Has a marketing tool stack ever been this cheap?
If I were building martech for a startup today, it would look like this:
- CRM: Twenty, free if you host it yourself
- Website: Astro on Cloudflare Pages, free
- Analytics: PostHog, free
- Search visibility: Google Search Console, free
- Newsletter: Buttondown, free up to 100 subscribers
- Contact data: Apollo, free basic tier
- Automations: n8n, free if you host it yourself
- Meeting booking: Cal, free
- Text, images and nearly everything else: Claude, $20 a month
Total: about $20 a month.
Claude, n8n and Twenty each get their own section below. Here is what the other six do, and where their free versions stop:
| Tool | What it does | Where the free version stops |
|---|---|---|
| Astro on Cloudflare Pages | Astro is an open-source framework for content-driven websites. Cloudflare Pages hosts the site | Astro is free. Pages allows 500 builds a month and 20,000 files per site |
| PostHog | Shows what visitors do on your site, including session recordings | 1 million events and 5,000 session recordings a month |
| Google Search Console | Shows which searches bring people to your site, with clicks, impressions and position | It’s a free service from Google |
| Buttondown | Sends your newsletter and keeps the subscriber list | 100 subscribers. Analytics and automations are paid add-ons |
| Apollo | A database of B2B companies and contacts for building prospect lists | Meant for light use. Credits arrive each month and expire unused |
| Cal.com | A booking page connected to your calendar | One user. Teams cost $12 per user a month, billed yearly |
Free tiers have limits, and hosting things yourself needs someone who knows how. But a few years ago, marketing automation alone meant an annual contract and an implementation project.
So the budget question has changed. Tools are cheap. The expensive mistake now is paying for tools that don’t change how anyone works.
This guide covers the AI marketing tools we think are worth paying for, the ones that usually disappoint, and how to decide.
Fix four things before you buy any tool
What kind of AI marketing stack should you build? It doesn’t matter much until you can answer these four questions:
- What value do you create?
- For whom?
- What do you do better and differently than others?
- Who are those others? In other words, what’s your category?
Once you can answer them, you can start thinking about where to get account lists, which channels reach your buyers, how to write the messages and what the call to action is.
Established companies often have their positioning in a brand book from a few years ago. The target group is “decision-makers in medium and large companies.” That’s true, but it doesn’t help anyone when you set out to grow internationally.
Try answering each of the four questions in one sentence. If that’s hard, start there. Our AI marketing strategy asks for two of the same answers in its first two weeks: who you sell to and why customers choose you.
Worth the budget
The short version first. Vendor prices are list prices, and the Claude Code figure is what our own team pays.
| Tool type | Examples | What we use it for | Cost |
|---|---|---|---|
| Frontier AI model | Claude, ChatGPT, Gemini | Writing, research and analysis for every marketer | About $20 a month per person |
| AI that works in your files and systems | Claude Code | Proposals, automations, SEO analysis and CRM updates | Roughly €1,000 a month for us |
| Automation layer | n8n, Make | Connecting everything else | n8n is free if you host it yourself |
| CRM your AI can read and write | Twenty | Our agents update it themselves | Free if you host it yourself |
| Brand in a format AI can read | A git repository with Claude skills | On-brand banners and copy | Time, not licenses |
| Custom tool | A web app built for one task | Your highest-volume task | A one-off build |
| Data enrichment | Clay or your own code | Company and contact data | Clay uses credits. Own code needs maintenance |
1. A frontier AI model for every marketer
Claude, ChatGPT or Gemini, at around $20 a month per person.
Practically every marketing team we work with uses at least one of these, often alongside whatever tool the company officially provides. When people keep using a tool the company didn’t give them, that’s what real demand looks like.
2. An AI tool that works inside your files and systems
Claude Code is the most important tool we pay for. It costs us roughly €1,000 a month.
It writes proposals from client meeting transcripts. It builds automations in n8n. It analyzes thousands of keywords for SEO projects, writes content plans, sends emails and updates our CRM. A first draft of a proposal is ready about 20 times faster than when we wrote it from scratch.
Compare that cost with one extra employee, or with outsourced development.
The way to get the most out of it is skills. A skill is an instruction file the AI reads before it does a piece of work: how the work is done, what to avoid and where to find the data. It can be a single file.
There are three common misunderstandings about skills:
- “You can just copy them from GitHub.” You can, but then you get someone else’s way of working. Another company’s skill is a bit like another company’s brand book. Interesting to read, but it isn’t yours.
- “Users don’t need to understand what a skill does.” They do. If you don’t know what a good result looks like, you won’t notice when the machine produces a bad one.
- “Everything should be automated.” Skills work best for work that repeats the same way and has a result you can check. Work that needs judgment stays with people.
We’ve written our own skills for sales, reporting and content. The best ones were always written by the person who does that work. Six of our sales skills are described in detail in Claude Code skills for sales.
3. An automation layer
n8n or Make. This is the glue that connects everything else. n8n is free if you host it yourself.
Singa’s inbound lead process runs on n8n connected to HubSpot. Every lead is enriched, scored and routed in under 2 minutes, and the morning sorting that took 45 minutes is gone. Read the Singa case, or compare the platforms in our workflow automation tools comparison.
4. A CRM your AI can read and write
Choose your CRM by its API. If your AI can’t read and write the data, you’ll be copying and pasting between windows for years.
That’s why we moved our own CRM to Twenty: every part of it is reachable through the API, and our agents update it themselves. We still use Airtable for operations: content calendar, project management and our proposal database. As a CRM, though, it starts to break down once you have thousands of contacts.
5. Your brand in a format AI can read
What will brand books look like in 2030? Probably not PDF files. Ours is already a git repository.
It holds our positioning, target group description, colors, fonts, design tokens and tone of voice. The same folder holds our workflow instructions as Claude skills. Our banner skill makes about 20 banner variations in roughly 5 minutes, and every one of them follows the brand because it reads from the same source. A person then picks the best ones.
Barona’s marketing team took this furthest. Their marketing OS is packaged as a Claude plugin that every marketer installs once. Its context-update skill takes raw material, like a sales call transcript or a new strategy deck, and turns it into structured knowledge the AI can use. A new hire can write a campaign brief on day one. The Barona case explains how it works.
This costs time more than licenses. And when someone changes a color or the tone of voice in the repository, the change shows up in everything new right away.
6. A custom tool for your highest-volume task
Olo-apteekki, a Finnish online pharmacy, used to build every campaign’s banners and animations by hand in Canva and Adobe. One full set took 20-30 minutes, across dozens of campaigns a month.
We built them a custom web app instead. It uses real product images, and AI handles only the layout. One click generates every banner size, and a second click produces the animations. The Olo-apteekki case has the details.
When one task repeats hundreds of times a year, a tool built for exactly that task beats a general one.
7. Data enrichment: Clay or your own code
Clay is still the handiest way to gather company and contact data without a developer. Claude Code can now do the same job in code, which gets cheaper once a lot of data runs through it every week, but someone has to maintain it. We compare the two in AI for lead generation.
Usually not worth it (yet)
Licenses without a process change
We’ve seen companies buy Copilot licenses for the whole organization, then find six months later that the number of active users is the same as one month after launch. The usage report goes to leadership every month, and nobody can name a process that changed.
A license pays off once you know which task it changes and who changes it. Until then, it’s a cost line.
The money goes to licenses. The return comes from changing the work.
Copilot Studio for demanding builds
We’ve done two projects where we tried to build sales tools with Copilot Studio. The first one was finished, and nobody uses it. In the second, Copilot Studio didn’t perform well enough. We moved to Power Automate calling AI models through Azure, which was too slow, and finally built the whole thing on Azure AI services.
Copilot Studio does work for focused jobs inside a Microsoft environment. At S-Group, the marketing team’s annual plans now start from a draft by a Copilot Studio agent. What makes it work is the rules: it may only use concepts that already exist in SharePoint, it keeps brand portfolios separate and it asks when something is unclear. See the S-Group case.
A built-in agent in every tool
Notion, Linear, Salesforce and HubSpot all ship their own agent now. Each one is a chat window in the corner of its product, and each sees only that product’s data.
Pay for tools your own AI can use through an API. Treat the built-in agent as a bonus. For sales tools, the four questions in AI sales agents make a quick test.
How to decide
Run every tool through five questions:
- Does it change a task that repeats every week?
- Can you check its output?
- Can your other tools and your AI reach it through an API?
- Who maintains it when something breaks?
- What does it cost per month compared with the hours it saves?
The last one is simple math. If a tool saves your team 10 hours a week, that’s roughly 500 hours a year. At a loaded cost of £50 an hour, that’s £25,000 a year. Most tools on the “worth it” list cost a fraction of that.
Our AI-powered marketing work starts with the same questions: what the team repeats every week, which tools can talk to each other and where the hours actually go.
The cheapest line in our stack costs $20 a month. The most expensive one is a license that changed nobody’s work.
