How our AI automation stack works
Built on Claude, n8n, Supabase and Obsidian, but this page is not about the tools. It is about the numbers they move: hours of repeat work removed, leads answered in minutes instead of days, and buyers who find you when they ask an AI. Every figure below is measured, and every figure carries its source.
One Integrated System, Not a Collection of Tools
Most agencies cobble together disconnected SaaS tools. We build a unified operating system where every component talks to every other, in real time.
Claude is the intelligence layer. It reads your documents, writes your queries, and orchestrates your workflows via MCP connections.
n8n is the automation engine. It moves data between platforms, triggers actions, and runs scheduled operations without manual intervention.
Supabase is the data layer. A production-grade PostgreSQL database that stores everything from CRM records to inventory levels, accessible via real-time APIs.
Obsidian is the knowledge hub. Your vault of SOPs, project context, and documentation that Claude can read directly to stay informed about your business.
What that buys you, measured
Independent research on what happens when businesses put systems like these to work. These are category results from published studies, not CH-ISE client figures.
+14%
more output per person, per hour
Support agents given an AI assistant resolved 14 percent more issues per hour in a study of 5,179 agents. The newest staff improved the most, at 34 percent.
Source: Stanford and MIT economists, National Bureau of Economic Research
Under 12 months
typical payback on automation
Organisations that deployed process automation reported the investment paying for itself in under a year.
Source: Deloitte global automation survey, 400+ organisations
90%
report better quality and accuracy
Nine in ten organisations said automation improved quality and accuracy, and 92 percent said compliance improved. Machines do not mistype on a Friday afternoon.
Source: Deloitte global automation survey
Departments We Automate
The same stack adapts to whichever team has the bottleneck. Here's where we've shipped real systems.
Sales
- AI call scoring against behavioural metrics with full transcripts
- Lead gen pipelines, personalised outreach, follow-up SMS
- Real-time agent leaderboards and pipeline dashboards
HR & Recruitment
- CV and resume screening against custom hiring criteria
- Candidate scoring pipelines from inbox to dashboard
- Knowledge base agents that answer policy and SOP questions
Finance & Accounting
- AI invoice and receipt parsing into structured ledger entries
- Multi-tenant accounting sync across Xero and CRMs
- Per-vehicle, per-project, or per-client profitability dashboards
Operations & Supply Chain
- Predictive inventory and reorder-point engines across warehouses
- Multi-tenant ERP and intercompany purchase-order workflows
- Real-time monitoring and breach alerts via Slack, email, or SMS
Transparent Infrastructure Costs
The entire stack runs on tools that cost less than a single software license.
Which part of your business is the bottleneck?
The stack is the same every time. What changes is where we point it. Tell us where the hours disappear and we will tell you whether this is worth building for you.
A 10-minute call. We look at how you work now and whether this is worth doing for you.
Automated Lead Gen & Personalized Outreach
Manual prospecting is slow. We automate the entire pipeline from finding leads to booking meetings.
Our system starts by identifying high-value targets matching your ICP. It then extracts key details (Name, Company, Role) and uses AI to clean and format this data.
The Secret Sauce: We don't send generic spam. Our AI analyzes each lead's LinkedIn profile, company news, and job description to generate a hyper-personalized icebreaker.
Finally, it verifies the email address to protect your domain reputation and automatically launches a multi-step outreach campaign.
Extracts: Name, Company, Role, Summary
"Hey John, congrats on leading that chatbot project at XYZ Corp! Figured I'd reach out..."
Why speed and personalisation pay
Outreach is a numbers game with two known levers: answer fast, and make it personal. Both are measured, and both are what the machine does tirelessly.
7x
more likely to qualify a lead
Companies that contacted a new web lead within an hour were nearly seven times as likely to qualify it as those even an hour slower. 23 percent of firms never replied at all. An automated pipeline answers in minutes, every time.
Source: Harvard Business Review audit of 2,241 companies
+32.7%
more replies when personalised
Across 12 million real outreach emails, personalised messages earned roughly a third more replies. Generic cold email gets a reply just 8.5 percent of the time. Personalising by hand does not scale; AI research does.
Source: Backlinko study of 12 million outreach emails
40%
more revenue for personalisation leaders
Companies that excel at personalisation generate 40 percent more revenue than average players in their industry.
Source: McKinsey, Next in Personalization, 2021
Want this pointed at your ideal customer?
We build the pipeline around your ICP, your tone and your inbox, then hand you the meetings. Bring a rough idea of who you want to reach and we will map it out on the call.
A 10-minute call. We look at how you work now and whether this is worth doing for you.
Workflow Automation
We automate repetitive manual data entry and pipelines. If you're doing it manually more than twice, we can automate it.
This isn't necessarily AI, but pure automation for moving data between different platforms within a company via APIs. No more copy-pasting from spreadsheets to CRMs.
Common Use Cases:
Document Generation: Auto-populate Google Docs from CRM data.
Instant Invoicing: Trigger Xero/QuickBooks from CRM status.
Client Onboarding: Auto-create channels, folders & emails.
Run the numbers on one repetitive task
Thirty minutes a day of copy-paste work, one person. This is arithmetic, not a study. Swap in your own rates; the shape stays the same.
11 hours
lost every month
130+ hours
lost every year
R39,000+
a year at R300 an hour, for one task
Most businesses run five to ten of these tasks without noticing. Deloitte's survey of 400+ organisations found automation typically paid for itself in under 12 months, with 90 percent reporting better accuracy. Source: Deloitte global automation survey.
Doing something manually more than twice a week?
That is usually the cheapest thing to automate and the fastest to prove. Bring one repetitive task to the call and we will scope it there and then.
A 10-minute call. We look at how you work now and whether this is worth doing for you.
Your next client may never click a link.
People increasingly ask ChatGPT, Claude, Perplexity or Google's AI overview instead of scrolling a page of blue links. The numbers are already stark: 58.5 percent of US Google searches end without a single click, and when Google shows an AI answer, clicks to the number one result drop by roughly a third. The assistant reads a handful of sites and answers on the visitor's behalf. If it cannot read yours, you are not in the answer, and the visitor never learns you exist.
Here is the part most people miss. Sites built with modern page builders and AI tools usually assemble themselves in the visitor's browser. Google will wait for that. Most AI crawlers will not. They fetch the page once and read whatever came back, which on a lot of sites is close to nothing.
The traffic that does still arrive has changed too. Visits referred to US retail sites by AI assistants grew 1,200 percent in eight months, and those visitors arrive having already asked their question. Fewer idle browsers, more people closer to buying and far less patient.
So a site now has two jobs: be readable by machines, and convert the humans they send. We build for both, in that order.
Sources: SparkToro and Datos zero-click study, 2024. Ahrefs analysis of 300,000 keywords, 2025. Adobe Analytics, 2025.
What changed
Four shifts that turned websites from brochures into revenue engines.
AI answers instead of linking
Fewer casual visitors, more intentional ones. Being the source the assistant quotes matters more than ranking tenth.
Structure beats keyword stuffing
Machine-readable markup, real questions and answers, and clear authority signals are what let an assistant cite you with confidence.
One page per offer
A single homepage cannot serve every audience. You need a system of pages, one per offer and per campaign, each answering one question properly.
Everyone can ship a pretty site now
Builders and AI tools made good-looking easy, so looking good is no longer the differentiator. Being found and converting is.
What we build into every page
None of this is visible to a visitor. All of it decides whether a machine can find you, understand you, and repeat what you said accurately.
Readable with JavaScript off
Every page delivers its full text in the first response. Crawlers that do not run scripts still get the whole argument, not an empty shell.
Structured data that matches the page
Organisation, service, breadcrumb and question-and-answer markup, describing exactly what a visitor can see. No markup for content that is not there.
A written brief for AI crawlers
An llms.txt file telling assistants who you are, what you sell and which claims belong to whom, so they summarise you correctly instead of guessing.
Claims worth quoting
Every statistic carries its source on the page. Assistants prefer citable pages, and so do buyers who check.
One page, one question, one address
A unique title and description per page, a canonical address that does not redirect, and a sitemap that matches. Sounds obvious. Most sites fail it.
Fast, and stable while loading
Speed is a revenue number. A 0.1 second improvement lifted retail conversions by 8.4 percent and order values by 9.2 percent across 30 million sessions. Source: Deloitte and Google, Milliseconds Make Millions.
Do not take our word for it. Test this page.
We rebuilt this site against everything above and published the numbers. Run it through Google yourself, or view the page source and see the full text sitting there before a single script runs.
Google Lighthouse, mobile, homepage. Performance sits in the mid 80s and moves a few points between runs, which is normal on a live connection. Before the rebuild this site served 25 words to a crawler without JavaScript. It now serves over 900.
(Opens Google PageSpeed Insights in a new tab)
Being found is half of it. The page still has to sell.
Perfect markup on a page that confuses people is a well-indexed failure. Most websites are digital art projects: they look pretty and put nothing in the bank.
We don't do "empowering digital transformations". We build sales machines.
The principle is simple: don't make them think.
If a visitor lands and has to guess what you do, you have lost them. If the text is hard to read because of a clever design choice, they leave. If you talk about yourself instead of their problem, they do not care.
Design serves the copy. The copy sells the product. Everything else is a distraction.
Why our structure works
Readability is SEO
Google ranks pages that people actually read. We use high contrast and short paragraphs. No walls of text.
The "grandma test"
Navigation so simple your grandma could use it. If we confuse the user, we lose the sale. Clear buttons, clear offers.
Copy-first design
We don't pick a template and stuff words into it. We write the sales argument first, then build the design to highlight it.
The "no-fluff" design checklist
We use this internal checklist to audit every page. If it doesn't pass, it doesn't go live. Feel free to use it on your current site.
Readable without JavaScript?
Turn scripts off and reload. Whatever remains is what an AI crawler sees.
Can a machine tell what you do?
Structured data and a clear entity definition, not just pretty headings.
Talking too much?
Simple websites convert better. Don't overload with info.
Too much branding?
Clients care about results, not your logo.
Attention Grabbing Headline?
First impressions matter.
Visuals helping sell?
Images should relate to the struggle and tell a story.
Conversational Tone?
Avoid salesy, creepy, or aggressive language.
Distractions?
Get them to contact you ASAP. No links to socials.
Real Pictures?
Pictures should depict the text, no random art.
Contrast?
Ensure text is readable against the background.
Message > Beauty
Even art requires salespeople.
Client Results Focus
No 'about us' sections. Focus on them.
AI Art Quality
Avoid obvious flaws in AI generations.
Call to Action
CTA + Free value + Button to contact form.
Is your site readable by the assistants your buyers now ask?
Most are not, and there is no warning when it happens. On a short call we open your site the way a crawler sees it, tell you what is missing, and whether a rebuild is worth it for you.
A 10-minute call. We look at how you work now and whether this is worth doing for you.
Natural Language Query (NLQ)
In practice, this is the difference between asking your ops manager for a number on Friday and having it yourself in ten seconds.
Traditional chatbots often fail with structured data (like sales figures or inventory counts) because they rely on "fuzzy" semantic search.
We use NLQ (Natural Language Query). Our agents are given the context of your database schema. When you ask a question, the agent writes a precise SQL query, executes it against your database, and returns 100% accurate, quantitative results. No hallucinations, just hard data.
Why it matters:
- Accuracy: Get exact numbers, not estimates.
- Speed: Direct database querying is faster than reading through documents.
- Complex Logic: "Show me sales from Q3 vs Q4 grouped by region" becomes a simple query.
FROM sales
WHERE quarter = 'Q3';
Chat to your Company's Documents
Traditional RAG (Retrieval Augmented Generation) is dumb. It treats every question as a search query. If you ask for a calculation, it fails.
We build Agentic RAG systems. The agent acts as a "router". It analyzes your request first:
• Need specific data? → Route to SQL Database (NLQ)
• Need conceptual info? → Route to Vector Store (Semantic Search)
This "Best of Both Worlds" approach ensures you get the right tool for the job, every time.
Contextual Retrieval & Hybrid Search
Standard vector search has a "loss of context" problem. If you search for "revenue growth", a chunk saying "it grew by 3%" is useless if you don't know which company or which quarter it refers to.
We implement Contextual Retrieval (a technique pioneered by Anthropic). Before indexing, we use AI to prepend context to every chunk of data."It grew by 3%" → "In Q3 2023, ACME Corp's revenue grew by 3%..."
We combine this with Hybrid Search (Keyword + Semantic) to ensure we catch both specific terms (like error codes) and general concepts.
"The revenue increased by 5%."
Missing context: Which year? Which product?
Contextual Enrichment
AI analyzes the parent document and prepends context to the chunk.
[Context: Q3 2024 Financial Report for Product X]
"The revenue increased by 5%."
What finding answers actually costs
The expensive part of company knowledge is not storing it. It is every hour someone spends looking for it, and every decision made without it.
19%
of the workweek spent hunting for information
McKinsey estimated knowledge workers spend almost a fifth of their time searching for and gathering information. That is close to a full day per person, per week, before any productive work happens.
Source: McKinsey Global Institute, 2012
+34%
faster ramp for your newest people
When answers are on tap, junior staff gain the most. Novice agents with an AI assistant performed 34 percent better in a 5,179-person study, because the system carries the knowledge of your best people to everyone else.
Source: National Bureau of Economic Research
Your team already has the answers. They just cannot find them.
Policies, SOPs, contracts, past quotes. We make them answerable in plain language, with the numbers pulled from your actual data rather than guessed at.
A 10-minute call. We look at how you work now and whether this is worth doing for you.