Short answer: most AI automation agencies run a layered stack: a workflow engine (n8n, Make or Zapier, sometimes custom code) at the center, large language models from one or more providers for the thinking, specialist platforms for voice (Vapi and similar), video (HeyGen and similar) and messaging (WhatsApp Business Solution Providers), a data store for memory and records (a database, vector store or spreadsheet), and monitoring and logging to see what the automation did. The tools vary by agency; the layers rarely do.
That matters when you're hiring one. An agency's AI automation tools stack decides what you can maintain yourself later, what you'll pay monthly after the build, where your data sits, and how hard it is to leave. This guide describes each layer, names tools only as examples (no rankings, and check current docs and pricing for any of them), and ends with the questions to ask before signing. If you're earlier in the process, start with how to hire an AI development agency.
Why think in layers rather than tools?
Tool names date quickly. A platform that's popular this year may be acquired, repriced or overtaken next year. Layers don't change: something has to trigger and orchestrate the work, something has to reason, something has to talk to customers on each channel, something has to store state, and something has to tell you when it breaks.
Thinking in layers also makes agency proposals comparable. Two agencies quoting the same project with different tools can be judged on the same questions: who owns each layer, what does it cost to run, and what happens if we swap it?
Here's the stack at a glance:
| Layer | Job | Example tools (not a ranking) |
|---|---|---|
| Workflow engine | Triggers, steps, branching, retries | n8n, Make, Zapier, custom code (Python, Node) |
| Models | Reasoning, classification, extraction, drafting | APIs from OpenAI, Anthropic, Google, Mistral; open-weight models self-hosted |
| Voice | Phone agents, speech in and out | Vapi, Retell AI, Bland AI; voices from ElevenLabs and others |
| Video | Avatar and generated video | HeyGen, Synthesia and similar |
| Messaging | WhatsApp, SMS, email, chat | WhatsApp Business Platform via a BSP, Twilio-type APIs, email providers |
| Data | Memory, records, documents, embeddings | Postgres (often via Supabase), Airtable, Google Sheets, vector databases |
| Monitoring | Logs, traces, alerts, cost tracking | Workflow execution logs, LLM observability tools, uptime monitors, dashboards |
What sits in the workflow layer?
This is the backbone. The workflow engine receives triggers (a form, a webhook, a schedule, a new email), runs steps in order or in branches, calls the AI and the other layers, and handles retries and errors.
Agencies pick based on the client and the job:
- n8n for teams wanting self-hosting, code steps alongside visual nodes, and agent-style workflows.
- Make for visual branching, data mapping and a cloud service with many connectors.
- Zapier for the widest app coverage and simple, linear automations.
- Custom code when volume, complexity, testing or compliance outgrow the no-code tools.
Many agencies run more than one. What matters to you is which one your automations will live in, whether you'll have admin access, and whether your team could edit a workflow after the agency leaves. Our comparison of n8n, Make and Zapier covers the trade-offs, and when to move to custom code covers the exit.
What sits in the model layer?
The model layer is the language models the workflows call. A few patterns are common:
- API models from the major providers for most tasks, often with different models for different jobs: a small, fast one for classification, a larger one for drafting or reasoning.
- Open-weight models self-hosted or via a hosting provider, where data residency, cost at volume or customization matters.
- Retrieval (RAG) to ground answers in your documents rather than relying on the model's memory.
- Structured output and tool calling so models return fields and trigger actions reliably.
Questions to ask: whose API account are the models billed to (yours, ideally), is customer data used for training under the provider's terms, and how will model changes be tested?
What about voice, video and messaging?
These are the customer-facing channels, and each has its own specialist platform and its own rules.
Voice. Platforms such as Vapi, Retell AI or Bland AI orchestrate phone calls: telephony, speech-to-text, the model, text-to-speech, and tool calls back into your systems. Voice providers like ElevenLabs supply the speaking voice. Compliance for outbound calling (consent, do-not-call registries, recording notices) sits with you, not the platform.
Video. Avatar platforms such as HeyGen or Synthesia generate presenter videos from scripts, often through an API. Consent for anyone's likeness or voice is mandatory, and external videos should be disclosed as AI-generated.
Messaging. WhatsApp automation goes through the WhatsApp Business Platform, usually via a Business Solution Provider (BSP), with opt-in and template rules that get accounts banned when ignored. SMS and email have their own deliverability and consent rules. Our overview of WhatsApp AI agents for business covers that channel.
For each channel, ask: who holds the account and phone numbers, who is responsible for policy compliance, and what the per-minute or per-message running costs look like at your volume.
What sits in the data layer?
Automations need somewhere to keep state: customer records, conversation history, documents, embeddings for retrieval, logs of what was done.
Common choices range from a managed Postgres database (Supabase and similar services are popular with agencies) to Airtable or Google Sheets for lighter workloads, plus a vector database or the database's own vector support for retrieval. Your existing CRM, ERP or accounting system is often the real system of record, with the automation reading from and writing to it.
The questions that matter: where does the data live (which country, which vendor), who owns the account, how is personal data protected and retained, and can you export everything if you leave?
What sits in the monitoring layer, and why do agencies skip it?
Monitoring is the layer most often missing from a cheap build, because it doesn't show in a demo. It includes:
- Execution logs from the workflow engine: what ran, what failed, what each step received and returned.
- LLM observability: prompts, outputs, latency, token costs per run, so you can see quality drift and spend.
- Alerts: a failed workflow, a spike in errors, a cost threshold crossed, a voice agent with long silences.
- Dashboards for the business metrics that matter: leads handled, bookings made, hours saved, escalation rate.
- Review queues where a human looks at samples and low-confidence cases.
An automation without monitoring fails quietly. Ask every agency how you'll know when something breaks, and who gets the alert. Our guide to AI agent monitoring and observability covers what to track.
How do you evaluate an agency's stack?
Ten questions, in rough order of importance:
- Who owns the accounts? Workflow engine, model API keys, phone numbers, WhatsApp account, database, domains. Your name on each, or a clear transfer plan.
- What are the monthly running costs at our volume, layer by layer, after the build is done?
- Can our team edit workflows, or is every change a ticket to the agency?
- Where does customer data live, which vendors see it, and how long is it kept?
- How are credentials stored and scoped? Shared logins and keys in prompts are red flags; see keeping API keys and customer data safe in automation tools.
- What's the human checkpoint on consequential actions, and who staffs it?
- How is it tested before and after changes, including model updates?
- How will we know when it breaks, and what's the response time?
- What's the exit? Export formats, documentation, handover sessions.
- Why these tools for us, rather than the tools the agency always uses?
A good agency answers all ten without defensiveness. A great one raises half of them before you do.
What are the risks in a typical agency stack?
- Vendor sprawl. Seven layers can mean seven subscriptions, seven sets of credentials and seven places data flows. Fewer, well-chosen tools beat a long list.
- Lock-in by neglect. Not because the tools trap you, but because nobody documented the workflows or exported the data.
- Cost shape surprises. Per-task, per-minute and per-token billing all scale with usage. Ask for a cost model at three volume levels.
- Policy violations. WhatsApp bans, outbound-call complaints and undisclosed synthetic media are business risks that land on you, not the agency.
- Security. Shared accounts, unscoped keys and logs full of personal data are common in fast builds. Insist on least privilege.
- Over-automation. Removing humans from decisions that need judgment. Keep approvals on money, customers and anything regulated.
How should a small business use this?
Before talking to agencies, sketch your own version of the table above: which layers your project needs, which systems you already have (CRM, accounting, phone, WhatsApp), and where your data must stay. Then ask each agency to map their proposal onto it. The conversation gets concrete fast, and it's much harder to be sold a stack that suits the agency more than you.
Frequently Asked Questions
Which automation platform do most AI agencies use?
There's no single answer; n8n, Make, Zapier and custom code are all common, and many agencies use more than one depending on the client. What matters is whether the platform suits your volume, data requirements and your team's ability to maintain it.
Should I own the tool accounts or let the agency own them?
Own them, or have a written transfer plan. Accounts for the workflow engine, model APIs, phone numbers, WhatsApp and databases hold your data and your running automations. If the agency owns them, leaving means rebuilding.
How much does an AI automation stack cost to run?
It depends on volume and layers. Expect a mix of subscriptions (workflow engine, database, monitoring) and usage-based charges (model tokens, voice minutes, messages, video credits). Ask for a layer-by-layer estimate at your current and expected volumes, and check current vendor pricing yourself.
Do I need all seven layers?
No. A lead-triage automation might need only the workflow engine, a model, your CRM and basic monitoring. Voice, video and messaging layers appear only when those channels are part of the job. Be wary of proposals that include layers your use case doesn't call for.
Can my team maintain the stack after the agency leaves?
Only if that was designed in: your accounts, documented workflows, readable prompts, a test set, and a handover. Ask for these as deliverables, not favors.
Want a stack that stays yours?
If you're comparing agency proposals or planning an AI automation build and want a second opinion on the tools and the ownership model, tell us what you're trying to automate. We build agentic AI systems in the client's accounts, documented and monitored, so what we build keeps working after we step back.