You evaluate machine learning readiness by testing one specific use case against six areas: a clear prediction target, enough labelled historical data of known quality, infrastructure to train and run a model, the skills to build and maintain it, governance for its decisions and a business case that survives a simple baseline. Skyo Digital, an engineering-led AI consultancy in Perth, Western Australia, applies this method with industrial clients across Australia, the UK and Europe.
This article is written for engineering managers, operations leaders and technical leads in energy, resources and engineering businesses who have a machine learning idea on the table, such as predicting equipment failures or forecasting production, and want to know whether it can work before funding a pilot. A short readiness evaluation is far cheaper than a pilot that stalls because the data, people or business case were never tested.
How is machine learning readiness different from AI readiness?
Machine learning readiness is narrower than AI readiness, because it asks whether your own data can train a model that predicts something useful, while AI readiness asks whether the whole organisation can adopt AI safely and profitably.
Machine learning readiness
A machine learning model learns patterns from your historical records, so readiness depends heavily on the volume, labelling, quality and history of that data. A classic industrial example is a model that predicts pump or compressor failure from historian and work order data.
Generative AI readiness
Generative AI tools such as Microsoft Copilot and Claude arrive already trained, so readiness depends more on access to well-organised documents, security controls, licensing and user skills. A business can be ready for generative AI to draft reports from engineering documents long before it is ready to train a predictive model on sensor data.
AI readiness
AI readiness covers both, and adds strategy, knowledge management, workforce competency and cybersecurity across the organisation. Skyo Digital's AI readiness assessment scores eight areas at that organisational level, and a machine learning readiness evaluation sits inside it as a deeper test of individual use cases.
What are the six areas to evaluate?
Each area should be scored against the specific use case in question, because a data set that is ready for one model can be useless for another.
1. Use case and prediction target
The use case must name the decision the model will improve, what it will predict and how far ahead. "Predict which conveyor idlers will fail in the next 30 days so maintenance can plan replacements" is a usable target, while "use machine learning on our maintenance data" is a direction that still needs work.
2. Data volume, labelling, quality and history
This is usually the deciding area. A supervised model needs examples of the outcome it is predicting, so a failure prediction model needs enough recorded failures, with consistent failure codes, linked to the sensor and operating data from the period before each one. The history must also reflect current conditions, because data from before a major plant modification or a change in operating philosophy may teach the model the wrong patterns.
3. Infrastructure and data access
The team needs a way to extract, join and store the relevant data, an environment to train and test models, and a route to put predictions in front of the people who act on them. In industrial settings this includes getting historian or SCADA data out of operational networks safely, without weakening the boundary between operational technology and corporate IT.
4. Skills and ownership
A model needs a data scientist or engineer to build it, a domain expert who understands the equipment or process, and a named owner who will monitor performance and retrain it as conditions change.
5. Governance and risk
Governance covers who approves the model, how its accuracy is tested, how its decisions are explained and when a human must stay in control. For businesses selling into Europe, Article 10 of the EU AI Act sets out data governance and quality requirements for high-risk AI systems, including relevant, representative and as far as possible error-free training data, and in Australia the Government's Guidance for AI Adoption sets out essential practices for accountability, testing and human control.
6. Business case and baseline
The business case should compare the model with the simplest alternative, such as a fixed threshold alarm, a rule of thumb or a better maintenance schedule. Google's Rules of Machine Learning opens with the advice to launch without machine learning first, and a good evaluation measures how much better the model needs to be than that baseline to justify its build and running costs.
What does a machine learning readiness checklist look like?
The checklist below can be run by an engineering or operations team in an afternoon, and every "no" answer points to a specific gap to close before a pilot.
We can describe the decision the model will improve, what it predicts and how far ahead.
We know what the decision is worth, in downtime, throughput, safety exposure or cost.
We have recorded examples of the outcome we want to predict, ideally dozens to hundreds for each type of event.
Those outcomes are labelled consistently, for example with standard failure codes rather than free text.
The input data, such as historian tags or inspection results, can be linked to each outcome by asset and time.
The history covers current operating conditions and enough seasonal or production cycles to show normal variation.
We can extract the data without manual spreadsheet work every time.
We have somewhere secure to train, test and run a model, and a way to show its output to the people who act on it.
We have the data science skills in house or through a partner, plus a domain expert with time to help.
A named owner will monitor the model after launch and decide when to retrain or retire it.
We know who approves the model and what level of accuracy is acceptable for the decision.
We have measured or estimated how well a simple rule or threshold performs today.
A business that can answer yes to most of these questions is ready for a time-boxed pilot. A business with several gaps in questions 3 to 7 needs data foundation work first, and that work usually pays back on its own through better reporting.
What does a readiness evaluation look like in practice?
A practical evaluation tests one or two candidate use cases against real data samples rather than interviews alone.
Consider an illustrative example of a processing plant that wants to predict failures on its slurry pumps. The evaluation starts with the maintenance system and finds five years of work orders, but failure codes were applied inconsistently and many entries say only "pump repaired". The historian holds good vibration and pressure data, although tags were renamed after an upgrade two years ago. The conclusion is that a failure prediction model is premature, while a condition alarm based on vibration thresholds is achievable now, and a twelve month fix to failure coding would make a proper model viable later.
This outcome is common and valuable, because it stops a pilot that would have failed and turns a vague ambition into a funded sequence of work.
What this looks like in practice at Skyo Digital
Skyo Digital evaluates machine learning use cases as part of its AI readiness review, following a set method: agree objectives, scope and sites, interview stakeholders across leadership, operations, engineering, maintenance, IT and OT, review real systems and data samples, identify and size use cases, score the eight readiness areas, then present findings, a roadmap and recommendations.
Our work with Santos shows why the data foundations come first. The Santos engineering function had no clear view of incoming work against capacity, so we built a lightweight task capture process and two dashboards that gave daily and weekly visibility and moved prioritisation from intuition to data. With consistent records in place, the next phase uses AI to capture and categorise engineering requests from emails. The full story is in the Santos case study, and our five layer AI readiness framework explains why data governance and integration sit underneath every successful use case.
How long does it take and what does it cost?
A Skyo Digital AI readiness review typically runs four to six weeks, depending on the number of sites, teams and systems in scope. For a small team with a focused domain, such as an engineering team, it costs around A$7,000 to A$9,000. Work with larger teams and across a whole organisation generally lands around A$15,000 to A$20,000, depending on the size and complexity of the organisation.
How Skyo Digital can help
Skyo Digital runs AI readiness assessments that test machine learning, generative AI and automation use cases against your real data, and where the data is the gap our data foundations and analytics strategy work sets out how to close it. We work from our Perth head office across Australia and with a global team and UK associates across the UK and Europe. Our services page also describes what an AI readiness assessment covers for mining and heavy industrial firms.
Frequently asked questions
How much data do you need for machine learning?
The amount depends on the problem, and the number of recorded outcomes matters more than the total volume of data. A model that predicts equipment failure needs enough labelled failures of each type to learn from, which often means dozens to hundreds per failure mode, while years of sensor data with only a handful of failures will struggle. Where outcomes are rare, anomaly detection or simple threshold rules are often the better starting point.
Do we need a data scientist before we start?
You do not need one to evaluate readiness, because the first questions are about the decision, the data and the business case, and engineers and operations staff can answer most of them. You will need data science skills to build and maintain a model, either in house or through a partner, together with a domain expert who understands the equipment or process and a named owner who will look after the model once it is live.
Is machine learning readiness the same as generative AI readiness?
The two overlap but test different things. Machine learning readiness asks whether your own historical data can train a model that predicts something useful. Generative AI readiness asks whether your documents, security controls, licences and people are ready to use pre-trained tools such as Microsoft Copilot or Claude. Most engineering businesses are ready for generative AI use cases sooner, while predictive models usually need data foundation work first.
What should we do if we are not ready yet?
Fix the gaps that block your highest-value use case, starting with consistent labelling and data capture, and use a simple rule or threshold in the meantime to deliver some value straight away. Data foundation work such as standard failure codes, linked asset records and automated data extraction usually improves reporting and decision-making on its own, so the investment pays back before any model is built.
Does the EU AI Act affect machine learning projects in industrial businesses?
It can, depending on how the system is used. The EU AI Act places data governance and quality obligations on high-risk AI systems, and some industrial uses, such as safety components in critical infrastructure, may fall into that category. Businesses operating in or selling into Europe should check the classification of each use case early, and the readiness evaluation is a sensible place to do it.
To find out whether your machine learning idea is ready for a pilot, email info@skyodigital.com or book an introduction with the Skyo Digital team.
