AI automation services an agency can sell, built on open source

What does an AI automation agency actually sell?

Clients do not buy “automation”. They buy fewer missed leads, faster invoice processing or a support queue that stops growing. The agency’s job is to turn that outcome into a fixed package with a clear price, a delivery date and a monthly maintenance fee.

Open source matters here for two reasons. It keeps your delivery cost low across many clients, and it lets you offer self-hosting to clients who cannot send data to third-party SaaS tools.

Which automation services can an agency package?

These offers repeat across industries, so you can build a template once and adapt it per client.

Which open-source stack should an agency standardize on?

Pick one tool per layer and reuse it everywhere. Every extra tool you support multiplies maintenance across every client.

• Workflow engine: n8n is the common choice; read its sustainable-use licence, which limits some hosted resale models. Windmill and Node-RED are alternatives.
• LLM access: a hosted API for quality, Ollama for private or cheap tasks, behind one adapter.
• Documents: Paperless-ngx for archiving, Docling or Unstructured for parsing, Tesseract for OCR.
• Data: PostgreSQL with pgvector, or Baserow and NocoDB when clients want a spreadsheet-like view.
• Monitoring: Langfuse for LLM traces, Uptime Kuma for simple health checks.

Who pays, and how should you price it?

The best buyers are owners of 10 to 200 person companies who feel the pain personally and can sign without procurement. Operations managers are good second contacts because they own the broken process.

Price the setup as a fixed project and the running system as a retainer that covers hosting, monitoring, prompt updates and small changes. Avoid pure hourly billing, which punishes you for building reusable templates.

What is the hard part of running an AI automation agency?

Delivery is easy compared with maintenance. Client APIs change, a form gets a new field, a model update changes output style, and your automation breaks silently at 2 a.m.

Build for failure from day one: log every run, alert on errors, and add a human review step wherever an AI decision reaches a customer. The agencies that survive treat automations like software, with versioning and tests, not like one-off scripts.

• Scope creep: every client wants “just one more” branch in the workflow.
• Data access: getting credentials and permissions often takes longer than building.
• Trust: clients need to see what the AI did and why, in plain language.

How do you scope an MVP offer?

• Step 1: choose one service from the table and one industry you can reach.
• Step 2: build it once for yourself or a friendly client, with real data.
• Step 3: write a one-page scope: inputs, outputs, what is excluded, response times.
• Step 4: template the workflow so a second client takes a fraction of the time.
• Step 5: add a monthly report that shows runs, errors and time saved in the client’s own terms.

Three packages in more detail

Lead qualification and routing is the easiest first sale. A web form, email or chat message arrives, an LLM classifies intent, budget signals and urgency, and n8n routes it to the right person with a short summary. The MVP covers one inbound channel and one CRM. The hard part is agreeing on what a “qualified” lead means, so run a workshop with the sales team before writing any prompt.

Invoice capture is a steady retainer business. Paperless-ngx watches an inbox, Tesseract or Docling extracts text, an LLM maps fields and n8n posts a draft entry to the accounting tool for an accountant to approve. Accuracy on messy scans decides whether the client keeps paying, so keep a correction queue and review its contents every month.

An internal knowledge assistant looks impressive in demos but needs the most care. Dify or Flowise gives you a retrieval pipeline quickly, yet the real work is cleaning source documents, handling permissions so staff only see what they may see, and refreshing the index when files change. Scope the MVP to one department and one document set.

How much does it cost to run client automations?

Your costs are hosting, model usage and monitoring time. Self-hosted workflow engines and databases run comfortably on modest servers for small clients, while model usage scales with volume and varies by provider, so measure it per client in the first month instead of guessing.

Build model spend into the retainer with a fair-use cap, or pass it through at cost with a clear line on the invoice. Either way, the client should never be surprised by a bill driven by a loop that ran all night.

• Set hard limits on retries and batch sizes in every workflow.
• Separate each client into its own workspace or instance for clean billing and data isolation.
• Keep credentials in a secrets store, not inside workflow nodes.

Where it breaks: operating many clients at once

Ten clients with five workflows each means fifty things that can fail. Without central logging you learn about failures from angry emails. Route every workflow’s errors to one channel, tag them by client, and review the week’s failures in a fixed slot so fixes become templates rather than one-off patches.

How do you hand over or offboard a client?

Every engagement ends eventually, and a clean exit protects your reputation. Keep each client’s workflows, prompts and credentials in a separate, exportable workspace from the start, with a short runbook that explains what each automation does and who to call when it fails.

Open-source tooling helps here: the client can keep running the same n8n, Paperless-ngx or Chatwoot instance without buying new licences. Agree in the contract who owns the workflows and templates, and which parts remain your reusable intellectual property.

Which industries respond best to these offers?

Industries with high inbound volume and thin admin teams respond fastest: trades and home services, clinics and salons, logistics firms, property managers and small accounting practices. They feel every missed call and every late invoice, and the owner usually makes the buying decision alone.

Pick one industry after your first few projects and learn its tools, vocabulary and busy seasons. A package called “lead routing for roofing companies” sells more easily than “AI automation for businesses”, and your templates improve with every similar client.

Common mistakes agencies make

• Selling custom builds to every client instead of productized packages.
• Running every client on your personal accounts, which makes handover and billing a mess.
• Letting AI send messages to customers without a review mode during the first weeks.
• Ignoring data protection rules; clients in regulated sectors need processing agreements and clear hosting locations.
• Using a tool whose licence forbids the way you host it for clients.
• Promising accuracy numbers you have never measured on the client’s data.

How to choose your first service

Start with the offer where you can show a before-and-after in a single demo call: lead routing and meeting summaries are good openers because the result is visible immediately. Document-heavy offers pay better but need more testing.

RepoLoot’s catalog lists open-source automation and agent projects with notes on business value, which helps you pick a base you can maintain for years. Standardize, template, then specialize in one industry once you have three similar clients.

Frequently asked questions

Can I resell n8n workflows to clients?

You can build workflows for clients, but n8n uses a sustainable-use licence rather than a classic open-source one, and it limits certain commercial hosting models. Read the current licence text and, if you plan to host n8n as a service for many clients, check whether you need a commercial agreement.

Should client automations run on my server or theirs?

Both work. Hosting yourself is simpler to maintain and supports a retainer model. Hosting on the client’s infrastructure suits regulated sectors and makes handover clean. Many agencies offer both, with a higher retainer for managed hosting and a lower one for support only.

Do I need to be a developer to run an AI automation agency?

Low-code tools lower the bar, but someone on the team must be comfortable with APIs, webhooks, JSON, authentication and debugging. Most painful incidents come from edge cases that visual tools hide. At least one technical person is essential for reliable maintenance.

How do I prove the automation is worth the retainer?

Send a short monthly report: number of runs, errors caught, items reviewed by humans and the manual steps that no longer happen. Use the client’s own measures, such as response time or invoices processed, and avoid claiming savings you cannot trace back to logs.