AI for Manufacturing: A Practical Guide for UK Manufacturers

If you read the trade press, you would think every factory in Britain is now run by AI. The reality on the shop floor is more sober. Make UK reports that only 2% of UK manufacturers say AI is widely embedded across their operations, and that where AI is used at all, it is mostly in HR, finance and admin rather than in production, quality or supply chain. The opportunity in industrial AI is real, but it sits behind some stubborn obstacles: legacy machines, fragmented data and a skills gap that no algorithm fixes on its own.

This guide is written for owners, operations directors and plant managers at UK manufacturers who want a grounded view of where AI actually pays off, where it stops, and what the law now expects of you. It is not a developer tutorial. It is the business case, the limits and the UK rules you need to know before you connect anything to a production line. If you want the wider context first, our pillar on Claude AI agents for business covers how these systems work across sectors.

Where AI earns its keep in manufacturing

The strongest use cases are the ones where data is reasonably clean and the AI informs a decision rather than taking control of a machine. In practice that means starting where the risk is lowest and the payback is clearest.

• Predictive maintenance on production assets. Models read vibration, temperature, current draw and runtime from sensors and PLCs to flag bearing wear or a failing motor before it stops the line. The business value is fewer unplanned stoppages and lower maintenance cost. Fluke's 2026 maintenance survey found UK predictive-maintenance adoption more than doubled, from 9% to 22%, while reactive maintenance fell from 42% to 26% year on year.
• Quality control and machine vision. Cameras plus computer-vision models inspect parts at full line speed for defects, surface flaws and dimensional drift, moving human inspectors to exception handling. This is manufacturing's most mature AI application across automotive, electronics, metals and food. The value is lower defect-escape rates, fewer warranty claims and less scrap and rework.
• Demand forecasting. Machine-learning models combine historical sales, order patterns, seasonality and external signals to sharpen forecasts that feed stock and procurement, reducing overstocking and stockouts and freeing working capital. The honest caveat is that a forecast is only as good as the often fragmented ERP and order data behind it.
• Production scheduling and planning. Optimisation and ML tools sequence jobs across machines and shifts against changeover times, due dates, material availability and labour, replacing spreadsheet planning. The value is higher throughput and better on-time delivery, with planners reviewing proposed plans rather than building them by hand.
• OEE and downtime analysis. AI aggregates machine state, cycle time and stoppage data to attribute losses across availability, performance and quality, and to surface the real root causes of downtime, turning improvement effort into something targeted and measurable rather than guesswork.
• Supply planning and inventory optimisation. Linking demand signals to supplier lead times, safety-stock policy and bills of materials lets AI recommend reorder points and flag supply risk. Given post-Brexit and global supply volatility, the value is resilience against shortages alongside lower tied-up capital.
• Back-office automation. Generative AI and document extraction handle purchase orders, invoices, RFQ responses and supplier emails. Make UK's data shows this is where UK manufacturers actually use AI most, because the time savings are real and the data is less sensitive than anything on the shop floor.
• Shop-floor knowledge assistants. A retrieval-based assistant over maintenance manuals, SOPs, machine logs and historical fault records lets technicians ask questions in plain language and get sourced answers. That speeds up fault diagnosis and captures the knowledge of retiring skilled staff, addressing the capability gap Make UK identifies.

Where AI stops, and why that matters in manufacturing

Industrial AI fails quietly when the groundwork is missing. Before you commit a budget, it helps to be clear-eyed about the limits.

• Core production adoption is still shallow. Make UK reports AI use of roughly 11% in production, 7% in supply chain and logistics, and 6% in quality control, against around 83% in HR, finance and admin. Much shop-floor AI remains at pilot stage rather than running at production scale.
• Legacy and fragmented data is the hard ceiling. Make UK found around 65% of manufacturers still rely on older systems that are incompatible with modern AI, and about 47% cite data fragmentation across OT, legacy and ERP systems as a major obstacle. A model trained on incomplete or inconsistent data produces unreliable output, so data quality is the binding constraint, not the algorithm.
• Skills, not funding, is the main barrier. Make UK reports that over 50% of manufacturers name skills shortages as the primary obstacle, particularly at technician and operator level, and the Fluke survey reaches the same conclusion. Deployments stall without people who can validate, maintain and act on model outputs.
• Safety-critical work needs human oversight. AI is not yet trusted to run safety functions on its own. The HSE requires validation of AI in critical roles, and the Government's machinery response mandates certification and hard design limits for machine learning in safety functions. A compromised model can have direct physical consequences, so human oversight and OT security stay mandatory.
• IT and OT integration widens the attack surface. The NCSC notes that threat actors increasingly target OT directly and that IT disruptions cascade into operations. Modbus alone accounted for 57% of OT cyber-attacks in 2025, up from 40% the year before. Integrating across OT, legacy and ERP is slow, costly and security-sensitive, and is frequently underestimated in AI business cases.

The rules: manufacturing-specific obligations

The UK has no single AI Act in force as of 2026. Instead it takes a sector-led, pro-innovation approach, with existing regulators applying existing law. For manufacturers that means several bodies and several rules sit over any AI you deploy. This is the part most generic AI advice skips, and it is where the real risk lives.

Health and safety law (HSE)

The HSE's January 2026 position is that AI is not a separate category of risk: it sits within existing health and safety law under the Health and Safety at Work etc. Act 1974. Employers must carry out a suitable and sufficient risk assessment of any AI that affects safety, and reduce risks so far as is reasonably practicable. The HSE states specifically that AI used for predictive maintenance must be validated so it does not overlook critical faults, and that robotics must be assessed for collision and human-interaction hazards.

Machinery safety law

The UK Government's 25 February 2026 response to the Machinery Call for Evidence confirmed it will update the Supply of Machinery (Safety) Regulations 2008 with measures mirroring the EU Machinery Regulation. New requirements include designing machinery so that cyber-attacks and unauthorised interference cannot create a hazardous situation, treating malicious third-party action as reasonably foreseeable misuse, providing security updates throughout the machinery's expected lifetime, and retaining technical documentation for 10 years after a product is placed on the market.

Critically, machinery that uses machine learning for safety functions will require mandatory Notified Body certification, and such systems must be designed to keep the machine's safety characteristics within acceptable, defined limits even as the model adapts. Note the territorial split: Northern Ireland applies the EU Machinery Regulation from October 2026, while Great Britain takes the adapted Supply of Machinery (Safety) Regulations route.

Data protection and automated decisions (ICO)

Any AI that processes personal data, including vision systems that capture workers, HR analytics and customer data in back-office tools, must comply with UK GDPR and the ICO's guidance on AI and data protection, covering fairness, lawful basis, transparency and data minimisation. The rules on automated decision-making changed recently: from 5 February 2026, UK GDPR Article 22 was replaced by new Articles 22A to 22D under section 80 of the Data (Use and Access) Act 2025, moving from a general prohibition to a right-of-challenge model with safeguards. Where AI makes solely automated decisions with significant effects on people, most relevant in recruitment, workforce and performance contexts, individuals must receive meaningful information about the logic, meaningful human intervention and the right to contest. The ICO closed a 2026 consultation finding most UK employers using automated recruitment tools were non-compliant, often because their "decision-support" tools had no meaningful human involvement in practice. The ICO is one of the UK's lead AI regulators alongside Ofcom and the FCA.

Operational technology security (NCSC)

In September 2025 the NCSC, with CISA, the FBI and international partners, published "Secure Connectivity Principles for Operational Technology" and joint guidance on building a definitive asset inventory. These set the expected baseline for connecting AI and analytics to OT networks, including migrating from legacy industrial protocols to secure versions, for example Modbus to Modbus Security and OPC DA to OPC UA, and segmenting IT from OT. This is directly relevant the moment you wire AI into shop-floor systems.

Off-the-shelf AI or a custom agent?

Off-the-shelf tools fit the lower-risk, standardised problems first: machine-vision inspection platforms, vendor predictive-maintenance modules, ERP-native forecasting and scheduling add-ons, and generative AI for back-office admin. They are fastest to value where the data is already clean and the use case is not safety-critical, which is exactly where UK manufacturers use AI today.

Their weakness is the messy reality of a working plant: data fragmented across OT, legacy machines and ERP, plant-specific processes, and the IT and OT integration and security work that generic tools quietly assume someone else has done. A custom industrial agent earns its place when value depends on unifying those siloed sources, encoding your own scheduling and quality logic, and sitting safely alongside OT under NCSC connectivity principles and HSE and machinery safety obligations. It should never be given autonomous control of safety functions, because machine learning in safety-related control systems triggers Notified Body certification and must operate within designed limits. The practical pattern is AI that informs and recommends while validated controls and human oversight stay in charge.

For most mid-market UK manufacturers the honest answer is hybrid: buy the commodity capabilities, build a custom layer for the integration and the firm-specific decisions, and treat data quality and OT security as prerequisites rather than afterthoughts. The same logic applies in adjacent operations, as we set out in AI for logistics and supply chain.

How to start

The manufacturers who get value do not begin with a moonshot. They begin with one bounded task and a tight feedback loop.

• Pick one task with clean data. Choose a use case where the data already exists and the decision is not safety-critical, such as demand forecasting, back-office document handling or a knowledge assistant. Resist the urge to start on the production line.
• Govern it before you build. Decide who owns the model, what it is allowed to decide, and where a human signs off. If personal data is involved, confirm your lawful basis and the ADM safeguards now. If a machine safety function is involved, that is a certification question, not a pilot.
• Pilot small and measure. Run a contained pilot against a baseline you can defend, whether that is forecast accuracy, OEE or hours saved in the back office. Measure against real numbers, not vendor claims.
• Keep a human in the loop. Validate the output, especially for predictive maintenance where a missed fault matters. Plan for the skills to maintain and challenge the model, because that is the barrier most likely to stall you.
• Then scale. Once one use case is proven and governed, extend it. Treat OT security and data quality as the foundation you build on, not a clean-up job for later.

What it costs

SpotDev works to fixed packages so you know the cost before you commit. We start with an AI and Data Readiness Assessment at £5,000, which is where the data-quality and OT-integration realities in this guide get tested against your actual plant rather than assumed. Delivery then runs from £8,000 to £45,000 depending on scope, and a first rollout is typically live in two to three weeks. We are a UK consultancy specialising in Anthropic's Claude, with in-house engineers and more than 300 technology projects delivered. If you want to pressure-test a use case before spending anything, talk to a Claude-specialist engineer.

Frequently asked questions

Is AI in UK manufacturing actually delivering results yet, or is it hype?

Both things are true. Real value is being delivered in specific places, particularly machine vision for quality, predictive maintenance and back-office automation, where Fluke reports predictive-maintenance adoption more than doubled from 9% to 22%. But Make UK reports only 2% of manufacturers have AI widely embedded, and most shop-floor use is still at pilot stage. The hype is in the breadth of claims, not in whether AI can work at all.

Can AI control our machinery and safety systems automatically?

Not without serious obligations. The HSE requires AI in safety-critical roles to be validated, and the Government's February 2026 machinery response mandates Notified Body certification for machine learning in safety functions, plus design limits the model cannot exceed. The safe and common pattern is AI that informs and recommends while validated controls and human oversight stay in charge.

Why do AI projects stall in manufacturing?

Usually data and skills rather than technology. Make UK found around 65% of manufacturers still rely on legacy systems incompatible with modern AI and about 47% cite data fragmentation as a major obstacle, while over 50% name skills shortages as the primary barrier. A model trained on incomplete data is unreliable, and without people to validate and act on its output, even a good model goes unused.

What do we need to secure before connecting AI to the shop floor?

Treat OT security as a prerequisite. The NCSC's 2025 Secure Connectivity Principles for Operational Technology set the baseline: segment IT from OT, build a definitive asset inventory, and migrate from legacy protocols to secure versions such as Modbus Security and OPC UA. This matters because Modbus accounted for 57% of OT cyber-attacks in 2025, and a compromised model connected to a machine can have physical consequences.

Work with a Claude specialist

SpotDev designs, builds and deploys custom Claude agents and enterprise Claude rollouts for UK businesses, with fixed packages from £8,000 to £45,000 and a first rollout live in two to three weeks. Explore our Claude implementation packages or talk to one of our engineers.