Automation & AI Agents for Business Workflows

Start here

Automation is a task that runs itself

Something happens, and a fixed sequence follows — with nobody remembering to do it.

Automation — one path, run the same way every time.

A quotation request arrives. The details get saved, your team gets a WhatsApp alert, a draft quotation is prepared, and a reminder is set for two days later. Nobody typed anything.

That is the whole idea. It is worth doing when a task repeats, follows the same steps, and currently depends on somebody remembering. If a task is different every time, automation is the wrong tool — and we will say so.

The other half

An AI agent decides, then acts

Automation follows a path. An agent works out which path to take — and loops until it gets there.

An agent — the loop back is the whole difference.

A scripted chatbot follows a decision tree and gives up the moment a customer phrases something unexpectedly. An agent reads the actual question, works out what it needs, looks it up in your documents, takes an action, checks whether that was enough, and goes again if it was not.

The rule we build into every agent

Unknown and wrong are not the same thing. If it does not know, it says so and hands over to a human with the full context. An agent that guesses confidently is worse than no agent at all — it costs you the customer and you never find out why.

The tools

Four ways to build it, and what each one costs you

Most shops know one platform, so every problem looks like that platform's problem. Here is the honest trade on each.

At a glance — the four, compared
	n8n	OpenClaw	Hermes Agent	Custom code
What it is	Workflow builder	Autonomous agent	Agent that learns	Your own software
Licence	Source-available	MIT	MIT	Yours outright
Reach for it when	The path never changes	It must act on a real machine	It should improve with use	It must be exactly right
Runs on	Cloud or your server	Local-first, your machine	Your machine or server	Anywhere
Time to live	Fastest	Fast	Medium	Slowest

n8n

A workflow builder. You draw the steps; it runs them on a schedule or a trigger.

Best forMoving data between services you already pay for, on a path that does not change.

Strengths

You can see the whole workflow as a picture, and so can your staff
Hundreds of ready-made connectors
Self-hostable, so it can sit on your own server
Quickest of the four to get something live

Costs and limits

Long branching logic gets messy fast
Source-available under a Sustainable Use Licence — not OSI open source, so read the terms if you plan to resell it
Judgement calls still need a model behind it

OpenClaw

An autonomous agent that lives on a machine and answers in the chat apps you already use.

Best forA hands-on assistant: run a command, check a file, reply on WhatsApp or Telegram, act without being asked.

Strengths

MIT licensed, genuinely open
Local-first — its memory is plain Markdown files on your own disk
Reaches shell, browser and files, not just APIs
Speaks WhatsApp, Telegram, Slack, Discord, iMessage and Signal
A heartbeat scheduler lets it act unprompted

Costs and limits

Real reach over a real machine — it needs scoped permissions and supervision, which is work we do at setup
Young project moving very fast; versions change under you
Not a business-process tool on its own — the guardrails are the build

Hermes Agent

An agent from Nous Research whose point of difference is that it learns between runs.

Best forWork you repeat with variation, where you want the thing to get better at your process instead of resetting each time.

Strengths

MIT licensed and self-hosted
Model-agnostic — swap the brain without rebuilding the agent
Keeps memory and builds reusable skills across runs
One agent and one memory across terminal, desktop and messaging

Costs and limits

Newest of the four; smallest body of production experience
A learning loop is a moving target — it needs review, not blind trust
Heavier to set up than a workflow you could draw in an afternoon

Custom code

Python or Node written for your process, with no platform in the middle.

Best forAnything that must be exactly right, run at volume, or do something no platform supports.

Strengths

No licence, no seat count, no platform limits
Testable, version-controlled, reviewable
Fastest at scale and cheapest to run once volume is real
Where our own systems live: the measurement backend, the tracking pixel, the daily content pipeline

Costs and limits

Costs the most to build
Nothing to look at — no visual canvas your staff can read
Needs someone to maintain it, which is either us or a developer you hire

We are not paid by any of them. n8n, OpenClaw and Hermes Agent are other people's tools; we pick between them per job and tell you which one you got and why. Most real builds use more than one — a workflow for the plumbing, an agent where judgement is needed, custom code for the part that has to be right.

The brain

Rented model, or one that is yours

Every agent needs a model behind it. The real question is whether your data leaves the building to reach it.

Hosted API

Someone else's model, called over the internet. The strongest models available, billed per use.

Strengths

Best quality available today
Nothing to install or maintain
Scales instantly

Costs and limits

Your text goes to a third party
Cost grows with every call, forever
Their outage is your outage
Prompts and prices can change under you

Open weights, your machine

A model you download and run yourself — Ollama to run it, Hugging Face to get it. Hermes models are a common pick because they are tuned for tool calling.

Strengths

Data never leaves your side
Fixed cost — the hardware, not the usage
Works with no internet
The version you tested is the version that keeps running

Costs and limits

Needs real hardware
Below top hosted models on the hardest reasoning
You own updates and uptime

We run this ourselves, not just recommend it

Our own research pipeline classifies text with a model running locally, specifically so the source material never reaches a third party. It is the same setup we would build for you — which is why we can tell you where it is genuinely good enough and where it is not.

The machine

It can run in your office or in a data centre

Local means nothing leaves. Server-side means it never sleeps. Which one you need is a business question, not a technical one.

Where the model lives. The trade is speed to start against what leaves your building.

Your existing PC or Mac

Cheapest honest start

A Mac mini or a desktop with a modern GPU already runs small and mid-size models well. For summarising, classifying, extracting and tagging — most of what business automation actually needs — this is usually enough.

A workstation GPU

More headroom

An RTX-class card gives you fast local inference for bigger models and room to run several jobs at once. The practical ceiling is how much video memory the card has.

NVIDIA DGX Spark

Serious local capacity

A desktop AI machine built for this: 128GB of unified memory shared between CPU and GPU, so large models load without shuffling data back and forth. Runs models up to around 200 billion parameters locally and fine-tunes up to about 70 billion. Two units link together for larger models still.

A server, yours or ours

Always on

The same software on a machine in a data centre instead of your office. Reachable from anywhere, backed up, and it does not stop when someone unplugs the wrong thing.

You do not need the biggest machine on this list. Most business automation is summarising, classifying, extracting and tagging — work a model that fits on hardware you already own does perfectly well. We would rather size it honestly than sell you a box you will not use.

The decision

How we pick, in four questions

You should be able to predict our recommendation before we make it.

The path never changes

n8n workflow

A form fills a sheet, a sheet alerts WhatsApp. Drawn, not coded.

Someone has to judge something

An agent

Reading an enquiry and deciding what it is needs a model, not a rule.

The data must not leave

Local model

Open weights on your own machine. Nothing goes to a third party.

It has to be exactly right

Custom code

Testable, reviewable, and it does not drift between runs.

The work

What we actually build

Pick the one that hurts most. The rest connects to it later.

Talking to customers

AI agent on your website, answering from your own documents
WhatsApp auto-reply that escalates to a human when it should
Support tickets organised so nothing is answered twice
Bookings, confirmations and reminders
Review requests sent at the right moment

Sales and leads

Every enquiry captured from website, WhatsApp, ads and forms
Instant WhatsApp or Telegram alert to your team
Follow-up reminders so nothing goes cold
Lead scoring and routing to the right person
A simple CRM without the enterprise price
Quotations drafted from your own templates

Back office

Invoices, forms and contracts generated automatically
Accounting and invoice sync between systems
Inventory and order sync across store, sheet and supplier
An internal assistant that answers staff from your SOPs
Onboarding sequences that run the same way every time

Data and reporting

Daily, weekly or monthly reports that arrive on their own
Live dashboards instead of five tabs and a spreadsheet
Pipelines that clean and combine sources that never spoke
Monitoring of prices, listings or competitors, within what each site permits
Custom integrations where no connector exists

Running it

Deployment on your cloud account, your office machine, or hosted by us
Local model setup with Ollama and open weights
Permission scoping and guardrails before anything touches live data
Monitoring, and fixing it when it breaks

Yours to keep

You own what we build

Workflow files, source code, and the account it runs on.

Whether it sits on your cloud account, a machine in your office, or a server we host, you get the workflow files and the source. If you ever want to move it in-house or hand it to another team, it goes with you. Nothing here is locked to us — which is the point. An automation you cannot leave is not an asset, it is a subscription.

Tell us one task you repeat every day

We will tell you which of the four it belongs on, whether it needs a model at all, and roughly what that costs. If automating it is not worth it, we will say that too.

Chat on WhatsApp Send an enquiry