RPA and rules-based tools work when every input looks the same. They break on an invoice in a new format or an email that doesn't fit the template.
That messy work is where most manual hours live. We use rules for the routine cases and a model for the rest, so only real exceptions reach a person.
RPA follows fixed rules, so it stops at an invoice in a new layout or a docket a driver photographed at an angle. Intelligent automation, also called AI automation, adds a model for the steps that need reading, sorting or deciding and keeps rules for the rest, so the whole task gets finished.
We build it for Australian mid-market operations teams in logistics, manufacturing, food and services. 22% of medium Australian businesses used AI in 2024-25, up from 3% in 2021-22, according to the ABS (2026). Rules handle the routine, a model handles the exceptions, and a person steps in wherever a wrong call is costly.
Four kinds of work where people are still the glue between systems.
Approvals, hand-offs, status updates and data moving between systems, so nobody has to chase a job through the business. We follow the work with the people who do it, cut the steps that add nothing, then automate one workflow end to end with the exceptions designed in.
Invoices, forms, emails, and reports that arrive in a different shape every time. We read, classify, and extract what matters from unstructured inputs, the judgement layer that rules-based tools cannot handle.
Where a task needs reasoning rather than a fixed rule, an AI agent handles it inside your process and escalates when it is unsure. The foundation for connected data and agent capability as your operations mature.
Disconnected tools that do not talk force people to be the integration. We connect your systems so data flows once and every team works from the same numbers, not five spreadsheets that disagree.
Take a freight business where two people spend each morning keying proof-of-delivery dockets that drivers photograph and email. A model reads each docket for the job number, signature and notes; rules check them against the transport system. When everything matches, the job closes and moves to invoicing.
Mismatches land in a short queue with the photo and the reason. The number agreed up front is the morning those two people get back and how much sooner jobs reach invoicing.
Workflow automation moves a job through its steps and its people so nobody has to chase it. In operations businesses these five pay back fastest.
Purchase orders, quotes, credit, variations. The slow part is not the decision, it is the chasing. Approvals route by rule, escalate when late and leave an audit trail.
New clients, staff, suppliers and sites. The checklist runs itself, collects the documents and holds back any step that is not ready.
Requests arriving by email, form and phone go into one queue, get sorted by a model and reach the right person with the right priority.
Inductions, inspections and reviews on a schedule. The calendar runs itself, chases the evidence and keeps the record an auditor asks for.
The Monday report rebuilt by hand is assembled from the source systems and sent, so the meeting is about decisions.
If a step adds nothing, remove it. The best workflow projects take out more steps than they automate.
We map the process as it actually runs, not as the manual describes it. Where does the effort concentrate, where do the exceptions cluster, and which steps are pure copy-and-paste. That map tells us what is worth automating.
We automate the cases that follow a pattern and design clean handoffs for the ones that do not. The aim is not zero humans, it is zero wasted human effort on work that never needed judgement.
Every automation is tied to a number: hours returned, errors cut, turnaround reduced. We measure it against that target and tune it. Automation that does not move the number gets reworked or removed.
Buy when your process looks like everyone else's. Build when the process is how you compete. Most clients end up with both.
Standard approvals, simple forms and light routing. A no-code workflow tool is cheap and fast, and your team can change it without a developer. Where a $30-a-month tool does the job, we will tell you to buy it.
Rules nobody else has, systems that do not connect cleanly, volumes that make per-user licences expensive. We build a custom layer for those two or three steps and keep the tool for the rest.
The honest answer comes from mapping the process first. As a rule, anything high-volume and pattern-based is a strong candidate: data entry and reconciliation, document processing, status chasing, and moving information between systems. We start by mapping where the hours go, then automate the work that is worth removing, not everything that can be automated.
RPA and tools like Zapier follow fixed rules and stop when an input changes shape. Intelligent automation adds a model for the steps that need judgement, so only true exceptions reach a person.
Not before we start, but readiness is part of the work. AI and agentic automation cannot be implemented unless the operational foundation is ready, clean processes and connected data. We assess that early, and where the groundwork is missing we build it as part of the engagement rather than bolting AI onto a broken process.
You measure hours returned by timing the job before it is automated and counting what still reaches a person after. Before we build, we sit with the people doing the work and time it: how many items a day, how long each one takes, how many need a second look. That gives a baseline in hours a week. Once the automation is live, every run is logged, so we can count how many items went straight through and how many landed in the exceptions queue. The difference is the hours returned, reported weekly against the number agreed at the start.
When the automation is unsure or wrong, the item goes to a person instead of going through. Each step has a confidence check. If the model cannot read a field, or the result does not match what your system expects, the item stops and lands in a review queue with the original document and the reason it stopped. Nothing is posted to your accounting or job system on a guess. Early on, we set the checks tight so more items go to review, then loosen them as the logs show where it is reliable. Every decision is logged, so you can always see why an item went the way it did.
The first automation is usually live within weeks, not months, because we start with one job rather than a programme. Mapping the process takes a few working sessions with the people who do it. The build follows, then a period where the automation runs alongside your team and every result is checked before anyone relies on it. The timeline depends most on access: how quickly we can connect to the systems involved, such as your accounting package or job system, and get sample documents. We confirm the timeline when we scope the job, before any build starts.
No. The aim is not zero humans, it is zero wasted human effort on work that never needed judgement. Automation returns hours to your people so they spend them on the work that actually requires a person, the work that grows the business. In practice, we involve the people who do the job from the first mapping session. They know where the exceptions hide, they help set the checks, and they work the review queue once it is live. That is usually what turns early worry into support: the automation takes the part of the job they liked least.
It starts with a Rapid Insight Workshop at a fixed $9,000 + GST, which defines exactly what is being automated and what it is worth in hours returned. The build is then quoted before any work starts, as whichever stream fits its size and complexity. You get a real number, not a rate card.
Automation that touches records or makes decisions needs to be auditable, so every run is logged with its inputs, its output, and the rule it applied. That log is what turns an automation into something an auditor or a regulator can review. Access is scoped so the automation reaches only the systems and fields it needs. Your data stays in an environment you control and is never used to train anyone else's model. If a wrong decision carries legal or safety exposure, we design a person into the path rather than automating it away.
Workflow automation moves a job through its steps and its people. RPA does one repetitive computer task inside a step. An approval that routes itself and escalates when late is workflow automation; the bot that then keys the approved order into the ERP is RPA. Most operations problems need both, built in the same project.
It should, and that is decided up front. Rules that change often, such as thresholds, approvers and pricing bands, go in settings your team controls. Rules that never change go in the build. Getting that split right is what keeps an automation in use years later.
They map the process as it actually runs before quoting. They recommend an off-the-shelf tool when one fits. They design for exceptions instead of ignoring them. And someone in your business can change the rules afterwards without paying a developer.
Python
Node.js
TypeScript
OpenAI
LangChain
n8n
AWS Lambda
REST APIs
Webhooks
PostgreSQL
Airflow
Docker
A Rapid Insight Workshop is two sessions with your key people that map how work really gets done and where time and money are lost. It costs a fixed $9,000 + GST, and about three weeks later you get a prioritised roadmap, a business case and a product specification you keep either way.
CodeFish builds intelligent document processing in Adelaide that reads invoices, dockets, forms and PDFs and writes the data into your systems without anyone typing it in. It copes when suppliers send the same information in fifty different layouts.
CodeFish builds RPA software in Adelaide that does the clicking, copying and re-keying your staff do between screens. We will also tell you plainly when a proper integration or an AI agent is the better answer.
An AI agent is software that owns a job end to end, such as invoicing or compliance checks, and only hands you the exceptions. A first agent is usually proven in an Ignite build over four to eight weeks, then run and improved under an Apex partnership (from $15K a month).
Most AI projects that stall were started in the wrong place. Know where your business stands before you spend: five free questions, two minutes, and you get a score, your weakest area and the first move that fits.
