how to build an AI agent operating system with Hermes Agent and Obsidian.

4. Build "skills" for each repeated job

A skill is two files: a markdown spec that tells Claude what the job is, and a small script or function that does it. For example:

  • invoice-skill.md — how to read unpaid invoices and generate reminders
  • invoice.py — fetches data from the database and writes PDF reminders
  • inbox-skill.md — how to classify, summarize, and draft replies to common messages

Store skills in a folder that Claude Code can read. Then turn them into scheduled routines.

5. Schedule routines, not chat sessions

The biggest mistake is trying to run your business by prompting Claude all day. Instead, create routines that run automatically:

  • Morning review routine (6 AM) — reads inbox, finance, analytics, and task tables; writes a summary to the dashboard
  • Inbox reply routine (every 15 minutes) — classifies new messages, drafts replies, marks them as "pending approval"
  • Invoice routine (last day of the month) — checks unpaid invoices, generates reminders, queues them for your approval
  • Content routine (Sunday evening) — reads analytics and memory, proposes next week's content plan
  • Finance routine (daily) — pulls bank and Stripe data, updates the transactions table

Use n8n self-hosted for the scheduler, or Supabase Edge Functions if you prefer to stay inside one platform.

6. Build a small dashboard for approve/reject decisions

The dashboard does not need to be fancy. It needs three things:

  • A list of pending approvals with one-tap approve/reject
  • A "voice reply" button that transcribes your spoken instruction into a drafted response
  • A live activity feed so you can see what the system is doing

This is where the "clicking beats chatting" principle matters. A founder who can process 20 decisions with taps will use the system every day. A founder who has to write prompts will not.

7. Add memory and feedback loops

Every time you approve, reject, or edit a system suggestion, log that feedback. Over time, the system learns:

  • The tone you prefer in client replies
  • The price floor below which you auto-approve expenses
  • The sponsors you always accept or decline
  • The content angles that performed well

Store these observations in the memory table and include relevant entries in every routine's prompt context.

The five highest-ROI automations to start with

Automation Time saved Key tool
Morning inbox + reply drafting 45–90 min/day Claude + Gmail MCP + dashboard
Daily cash position + P&L summary 20–30 min/day Supabase + bank/Stripe API
Weekly content plan from analytics 1–2 hrs/week Claude + memory + content table
Monthly invoice reminders 2–3 hrs/month Invoice skill + scheduled routine
Sponsor/lead negotiation replies 30–60 min/email Claude + memory of your rates

Start with inbox and finance. They are high-volume, repetitive, and relatively safe because every outgoing message still needs your approval.

Governance: keep a human in the loop

Autonomous back-office systems fail when they are given too much authority too soon. Set hard rules:

  • No spending or sending money without explicit approval.
  • No client-facing message goes out unreviewed.
  • Every routine logs what it did to routine_log.
  • Every model action has a rollback path.
  • Review and update memory monthly; delete stale facts.

For a deeper look at why agent projects fail and how to avoid it, read our guide on how to build Claude employees instead of chatbots.

What this means for you

If you run a small business, agency, or creator operation, a one-person AI back office is probably the highest-leverage project you can build in 2026. It will not replace your judgment, but it will remove the friction that keeps you from exercising it. Start with three repeated tasks, one database, and one daily review screen. Measure hours saved per week. Expand only after the first three routines are reliable.

FAQ

Q: Is this only for technical founders? A: The first version requires comfort with Claude Code, SQL, and a small web app. Non-technical owners can use Claude for Small Business, a no-code Anthropic offering that connects QuickBooks, PayPal, HubSpot, Google Workspace, and Microsoft 365 (Anthropic announcement).

Q: How much does Claude cost for this kind of workload? A: For interactive building and review, Claude Pro is $20/month and Max starts at $100/month. If you run heavy automation through the API, current rates are Sonnet 4.6 at $3 input / $15 output per million tokens, Opus 4.8 at $5 / $25, and Haiku 4.5 at $1 / $5 (Anthropic API pricing). Prompt caching can cut repeat-context costs by ~90%.

Q: Can I use n8n instead of Claude Code to build this? A: n8n is excellent for the scheduling and connection layer. Claude Code is better for the reasoning, memory, and skill-authoring layer. Most builders pair them: n8n for orchestration, Claude for decisions and drafts.

Q: What is MCP and why do I need it? A: MCP (Model Context Protocol) is an open standard that lets an AI model read from databases, files, and APIs through a common interface. It means you can connect Claude to your tools without writing a custom integration for every service (Anthropic MCP docs).

Q: What are the biggest risks? A: Data access overreach, stale memory, hallucinated replies, and runaway API spend. Mitigate them with read-only defaults, approval gates, monthly memory audits, and spending caps in your Anthropic console.

Q: How long until this actually saves time? A: Most builders see a net time loss in week one while setting up, break even in weeks two to four, and save 5–15 hours per week by month two once three core routines are running reliably.

Sources
Updates & Corrections
  • 2026-06-18 — Article published. Pricing verified against Anthropic, Supabase, n8n, Zapier, and Make.

Get the practical AI brief

Verified, no-hype AI tips you can actually use - in your inbox. Free.

No spam. We verify what we send. Unsubscribe anytime.

Discussion

0 comments