The board signed off on an AI scheduling tool. The budget was there, the supplier credible, the pitch persuasive. Three months in, the delegate running the project went to pull the historical job data needed to configure the model, and found it in a ring-bound diary. Two systems that had never been connected. A spreadsheet rebuilt from scratch when the previous operations manager left eighteen months ago.
This pattern shows up repeatedly in specialist manufacturing. The cause is almost always the same: the data those tools need has never been captured in a form the AI can use.
What is the data-readiness problem in manufacturing AI?
Manufacturing AI tools, whether for predictive maintenance, job scheduling, or demand planning, run on structured historical data. They need records of what was done, when, by whom, on which machine, in what condition. Many specialist manufacturers at this scale carry that knowledge in paper job books, standalone spreadsheets, and machine systems that have never been connected. The AI tool arrives with nothing to work from.
The software itself is rarely the limitation. The tools available now are capable, affordable, and deployed successfully in larger operations with established digital infrastructure. The binding constraint in specialist manufacturing is that the data those tools need often does not exist in a usable form.
Gartner research, cited in Schellman’s analysis of AI implementation failures, found that 77 per cent of firms name data quality as their biggest barrier to responsible AI use. In specialist manufacturing, that figure reflects something specific to how these businesses operate. Larger enterprises have typically spent years building systems that generate structured data as a by-product of normal operations. A 20-person precision machining firm has usually not. Job durations, machine uptime, fault logs, maintenance histories: if captured at all, they exist on clipboards, whiteboards, or in the heads of experienced engineers who are approaching retirement.
BuildOps’ research on skilled trades and construction found that many firms at this scale run with minimal digital systems, with paper job books being common rather than exceptional.
Why does missing data kill a manufacturing AI project?
When an AI model trains on absent or incomplete data, it fills the gaps rather than flagging them. A job scheduling tool with no reliable job-duration history will still produce a schedule, and that schedule will be wrong in ways that take months to diagnose. The outputs look credible enough that nobody challenges the model. The firm spends the budget and gets confident nonsense.
The British Chambers of Commerce found that half of UK owner-managed businesses are now using AI, but with limited impact on headcount or productivity. Data readiness is part of that explanation. Tools running on incomplete data underperform against expectations, and the narrative quickly becomes that AI does not work in this sector, when the honest answer is that the AI was never given what it needed.
The ROI timeline compounds this. Helium42’s analysis of AI business cases found that 59 per cent of chief executives expect measurable AI results within 12 months, while realistic implementation timelines for operational AI run to 18 or 36 months. At a 20-person manufacturing firm with a data problem, that gap widens further. The board already has unrealistic expectations on timeline. Telling them the first six months will be spent digitising job records rather than running predictive maintenance models is a conversation the delegate has to have early, or they end up managing a failing project mid-stream.
Where does missing data actually show up in practice?
The symptoms are specific enough to identify before you run a formal data audit. The scheduling tool recommends staffing configurations that assume job durations the firm has never measured. The maintenance model flags patterns in sensor data that correspond to nothing on the actual maintenance record. The demand forecast produces a figure nobody can explain to the production floor.
Schedules that used to be built on the operations manager’s institutional knowledge now get built by an algorithm that lacks that knowledge. The experienced engineer who knew a particular job type always ran fifteen minutes longer on the afternoon shift has retired. That knowledge was never captured. The algorithm has no way to know it.
Predictive maintenance tools need sensor data and maintenance history to find correlations between machine readings and failure patterns. If the machine data was logged on paper forms that were never digitised, the model has nothing to correlate. The Infor UK survey found that 45 per cent of UK businesses cite data integration as a primary barrier to operationalising AI. Specialist manufacturing is particularly exposed because machine data tends to be distributed across proprietary control systems that were never built to share information with anything external.
Identifying where the gaps are is straightforward once you know what to look for. Take any input field that an AI tool will need, trace it back to its source, and ask whether it exists in digital form and whether it is consistent across time periods. That trace typically takes a day. What it reveals may take considerably longer to fix.
When should you digitise before you deploy?
Digitise first when you cannot pull 12 months of clean job data in a morning. That test is more useful than any formal readiness framework because it gives you a concrete answer fast. If the data does not exist in digital form, or exists but is spread across systems that cannot be queried together, the AI deployment will stall. The sequence matters more than the tool selection.
McKinsey’s State of AI research found that only about a third of organisations have begun to scale their AI programmes. Smaller firms scale at roughly half the rate of larger ones, and data readiness is a significant part of the explanation. Larger firms have spent years building the digital infrastructure that makes AI viable. Many owner-managed manufacturers are at the start of that process.
The OECD’s research on AI adoption in smaller businesses flagged this divide directly. Firms that are scaling AI have typically already digitised their core operational data. Firms stuck in pilot phase frequently have not.
Digitisation at this stage means capturing job records digitally as they happen, logging machine sensor readings centrally, and keeping maintenance history in a queryable form, not commissioning a full ERP system. That is the foundation the AI tools need, and getting it in place before selecting a tool is almost always the faster path to a working system.
What does a realistic first phase look like?
A realistic first phase in manufacturing AI is two to three months of data work before a single AI tool goes live. That means identifying which data exists and where, getting it into a digital format, connecting or replacing the systems that have never communicated with each other, and running a basic quality check before anyone purchases an AI product.
This is the conversation nobody wants to have with the board. They approved an AI scheduling tool. They want to see it running. A delegate who turns up in month two with a report on data infrastructure has to reframe what progress looks like.
The framing that tends to land is that data preparation is the foundation of the AI project, and every AI use case that follows will depend on it. A firm that skips this phase and goes straight to the AI tool will likely spend six months discovering why the outputs cannot be trusted, then have to do the data work anyway. At that point the board’s confidence in the programme is already damaged.
Goldman Sachs’ research on AI adoption in smaller businesses found that the biggest constraint for many owner-managed firms is internal capability to deploy AI well, not access to tools. Data readiness is foundational to that capability. Getting it right in the first phase protects the investment and shortens the time to a working system.
A practical starting point is a one-page data audit. Identify what data the AI tool will need, where it currently lives, and what the gap looks like. That audit takes a week. What it surfaces may take two to three months to address, but the work is scoped and the board can see what they are funding. Book a conversation to work through where the data gaps sit and what the first phase looks like for your operation.
