The Tools AI Automation Agencies Use

When a client hires an AI automation agency, they often expect some magical black box: complex AI systems doing mysterious things behind the scenes.

The reality is more practical. Most AI automation projects use a surprisingly consistent set of tools, the same platforms I'm going to walk you through here. What makes a project work is how those tools are combined and configured around a specific business problem.

Whether you're evaluating agencies, considering doing it yourself, or just curious what's under the hood, this is what's actually in the toolkit.

How does a modern AI automation stack fit together?

A typical stack has four layers:

1. AI models that read, write and analyze (OpenAI, Claude and so on)
2. Workflow platforms that connect and coordinate everything (Make.com, n8n)
3. Databases that hold the information (Airtable, Supabase)
4. APIs and connectors into the tools you already run (CRM, email)

Almost every automation project touches all four. Data comes in, gets processed, triggers an action and produces an output, a bit like an assembly line.

Tool #1: Large language models

The models themselves are what make "AI automation" different from regular automation.

OpenAI (GPT-4, GPT-4o)

OpenAI is still the default choice for most automation projects. Its models power everything from chatbots to document analysis.

Agencies use it for:

• Content generation (emails, social posts, product descriptions)
• Document summarization and extraction
• Customer support responses
• Lead qualification and scoring
• Data categorization and tagging

You pay per token through the API, and the rates change often enough that you should check OpenAI's pricing page rather than trust any number in a blog post. For short tasks like tagging an email, the cost per run is usually a rounding error next to the labor it replaces.

It has the widest adoption, which means the best integrations, consistent quality on most tasks, fast responses from the newer models and solid function calling for structured output. Unless there's a specific reason to pick something else, it's where most projects start.

Claude (Anthropic)

Claude is OpenAI's main competitor, and plenty of agencies (us included) prefer it for certain work: long-form content and analysis, complex reasoning, long documents, writing that has to match a specific voice, and anything that needs careful, nuanced answers. Pricing is in the same range as OpenAI's.

In my experience it holds a consistent tone better, handles longer documents, follows complicated instructions well and writes better. We reach for it on content-heavy projects and document analysis.

Open-source models (Llama, Mistral)

For cost-sensitive or privacy-focused projects, open-source models run on your own infrastructure. Agencies use them for high-volume work, on-premise deployments where data can't leave the building, fine-tuning for a narrow domain, and the odd case where a commercial model doesn't fit.

You pay for the servers instead of per token, the data stays private, and you can fine-tune them. The tradeoff is that someone technical has to deploy and run them, and quality is generally a step below the top commercial models. They make sense for healthcare or finance projects with strict data requirements, or for volumes where API costs would get out of hand.

How do LLMs fit into an automation?

Say you want to categorize incoming support emails automatically:

1. An email arrives and triggers the workflow
2. The workflow sends the email to GPT-4 with a prompt: "Categorize this email as: billing, technical, general, or spam. Return only the category."
3. GPT-4 returns the category
4. The workflow routes the email to the right team or queue
5. An auto-response goes out if one applies

The LLM does one focused job, categorization, inside a larger workflow. That's the pattern: AI handles the judgment calls and traditional automation handles the plumbing. For which steps should stay deterministic, see where AI workflow automation pays.

Tool #2: Workflow orchestration platforms

These coordinate everything else.

Make.com

Make.com is my primary recommendation for most businesses. The visual builder hits the right balance between power and usability.

Agencies use it for multi-step workflows, connecting business apps, scheduling and triggering automations, error handling and logging, and conditional logic. As of November 2026 the Core plan is $12 a month for 10,000 credits, with enterprise plans above that (Make pricing).

The visual builder is something clients can understand and maintain, the error handling and debugging are good, and the OpenAI and Claude modules are solid. That's why we like it: clients can see what's happening, so when we hand off a project they can maintain it without us.

n8n

n8n is the choice for technical teams or projects with specific requirements. Agencies use it for self-hosted automations where data never leaves the client's infrastructure, technical workflows, high-volume processing (self-hosting means no per-execution fees) and custom integrations that need code.

Self-hosting is free, and the cloud Starter plan is $20 a month on annual billing (n8n pricing). You get complete control when you self-host, no execution limits, code nodes for deep customization and a strong developer community. We pick it for healthcare, finance or any project where data can't leave the client's infrastructure, and for high-volume projects where Make.com's pricing would get expensive.

Custom code (Node.js, Python)

Sometimes there's no substitute for writing actual code: logic that doesn't fit a visual builder, performance-critical steps, heavy data transformations, or APIs with no pre-built connector. We switch to code when a workflow gets complex enough that the visual version is harder to maintain than the code would be.

Tool #3: Flexible databases

Every automation needs somewhere to store data, and the choice of database matters more than most people realize.

Airtable

Airtable is the Swiss Army knife of automation databases, a spreadsheet and database hybrid that non-technical people can actually use. Agencies use it as the central data store, for client-facing dashboards and views, as a lightweight CRM or project tracker, for content calendars, and for approval workflows.

Pricing, from Airtable's pricing page on annual billing:

• Free: 1,000 records per base
• Team: $20 per seat per month
• Business: $45 per seat per month

Non-technical users can view and edit data, the interface has plenty of views, the API is strong, automations are built in, and it's excellent for prototyping. It has record limits on the lower tiers, slows down with very large datasets, has API rate limits and isn't suited to high-frequency writes. It's still our default for most projects, especially when the client needs to see or change the data.

Supabase

For projects that need more power, Supabase gives you a real PostgreSQL database with a friendly interface. Agencies use it for higher-volume storage, SQL queries, real-time data, user authentication and projects that might grow a lot.

Pricing, from Supabase's pricing page:

• Free: 500 MB database and 5 GB of egress
• Pro: from $25 a month
• Team: from $599 a month

You get real Postgres with proper queries, a generous free tier, real-time subscriptions, built-in auth, and room to scale. We use it for data-intensive projects, anything that needs real-time updates, and projects we expect to grow large.

Google Sheets

Don't laugh. Google Sheets is useful for quick prototypes, data the client edits all the time, reports that need to be shared easily, and clients who won't learn a new tool. Everyone knows how to use it, it's free and sharing is easy.

It also slows down as it grows, the API is clunky and it isn't a real database. We use it when simplicity matters more than capability, or for a quick pilot before building something sturdier.

Tool #4: Specialized AI tools

Beyond the general LLMs, a few specialized tools handle specific jobs better.

Document processing

Tools like Rossum, Docsumo or a custom GPT-4 Vision setup pull data out of documents: invoices, contracts, forms and receipts. You upload a PDF or image, the AI extracts structured data (vendor name, total, line items), and that data flows into your systems.

Voice AI (ElevenLabs, PlayHT)

For automations that involve audio, like customer calls, voice assistants or narration. Typical uses are automated voice responses, content narration, IVR systems and personalized audio messages.

Image generation (DALL-E, Midjourney via API)

For automations that need to create visuals: product image variations, social graphics, thumbnails and personalized visual content.

Embeddings and vector search (Pinecone, Weaviate)

For building AI that understands your own data: semantic search over documents, FAQ bots trained on your knowledge base, content recommendations and matching similar items.

Tool #5: Analytics and monitoring

You can't improve what you don't measure.

Custom dashboards

Most agencies build dashboards that show executions per day, week or month, error rates and failure points, time saved against the manual baseline, and business results like leads processed or emails sent. The usual tools are Metabase, Looker Studio, Retool, or the views built into Airtable and Supabase.

Error monitoring

Automations fail, and when they do you need to know right away. The built-in monitoring in Make.com and n8n covers most of it, with Sentry for code-based automations, custom Slack or email alerts, and PagerDuty for critical systems.

Business intelligence

To connect automation data to the rest of the business, agencies use Looker Studio (free, in the Google ecosystem), Metabase (open source, self-hosted) or Tableau and Power BI at the enterprise end.

What does it look like when these tools work together?

Here is how the pieces fit in a common build: lead qualification and routing.

Say a sales team is drowning in leads. Reps spend hours qualifying prospects who aren't a good fit, while the hot leads wait in the queue. The build looks like this:

1. A new Typeform submission triggers the workflow
2. The Clearbit API adds company data
3. GPT-4 scores the lead against your criteria
4. Make.com routes the lead based on that score
5. High-score leads go to a priority queue in HubSpot
6. Hot leads trigger an immediate Slack alert
7. Everything gets logged to Airtable for analysis

The flow:

Typeform → Make.com → Clearbit (enrich) → GPT-4 (score) →
Router (by score) → HubSpot + Slack + Airtable

Reps only see leads that pass the scoring step, and a hot lead gets a Slack alert the moment the form is submitted instead of waiting for someone to check the queue.

That's what "AI automation" usually means in practice: several ordinary tools wired together, with AI making the judgment calls a person would otherwise have to review.

How do you tell whether an agency knows what it's doing?

Now that you know the toolkit, here are the questions I'd ask.

Questions to ask

"What's your primary workflow platform?" Good answer: a clear preference with reasons (Make.com for accessibility, n8n for self-hosting). Red flag: "We built our own proprietary platform," which often means lock-in and no portability.

"How will we maintain this after the project?" Good answer: training, documentation, a handoff process and ongoing support options. Red flag: "Just call us whenever you need changes," which means you're dependent forever.

"What happens when the automation breaks?" Good answer: error monitoring, alerting and documented troubleshooting. Red flag: "It won't break." Everything breaks eventually.

"Can we see examples of similar projects?" Good answer: specific case studies with results and an architecture overview. Red flag: vague claims with no concrete examples.

Signs of a good agency

A good agency asks about your processes before it proposes tools, because the tools should fit the problem. It plans the handoff from day one, so documentation and training aren't afterthoughts. It's honest about limits and tells you when AI isn't a good fit. And it measures impact in time saved, errors reduced and revenue gained, rather than counting how many automations it built.

Signs to be careful

Be careful with an agency that promises magic ("Our AI will 10x your business") without specifics, keeps everything proprietary when you should own your automations and be able to move them, won't give you clear pricing up front, or doesn't ask questions. If they aren't curious about your business, they'll build the wrong thing.

Should you build it yourself or hire an agency?

Not every project needs an agency.

When to do it yourself

• The automation is straightforward (3 to 5 steps, one or two apps)
• You have time to learn the platforms
• Your needs might change frequently
• Budget is very limited

When to hire an agency

• The project is complex (multiple systems, AI components, custom logic)
• Speed matters more than cost
• You need it done right the first time
• The automation is business-critical
• You don't have technical people in house

The hybrid approach

Many businesses do both: simple automations in house, complex projects with agency help. You build internal skill and still get expert help where it counts.

FAQ

Q: How much does it cost to work with an AI automation agency?

It depends almost entirely on scope: how many systems are involved, whether there's an AI component, and whether you need ongoing support after launch. Ask any agency for a fixed estimate before you start. For reference, our own one-time sprints start at $9,999 and run 4 to 8 weeks, and ongoing work is on a monthly retainer (see pricing).

Q: How long do automation projects take?

A simple workflow between two apps can go live quickly. Projects that touch several systems take longer, because discovery, testing and handoff take longer. Whatever the size, ask for a timeline that covers discovery, testing and handoff as well as the build.

Q: What if the automation stops working?

Good agencies include monitoring and alerting. Most issues are minor, like an API change or an expired token, and get fixed quickly. Critical automations should have a fallback procedure.

Q: Do I need technical staff to maintain automations?

It depends on the stack. Non-technical people can maintain Make.com and Airtable with some training. n8n and custom code need more technical comfort.

Q: What's the typical ROI on automation projects?

It depends on what the automation replaces. Estimate it before you build: hours saved per week × hourly rate × 52 weeks, plus the cost of the errors it removes, compared against build and running costs.

Q: Can automations handle sensitive data?

Yes, with the right precautions. Self-hosted options (n8n, Supabase) keep data on your infrastructure, and the commercial platforms have enterprise security features. Always review how each tool handles data.

Where should you start?

Start by auditing your current processes. Where do you spend time on repetitive tasks? Where do errors happen? Where are the bottlenecks? Those are your automation opportunities, and whether you build in house or hire help, knowing this toolkit makes those decisions easier.

If you want help identifying opportunities or building your automation stack, that is what our AI workflow automation services cover.

If you are still deciding whether you need an agency at all, what an AI automation agency is explains what one does and how to tell a real one from a reseller. Everything we build as an AI automation agency is listed in one place.

That's all I got for now. Until next time.