AI Consulting for Manufacturing

No, but it changes the honest sequence and it changes the budget. Old machines are not the barrier people assume — a great deal of value in Australian manufacturing sits in the office rather than on the plant floor, in quoting, scheduling, purchasing and customer correspondence, none of which cares how old the CNC is. Where you do want machine data, the question is not the machine’s age but whether it exposes anything: many older machines have a controller with a serial or Ethernet port, a data logger, or at minimum a sensor that can be retrofitted for a few hundred dollars. What actually stops projects is not vintage, it is silence — a machine with no output and no practical way to add one. The mistake we see is a business deferring every AI decision until a capital replacement program that is five years away, while an entirely software-side opportunity sits unexploited in the front office. Sequence the office work first; instrument the plant when there is a specific question worth instrumenting for.

Sometimes, and it is oversold more than any other manufacturing AI use case. Predictive maintenance needs three things that most mid-sized plants do not have: continuous condition data from the asset, a history of failures labelled accurately enough to learn from, and enough failures to constitute a pattern. That last one is the killer. If a critical machine has failed four times in eight years, there is no statistical model on earth that will predict the fifth from that history — you have anecdotes, not data. Where it genuinely works is on populations: many similar assets, high duty cycles, continuous sensing, and a maintenance history that was actually recorded rather than reconstructed from memory. Where it does not, the honest alternative is usually condition monitoring with sensible thresholds, which is not AI and costs a fraction as much. A consultant who agrees with your predictive maintenance ambition before asking how many failures you have recorded is selling, not advising.

Most often in quoting and estimating, and it surprises people every time. Custom and semi-custom manufacturers live or die on quote accuracy and quote turnaround, and both are usually appalling — quotes take days because they wait for the one estimator who knows, and margins vary wildly because the estimate is judgement reconstructed under time pressure. Your historical job costing data contains the answer to what things actually cost, and almost nobody uses it. Beyond quoting: vision-based quality inspection where volumes are high and the defect is visual; demand forecasting where you have real order history and seasonal patterns; scheduling optimisation where the constraint set is complex; and the unglamorous back-office flow — purchase orders, supplier correspondence, delivery scheduling, compliance documentation. The consistent pattern is that the value is in the decisions surrounding production far more often than in production itself.

By treating the boundary as sacred, and by putting the burden of proof on anything that wants to cross it. Operational technology on a plant floor has different failure consequences from a spreadsheet: a badly behaved integration does not corrupt a report, it stops a line or hurts someone. So the working rule is that data flows out of OT and decisions flow back in only through whatever change and safety process your plant already runs — not through an API a vendor added last month. Read-only telemetry out is a fundamentally different risk proposition from writing setpoints in, and the two should never be scoped in the same conversation. Security matters too: connecting previously isolated plant networks to cloud services is a decision with a real threat model attached, and it deserves the same scrutiny your engineering changes get. Most of the value we find does not require crossing the boundary at all, which is a happy accident worth exploiting first.

A pilot is useful when it can fail visibly and cheaply. That means picking something with a clean measurement — quote turnaround time, first-pass yield on a specific defect, forecast error on a specific SKU family — rather than a vague ambition to modernise. It means running it beside the existing process rather than replacing it, so a bad week costs you nothing. It means a defined stop date and an agreed number that decides go or no-go, written down before the start, because otherwise the pilot runs forever and quietly becomes production without ever having been evaluated. And it means the people who do the work being genuinely involved, because a leading hand who thinks the system is nonsense will be proven right within a fortnight. The pilots that fail are the ones scoped to demonstrate success rather than test a hypothesis.

It opens with a free initial consultation — a conversation about the actual constraint in your business rather than a demonstration of anything. If there is a real opportunity, the next step is generally the AI Opportunity Audit at around $3,000: a written map of your workflows, what data genuinely exists versus what people believe exists, the systems it lives in, and a ranked shortlist with realistic effort and payback. The data reality check is the part manufacturers value most, because the gap between the two is usually large and nobody has written it down before. You keep the audit regardless, including the parts that recommend a data project rather than an AI project. Where it stacks up, we build and ship the first workflow into production rather than handing over a report. The audit generally earns its fee at around 20 or more staff.