State Zero
Automation & AI

AI and automation that earn their place — not because they're new, but because they remove real work.

Most "AI features" add noise. We build the parts that quietly do the job: an enquiry that lands in the right place, a question answered from your own procedures, a first draft a person still checks before it goes out. Useful outcomes are the objective — automation and AI are simply how we get there.

Independent trades business owner reviewing an AI-drafted job quote on a tablet before it goes out to the customer
The framing

AI is a capability. Not the product.

Somewhere in the last few years "AI" became the pitch instead of the outcome. We think that's backwards. Nobody wakes up wanting an AI assistant — they want the quote out the door, the enquiry actioned, the question answered without pulling someone off the tools to go and look it up.

We do not use AI to replace the people who make your business valuable. We use it to reduce the work they should not have been doing manually in the first place.

Where we do use it, it is grounded in your own information, scoped to a defined task, and reviewed by a person — until there is a proven reason to loosen that.

Grounded
Answers are built from your documents and records, not a guess
Scoped
Each automation or assistant can see only what it needs to do its job
Reviewed
A person checks before anything is sent, spent or committed, by default
The three layers

Automation, assistance and agentic action are three different things.

We treat them separately because they carry different risk and need different controls. Most businesses get most of the value from the first layer alone. Some benefit from the second. Very few need the third applied broadly — and we will tell you if that's where you are.

01Layer one

Workflow automation

This is the layer that does the most good for the least risk. No interpretation, no judgement calls — just consistent execution of rules you already follow, applied every single time instead of most of the time.

  • Integration between two systems that don't currently talk to each other
  • A notification the moment a job, order or ticket changes status
  • A task or job created automatically from a form, email or booking
  • A document moved, filed or routed to its next step without anyone dragging it there
  • Customer or job data kept in sync across platforms instead of slowly drifting apart
ENQUIRY → JOB → NOTIFICATION
Web enquiry received TRIGGER
Checked against your rules LOGIC
Job created & assigned ACTION
Runs the same way every time — nobody has to remember to do it.
knowledge.internal
What's our returns process for a faulty item bought in-store?

Faulty items can be returned within 14 days with proof of purchase. Check the item against the fault categories before offering a replacement or refund.

Source: Returns & Warranty Procedure
02Layer two

Knowledge & assistance

This layer answers questions using what your business already knows — procedures, past jobs, product details, policies — instead of what one person happens to remember, or what the open internet guesses.

It only has access to what you give it. It shows where an answer came from. And when it doesn't know, it says so, instead of filling the gap with something confident-sounding and wrong.

03Layer three

Agentic workflows

This is the layer people picture when they hear "AI agent" — a system that reads a situation, recommends what should happen, and in defined cases takes the action itself. It's also the layer that deserves the most caution, so we build it narrowly.

Every agentic workflow gets a specific task, a defined permission boundary, a clear point where a person approves it, and a full record of what happened and why. We are not in the business of handing a system the keys and hoping for the best.

Where this fits
Agentic workflows suit narrow, well-defined, repeatable decisions — not "run the business" scenarios. If a task doesn't fit that shape yet, we'll say so rather than force it.
TICKET → DRAFT → APPROVAL → ACTION
Support ticket received TRIGGER
AI drafts response & action SUGGESTED
Staff member approves HUMAN
Action taken & logged RECORDED
Nothing acts without an approval and a record of why.
How we keep it safe to rely on

The controls that make automation and AI safe to switch on

None of the above works without guardrails. These are the controls we build into every automation, assistant and agentic workflow, applied in proportion to the risk involved.

Scoped permissions

Every automation and assistant can only see and touch what it has explicitly been given access to — nothing more.

Human approval

Anything that sends, spends or commits the business waits for a person to say yes, until there is a proven reason not to.

Auditability

A record of what ran, what it was given, and what it did — so any action can be traced back and explained.

Source grounding

Answers are built from your actual documents and records, with the source shown, not generated from a guess.

Exception handling

When something falls outside the rules, it is flagged for a person instead of forced through anyway.

Data minimisation

Only the information a workflow actually needs is exposed to it, not your entire system by default.

Practical security

Sensible, standard safeguards — access control, encryption in transit, least-privilege credentials — not a marketing checklist.

Testing before autonomy

New automations run in the open, watched, before anything is allowed to act without supervision.

What this looks like in practice

A sense of the kind of problems we take on

Illustrative
These are representative examples of the kind of automation and AI work we build, not a list of completed projects. Our delivered client work to date has been website engagements — this is the shape we expect automation and AI work to take next.

Enquiry to task, automatically

A new website enquiry becomes a job in the right system, assigned to the right person, without anyone re-typing it.

Status notifications that make sense

A text or email the moment a job moves stage, instead of a customer ringing to ask what is happening.

A staff knowledge assistant

Grounded answers from your own procedures and product information, not a generic chatbot guessing.

A first-draft quote or reply

AI prepares the first version from the enquiry details. A person checks it before anything goes out.

Documents that move themselves

An invoice or job document filed, matched and routed to its next step without manual handling.

Exceptions surfaced, not buried

A mismatched order or an overdue approval flagged for a person, instead of quietly slipping through.

Common questions

Questions we get asked

Will this replace staff?
No. We build automation and AI to take on repetitive, low-judgement work — retyping, chasing, searching — so the people you already have can spend their time on the parts of the job that actually need a person.
How do you keep our data secure?
Every automation and assistant only sees what it is explicitly given access to. We apply standard, sensible security practices — access control, least-privilege credentials, encryption in transit — rather than treating security as an afterthought.
What happens when it gets something wrong?
It is designed to stop, not guess. Where a workflow is not confident, or a situation falls outside what it has been scoped for, it hands off to a person instead of pushing through anyway.
Do we need to be "AI-ready" before we start?
No. Most of the value comes from layer one — rules-based automation — which needs clear rules and reasonably clean data, not a data science team. We start from wherever your business actually is.
How do you decide what gets automated, assisted or agentic?
By risk and repetition. High-volume, well-defined, low-risk tasks are good automation candidates. Anything involving money, judgement calls or reputation gets a human approval step, at least until there is a track record that says otherwise.

What's the task nobody has time to do properly?

Tell us about the workflow, the question staff keep asking, or the approval that takes too long. We'll help work out which layer — if any — actually fits.