TL;DR
The Metavision runs its agency site and a portfolio of its own brands with a very small team, and the reason that works is not one clever AI tool. It is a boring, shared foundation (one content system, one hosting and deployment pattern, one payment provider) with AI doing specific, bounded jobs on top: drafting, image and voice production, checks, reminders and reports. The tools matter less than the rules around them. Every brand gets its own content project and inbox, every change is verified on the live site, and anything that spends money, publishes or talks to a customer waits for a human yes.
Why we are writing this
People ask us which AI tools they should use. It is the wrong first question, and the honest way to show why is to open our own toolbox. We run our own original brands (among them The Biker Babies, Cats On Crack, The Fuglys and Labrats), our own print-on-demand storefronts, and this agency site, all on one stack. Here is what is in it, what each piece is for, and what still goes wrong.
The foundation: one stack, many sites
Every site we run is built the same way:
- Astro for the front end. Pages arrive as plain HTML, so they are fast and easy for both search engines and AI crawlers to read.
- Sanity as the content system. Each brand has its own Sanity project, so a change to one brand's content can never leak into another's.
- Netlify for hosting, builds and small serverless functions (order handling, form handling, scheduled checks).
- Stripe for payments, and Printful to print and ship merchandise for the brands that sell physical products.
- MailerLite for newsletters and Resend for transactional email such as order confirmations.
The point of sharing one pattern is that a fix learned on one site applies to all of them. When something breaks, we already know where to look.
Where AI actually does the work
AI sits on top of that foundation in five specific places. None of them is "AI runs the business".
1. Planning and drafting. We use Claude as a team of specialist agents (content, marketing, commerce, analytics, finance and others), each with its own written brief and its own action log. They draft articles, check copy against our published prices, read analytics, and hand work to each other in writing. A person reads the output before anything goes live.
2. Image and video production. Higgsfield generates and refines visual work for our brands and for articles like this one. Every brand has a written style and a "never" list that prompts must respect, so a family brand's visuals never borrow from an adult one.
3. Voice. ElevenLabs produces voiceover for animated work and spots, so a character can sound the same across many short pieces.
4. Checks and reminders. Small scheduled functions do the dull, important things: notice that a customer has not approved a proof, notice that an approved order has no tracking number, flag a monitoring failure. These do not fix anything on their own. They tell a person, early.
5. Reports. Weekly reads of traffic, sales and the mailbox are summarised into a short list of decisions for the owner, rather than a dashboard nobody opens.
The rules that make it safe
The tools are replaceable. These rules are not.
- Verify on the live site, never assume. A successful build log is not proof. After any change we load the real page and check the expected text is there.
- Humans hold the gates. Publishing, prices, payments, anything sent to a customer and anything that deletes data all wait for a person to say yes. AI agents prepare; a person releases.
- One brand, one container. Separate content projects, separate inboxes and separate mailing lists per brand. It costs a little more admin and prevents the worst kind of mistake.
- Write decisions down. Every agent keeps a dated log, and every handoff is a short written note. When a question comes back three weeks later, the answer is on file.
What still goes wrong
Being honest about this is the most useful part of the post.
The weak points are the handoffs between systems, not the systems. Payments, printing and shipping each worked; what failed for us was a manual step in between that nobody was reminded about. The fix was not a smarter AI. It was a written status on each order and a scheduled check that asks "has this moved?" and tells a person when it has not.
AI drafts confidently, including when it is wrong. An agent once reported that none of our blog posts were indexed, based on a web search. When we connected Google Search Console, the direct check showed otherwise. Indirect evidence is a hint; the primary tool is the answer.
More tools is not more capacity. We review our subscriptions regularly and cut the ones that are not earning their place. A tool that is not part of a weekly workflow is a cost, not an asset.
How to apply this to your own business
You do not need our stack. You need the same order of operations:
- Fix the foundation first. One place for content, one way to deploy, one payment route. If those are messy, AI multiplies the mess.
- Pick one bounded job for AI. A draft, a check, a reminder, a report. Something you can describe in a sentence and verify in a minute.
- Put a human gate on anything external. Money, publishing, customers.
- Add a check before you add a feature. The cheapest automation we run is the one that notices something has stalled.
If you want the longer version of that framework, our guide to business automation with AI walks through it step by step, and the most common AI implementation mistakes covers the traps.
FAQ
What AI tools does a small business actually need?
Fewer than you think. Start with one tool for one bounded job, such as drafting or a scheduled check, on top of a tidy website and content system. Add a second only when the first is part of a weekly routine.
Can AI agents run a business on their own?
Not safely. In our setup agents draft, check and report, but publishing, prices, payments and customer messages wait for a person. The value is speed and coverage, not autonomy.
Is a headless CMS worth it for a small business?
It can be, if you run more than one site or brand. A headless content system like Sanity lets each brand keep its own content while sharing the same front-end pattern, so improvements carry across.
How do you stop AI images going off-brand?
Write the rules down before you generate anything: a short style guide and a list of things the brand never shows. Then check every image against it. The prompt is not the control; the written rules are.
What is the first automation worth building?
A reminder that something has stalled: an unanswered quote, an unapproved proof, an order with no tracking. It is cheap, it is easy to verify, and it protects customer trust more than any new feature.
Want this for your business?
We set up stacks and automations like this for UK businesses, sized to the team you actually have. See our workflow automation service, or book a discovery call and we will look at where AI would earn its place in your week.