What a Claude deployment actually looks like for a small business

“Claude South Africa” and “Claude packages South Africa” are the searches that bring the most new visitors to this site, which tells us people want to know what deploying Anthropic’s Claude in a small business actually involves. Not the demo. The Tuesday-afternoon reality. So here is ours, because we run our own studio on it.

What we use it for every day

Claude is not a chatbot in the corner of our office. It is wired into the tools we already use, with its own access, doing real work under supervision.

• Infrastructure operations. It has a dedicated SSH key on our servers. It reads logs, diagnoses a slow site, finds the process eating the database, drafts the fix and applies it when we say so. During a security incident it did the first sweep of the box, listed every suspicious file and cron entry, and wrote the timeline.
• Analytics and reporting. It queries Google Analytics, Search Console, Ad Manager and Merchant Center through their APIs with service accounts we created for it, so monthly client reports are pulled from the source rather than screenshotted from dashboards.
• Ad operations. It loads weekly creative bookings into Ad Manager from a brief, creates line items per booking, and checks they are serving.
• Client correspondence. It reads our inbox, drafts replies and reports in our tone, and never sends anything. We send.
• Code and content. Theme and plugin work, SQL, migration scripts, SEO metadata for a hundred pages, this article.

The rules that make it safe

None of that works without guardrails, and the guardrails are the deployment.

• Separate identities. One service account per client, per system. A key labelled for one client never touches another client’s property.
• Drafts, not sends. Email, invoices, anything outward-facing is drafted for a human to send.
• Destructive actions wait. Deleting, restarting, dropping: it stops and asks, with the evidence, before it acts.
• Memory with discipline. It keeps notes on every project: what was done, what broke, what not to do again. Those notes are reviewed and corrected like any colleague’s.
• Self-hosted where it matters. Sensitive data stays on infrastructure we control. The model sees what it needs for the task, not the company’s filing cabinet.

What a deployment looks like

For a small business the first week is less about the model and more about plumbing and policy.

1. Map the work. Which repetitive, judgement-light tasks eat your week? Reporting, first-draft correspondence, data cleanup, monitoring, documentation. We pick two or three.
2. Wire the access. Service accounts, API keys, read-only where possible, scoped per system. No shared passwords.
3. Write the rules. Tone, what it may send, what it must ask about, what it must never touch. These go in its standing instructions.
4. Run it supervised. Two weeks of it doing the work with a person checking every output. Corrections become rules.
5. Measure. Hours saved, errors caught, turnaround times. If the numbers are not there, we stop.

What it costs, honestly

The subscription is the cheap part. The cost is the setup and the first month of supervision, and the price depends entirely on how many systems it needs to reach. For a business that mostly needs correspondence and reporting it is a few days of work. For one that wants it operating infrastructure it is a proper project with the same security care as any other system with keys to your servers.

Who this is for

Owner-run businesses and small teams in South Africa who are drowning in admin and do not have an IT department to hand this to. You do not need a data scientist. You need someone who has already made the mistakes, which is why we start with our own studio. Here is what we offer.