The AI automation agency tech stack: what a small agency actually needs
Search for an AI automation agency tech stack and you get lists of thirty tools. A small agency needs about seven, one per layer, and the hard part is not picking them but keeping the stack small enough that your week goes to clients rather than plumbing. This guide goes layer by layer, says when to add a second tool, and is honest about which layers are worth building yourself.
Direct answer
What tech stack does an AI automation agency need?
Seven layers: a CRM for the pipeline, a workflow tool such as n8n, Make or Zapier for fixed integrations, an agent runtime with hosting for judgment work that runs around the clock, model access, a way to write proposals, invoicing, and a channel to reach clients. Keep one tool per layer until a client forces a second. Buy the layers that are not your product, such as hosting, isolation and invoicing, and build only the part clients pay you for: the instructions, skills and review process inside each agent.
The stack, layer by layer
| Layer | What it does for you | Start with | Add or upgrade when |
|---|---|---|---|
| CRM | Tracks prospects, pilots and renewals | A free CRM or a spreadsheet | You have more than a handful of open deals |
| Workflow tool | Moves data between apps along fixed steps | n8n, Make or Zapier, whichever you know | A client's volume makes usage billing hurt |
| Agent runtime and hosting | Runs judgment work around the clock | A managed platform, or one server you already run | Upkeep for several clients starts eating your week |
| Model access | The models your agents call | One key through one provider | A client wants usage on their own account |
| Proposals | Quotes the pilot and the retainer | The free proposal generator | You want quotes built from your real platform cost |
| Invoicing | Bills setup fees and retainers | The accounting tool you already use | Retainers repeat and need reminders |
| Client comms | Where clients hear from you and the agents | Email and one shared chat channel | Several clients each expect a weekly brief |
Build vs buy, by layer
| Layer | Buy when | Build when |
|---|---|---|
| Agent hosting and isolation | Hosting is not what clients pay you for | Host-level control or residency is part of the offer |
| Workflow integrations | A connector already exists | The client's system has an API nobody supports |
| Agent instructions and skills | Rarely: this is your product | Almost always: it is what makes your delivery different |
| Proposals and invoicing | Almost always | Only once your volume justifies it |
Start from what you deliver, not from a tool list
Every tool in your stack should trace back to something a client pays for. Write down the one or two jobs you sell, then ask what each layer must do for that job. A review reply service needs access to the client's review profiles, an agent that drafts on a schedule, a way to approve drafts, and a weekly brief. It does not need a vector database, three automation tools and a voice platform. Keep one tool per layer until a client forces a second, because every extra tool is another login, another bill, and another thing that can break overnight. The table below is the whole stack for a founder with a first few clients.
Workflow tools and agents do different jobs
Workflow tools such as n8n, Make and Zapier are the right choice when the steps are known in advance: a form comes in, a row is added, a message goes out. They are predictable and easy to reason about. Agents are for the work in between, where something has to read, judge and draft: which lead deserves a call, what a one star review needs in reply, why ad spend jumped overnight. Many agencies end up with both, and the split is simple: fixed paths in the workflow tool, judgment in the agent. Check how each tool bills and licenses before you build a client on it, because client volume and resale change the math.
- n8n's Sustainable Use License lets you build, run and charge for workflows for clients on your own instance, as long as the clients cannot create or edit them. Hosting n8n as a service for clients to build in, or white-labeling it, is not allowed.
- Make bills usage in credits, and each module action in a scenario counts as one credit, so a busy client scenario costs more each month.
- Zapier bills by tasks, counted each time it completes a unit of work, and offers a free plan to start on.
Sources: n8n License FAQ · Make pricing · Zapier pricing
Agent runtime and hosting: the layer that turns into a second job
An always-on agent needs a runtime such as OpenClaw, Hermes or Codex, and a machine that stays up. On your own servers that comes with homework for every client: keeping the machine running, storing secrets safely, applying runtime updates, reading logs when an agent goes quiet, capping spend, and answering what the agent touched. One client on a server you already run is manageable. Five clients is a part-time operations job nobody pays you for. A managed platform takes that layer off your plate. On Qoren each client runs in its own isolated environment, and keys you save to a client reach only that client's agents. Agents start from templates and run on schedules or triggers, and approval mode stops an agent before anything that changes state. The account runs inside one monthly budget, with alerts or an automatic pause.
Model access, and knowing what each client costs
Your agents call models, and model usage is the cost that moves with how busy a client is. Start with one route to the models rather than a key per provider. On Qoren you can use managed model access, paid from your credit balance, or bring your own key so the provider bills you directly. Qoren accepts keys for OpenRouter, Anthropic, OpenAI, Google Gemini and other providers, and OpenRouter itself puts hundreds of models behind a single API, which keeps it to one key. Whichever you choose, know what each client costs you before you renew them. In the console, Clients groups each client's environments, and cost per client shows what each one used over a period, split into hosting, models, web search and mailboxes. That figure is the cost line behind the retainer you invoice.
Sources: OpenRouter quickstart
CRM, proposals, invoicing and client comms
The business side of the stack is the easiest to over-buy. A free CRM or a spreadsheet is enough until you have more than a handful of open deals; HubSpot, for one, offers a free CRM. Quote your first pilots with the free proposal generator, then move to Proposals in the console once you run agents on Qoren, because it prices from your real platform cost and your margin. Invoice from the accounting tool you already use: Qoren bills you, never your client, so your invoice carries your price. For comms, email plus one shared chat channel per client is plenty. Agents can post their results to Slack, Telegram or Discord, or send them from a mailbox of their own on plans that include mailboxes, and you forward the brief under your name. You can also point agents at your own agency: one keeps your pipeline honest, another drafts payment reminders for your approval. If you are weighing an all-in-one agency suite instead of separate tools, the GoHighLevel comparison sets the two approaches side by side.
Sources: HubSpot free CRM
Build vs buy, and how the stack changes as you grow
Build the parts clients pay you for and buy the rest. Your product is the judgment written into each agent: its instructions, its skills, the rules for each client and the way you check its work. Hosting, isolation, invoicing and proposals are not, so buying them usually costs less than your hours. The exception is an offer where host-level control is the point, such as strict data residency: then own the stack on purpose. As you grow, the stack barely changes shape. At the first client you need a template and a trial. At a few clients you need one environment per client and cost per client to price renewals. Once you pass 5 clients you outgrow the founder plans' environment limit, and Ultimate allows up to 10 client environments plus API access, so you can script client setups with the Qoren CLI.
Related guides
Playbook
How to start an AI automation agency, and sell what SMBs actually buy
One guide to starting an AI automation agency: what small businesses actually buy, how to deliver it with managed agents, and real retainer pricing.
Read guideBuild vs buy
Build your own agent stack, or use a managed platform?
Build versus buy for agencies delivering AI automation: running your own servers and agent stack versus a managed platform with per-client environments.
Read guideFounder guide
How to get your first AI automation client, from zero to a signed retainer
Get your first AI automation client without ads or a portfolio: ask people you know, sell one narrow job, run a two week pilot, then convert it to a retainer.
Read guideFounder guide
AI automation agency niches: how to pick one that pays every month
AI automation agency niches judged by who pays, how often work recurs and how hard it is to deliver, each mapped to an agent template, crowded ones included.
Read guideTools, templates and next steps
n8n vs Qoren
A workflow tool you host, vs autonomous agents on managed environments.
OpenMake vs Qoren
A visual automation tool you build and maintain, vs autonomous agents on managed environments.
OpenZapier vs Qoren
App-to-app automation, vs long-running autonomous agents.
OpenGoHighLevel vs Qoren
An all-in-one white-label CRM and marketing suite for agencies, vs managed AI agents you run per client.
OpenSee what each client costs
See what each agency client cost you on Qoren over a period, split into hosting, models, web search and mailboxes, so you can bill your clients with confidence.
OpenAI agent running cost calculator
Estimate an agent's monthly token spend, hosting, and a safe cap.
OpenQoren pricing
Every plan, with what each one includes.
OpenFrequently asked questions
What tools does an AI automation agency need?
One per layer: a CRM, a workflow tool such as n8n, Make or Zapier, an agent runtime with hosting, model access, a proposal tool, invoicing, and a channel to reach clients. Start with free or existing tools where you can, and add a second tool in a layer only when a client forces it.
Should I use n8n, Make or Zapier?
Use the one you already know for your first client; speed matters more than the choice. n8n can run on your own server, and its license allows building and running workflows for clients on your instance as long as they cannot edit them. Make bills in credits and Zapier in tasks, so check how a busy client scales on each.
Do I need a workflow tool if I use AI agents?
Not always. Fixed paths between apps are cheaper and more predictable in a workflow tool, while triage, drafting and monitoring suit an agent. Many agencies run both, with each doing the part it is good at.
Should I self-host agents for my clients?
Only if you have the operations time, or host-level control is part of what you sell. Otherwise every client adds a server to keep up, patch and watch. The build vs buy guide for agencies puts numbers on that trade.
Which Qoren plan fits a new AI automation agency?
Starter or Pro while you have one to a few clients: Starter allows up to 3 environments and Pro up to 5, so each client can have its own. Past that, Ultimate allows up to 10 environments and adds API access. Every monthly plan starts with a free trial, and the pricing page lists current prices.
Run your first client's agent without running a server.
Start free, deploy one template for one client on the free credits every new account gets, and see what it really costs before you quote. Starter and Pro fit a founder with one to a few clients: up to 3 and 5 environments, so each client can have its own. Once you pass 5 clients, Ultimate is the plan for a roster, with up to 10 client environments and API access.