AI in back office
0%"
HR, finance and admin. The work already built to be queried.
UK Manufacturing · Data & AI Report 2026 · Catalyst BI
The ConnectedFactory
You are already using AI. The problem is it cannot see across your sites and factories, and that disconnect is where margin leaks and the numbers stop reconciling.
AI in production
0%
The part of the business that makes them a manufacturer.
AI readiness, of 10
0%
Manufacturing, last of five industries measured.
00
- The Argument
ADOPTION WAS NEVER THE PROBLEM
Ask whether you use AI, and the answer is yes. Ask where you use it, and the conversation changes.
In June 2026, Make UK asked its members where artificial intelligence is actually in use. The finding was not that manufacturers are behind. It was that manufacturers have deployed AI comprehensively in the parts of the business that look like every other business, and barely at all in the part that makes them a manufacturer.
Figure 1
Where UK manufacturers use AI
Share of manufacturers reporting AI in use, by function. The back-office lead is shown in green.
HR, finance and admin
83%
Production
11%
Supply chain and logistics
7%
Quality control
6%
Widely embedded across operations
2%
Source: Make UK, AI, Skills and the Future of the UK Manufacturing Sector, June 2026. Sample size and fieldwork dates not disclosed by the publisher.
Read that as a technology-adoption problem and you will fund more pilots. Read it as a data problem and it explains itself.
HR, finance and admin data already sits in systems built to be queried. It is structured, centralised, consistently named and owned by someone. Production data is none of those things. It sits in historians, SCADA systems, building management systems and an MES, at high frequency, named by whoever commissioned the line, in units nobody standardised, with no shared definition of what counts as a stoppage.
So the back office got AI because the back office had already done the data work, twenty years ago, when it bought an ERP. The factory floor has not. That is the whole argument of this report.
// The evidence says manufacturers are close, not far
This is not a story about an industry that cannot do it. Internationally, manufacturers are the sector most focused on operational efficiency from AI, and the most successful at achieving it. They are also, by some distance, the worst at proving the financial return.
90%
of manufacturers using generative AI report benefits in operational efficiency, the highest reading of any sector.
Omdia / Informa TechTarget, 2025
38%
mean quantified return on generative AI in manufacturing, the lowest of any industry against a 49% average.
Omdia / Informa TechTarget, 2025
5.8
manufacturing's AI readiness score out of 10, last of five industries measured.
Fivetran, May 2025
Getting the operational benefit and failing to prove the financial one is a measurement failure, not a capability failure. It happens when the improvement is real but nobody can trace it back through the data to a number finance will accept.
The pattern is not subtle. In a 2026 survey of UK food and beverage manufacturers, only 1% said they use the operational data they already collect very effectively (Barclays). Drowning in data and starved of insight is not a food and beverage problem. It is the problem this report is about.
// The UK has already run this experiment
And the evaluation is public, which is rarer than it should be.
The Made Smarter Adoption Programme spent £19.8 million between 2022 and 2025 helping manufacturing SMEs across five English regions adopt industrial digital technology. On its own terms it worked. More than 1,400 businesses built digital roadmaps, 60% above target, and 78% of the firms it supported went on to adopt industrial digital technology.
Ipsos UK then evaluated the programme for the Department for Business and Trade. The final impact evaluation was published in July 2026. Its more robust econometric analysis found no statistically significant effect on turnover, employment or productivity.
That is a government evaluation of a government programme, conceding that adoption did not convert into business results. It is the most honest account of this problem in print, and it is the reason this report is about foundations rather than technology.
£19.8m
Spent. No measurable return.
A government programme helped more than 1,400 manufacturers adopt digital technology. Its own independent evaluation found no statistically significant effect on turnover, employment or productivity. Buying the tools was never the hard part.
Source, Ipsos UK for DBT · Made Smarter Final Impact Evaluation · July 2026
Buying the tools was never the hard part.
Three chapters follow, each one a different instance of the same underlying problem. A cost you cannot allocate. A model that cannot move site. A figure you cannot reconcile. In each case the technology works and the data underneath does not carry enough shared meaning to prove it.
01
Chapter 01 · Energy
THE COST YOU CANNOT SEE
You know exactly what you spent on energy last month. You almost certainly cannot say what it cost to make a given product on a given line.
That gap matters more than it sounds. The site total is a number you can only react to. Energy cost per product, per process, per batch is a number you can act on, because it sits inside decisions you are already making about pricing, product mix and where to run what.
The data usually exists. It sits in the historian, the SCADA system, the building management system and the sub-meters. It just never meets the production order, so nobody can join the kilowatt hours to the thing being made.
// Why the site total stops being useful
Because it answers a question nobody needs answered twice.
You already know the bill went up. What you cannot see is whether it went up because of volume, product mix, a drifting compressor, a shift pattern, or one line running hot since a changeover in March. The monthly total treats all of those as the same event.
The reporting cycle compounds it. Energy is booked as an overhead, allocated across output on a flat rate, and lands in a finance pack weeks after anyone could have done something about it. By then the batch has shipped and the price is set.
Allocate energy as a flat overhead and every product looks equally efficient. None of them are. There is a fuller treatment of this in costing energy by product rather than by site.
// How much this is actually worth
Worth being straight about the numbers, because the honest version is more useful than the inflated one.
The Carbon Trust puts typical savings from monitoring and targeting at around 5% of energy spend. What energy costs you as a share of output depends heavily on what you make.
Table 1
How much this is actually worth
Electricity and gas as a share of gross output, by energy-intensive sub-sector. The highest-intensity sub-sector is shown in green.
| Sub-sector | Electricity and gas as share of gross output |
|---|---|
| Industrial gases, inorganics, fertilisers | 11.07% |
| Petrochemicals | 3.68% |
| Paper and paper products | 3.25% |
| Glass, ceramics and stone | 3.24% |
| Basic iron and steel | 3.08% |
| Other basic metals and casting | 2.71% |
Source: ONS, The impact of higher energy costs on UK businesses, May 2025. Intermediate consumption, 2017 to 2019 average.
Do that arithmetic and the direct saving on the bill is real but modest. If that were the whole case, it would be hard to justify the work.
It is not the whole case. The larger number is the one you are currently mis-allocating. When energy is spread evenly across output, high-intensity products are subsidised by low-intensity ones on every invoice you send. You are not losing money on energy. You are losing margin on the products you price as though they were cheaper to make than they are.
That is a mix and pricing question, and it belongs to the CFO rather than the energy manager.
// What changes once you can cost energy by product
Four decisions get better, and none of them are about energy.
-
Pricing and quoting. A quote built on true cost to produce rather than average cost to produce. On energy-intensive lines, that difference is large enough to change whether you should win the work.
-
Product mix. You find the lines that look profitable on paper and are not once their real consumption is attached. Some are worth repricing. Some are worth declining.
-
Maintenance priority. Rising energy per unit on a single asset is one of the earliest signals that something is degrading, often ahead of vibration or temperature thresholds.
-
Capital cases. An efficiency investment argued on product margin lands very differently in a board pack from one argued on utility spend.
None of that needs new hardware on the shop floor. The meters are usually already there. What is missing is the join.
// Why you probably do not have this already
Because the join is genuinely hard, and it is a data problem rather than an engineering one.
Consumption data lives at plant level in historians, SCADA and building management systems, typically at high frequency and long retention. Production data lives in the ERP and MES, organised by order, batch and product code. The two were never designed to meet.
Then multiply that by sites. Tag naming conventions differ between plants. Units differ. One site logs at fifteen-minute intervals, another at one. A compressor is called three different things in three historians. Before you can compare energy per unit across a group, somebody has to decide what the standard looks like and apply it everywhere.
This is why energy dashboards so often stall after the first site. The pilot works because one engineer knew that plant's tags. The rollout does not, because nobody standardised them.
Internationally, tesa built a single source of truth across roughly 7,000 products, giving it quality and efficiency insight at product level rather than site level. That is the shape of the outcome: not a better energy report, but production data granular enough to make commercial decisions from. Bonfiglioli took a similar route to a cloud data platform and got faster data loading and extraction as a result. Neither started with an energy project. Both started with fragmented operational data, and product-level visibility fell out of fixing it.
02
Chapter 02 · Predictive maintenance
WHY THE PILOT DOES NOT TRAVEL
Your predictive maintenance model worked at one plant. You moved it to the second, and it stopped. This is the most common way manufacturing AI stalls, and it is rarely a model problem.
It is a semantics problem. The model learned one plant's tag names, one plant's units, one plant's idea of what counts as a stoppage. None of that travelled.
You will be tempted to read that as a technical setback and fund a second build. It is cheaper to read it as evidence that your operational data model was never standardised in the first place.
// Why pilots succeed and then stop
Because the pilot is usually carried by one person's knowledge rather than by the data.
Someone at that plant knows which historian tag corresponds to which motor, that the temperature sensor on line three reads in Fahrenheit for historical reasons, and that a stoppage under two minutes is not logged as a stoppage. That knowledge makes the pilot work. It is also why the pilot is not a template.
Move to the second site and the asset hierarchy is different. Tags are named by a different convention, sometimes by a different contractor. Units differ. Sampling intervals differ. The MES defines a batch one way and the ERP another.
So your project team does the only thing available to it, which is to rebuild. The second model costs nearly what the first one did. By the third site, the programme is quietly reframed as a series of local initiatives, and the scale case that justified the investment has gone.
82% of manufacturing executives cited data-related problems as a critical factor that could jeopardise their company's future AI success. MIT Technology Review Insights, survey of 600 CIOs, February 2023
This is not a niche observation. The Government's AI Champion for Advanced Manufacturing, writing in the AI Adoption Plan published on GOV.UK in June 2026, put it plainly: too many promising AI projects remain trapped in pilot phases rather than delivering sustained impact. The barriers named are legacy systems, uncertain return on investment, limited access to test environments, workforce capability gaps, and integration and safety concerns.
The plan also makes a point worth holding on to. Industrial AI is harder to roll out than most other kinds of AI, because the environments are complex, highly engineered and often safety-critical. A model that is 90% right is a useful assistant in a marketing team and an unacceptable one on a production line. That raises the bar on the data underneath it, not lowers it.
// What AI-ready actually means on a factory floor
Not more data. You almost certainly hold far more than you use already, at high frequency and long retention.
AI-ready means the data carries enough shared meaning to be used somewhere other than where it was created. In practice that is five things.
-
A consistent asset hierarchy. Site, area, line, cell, asset, named the same way everywhere, so a model knows what it is looking at without a human translating.
-
Agreed tag naming and units. Applied on ingest rather than patched in the dashboard, because every downstream fix has to be repeated by everyone downstream.
-
Shared event definitions. What counts as a stoppage, a changeover, a reject, a good unit. These are operational decisions, not technical ones, and they belong to operations.
-
Production context attached. Sensor readings joined to order, batch, product and shift. Without that, a model can see that something changed but not what was being made when it did.
-
Lineage you can show. When a model recommends deferring maintenance on a critical asset, somebody will ask where the number came from.
None of that is exotic. All of it is unglamorous, and it is the reason the second plant does or does not work. We have written separately on what AI-ready actually means on a factory floor.
// The honest return on predictive maintenance
Worth being precise here, because this is where business cases get built on figures that will not survive a challenge.
Deloitte's published ranges are the ones to plan against: a 10 to 20% increase in uptime, a 5 to 10% reduction in maintenance cost, and a 20 to 50% reduction in maintenance planning time. In Deloitte's own case material, one manufacturer cut unplanned downtime by half. That is the ceiling rather than the expectation, and it is worth knowing which of the two you are putting in a business case.
The planning-time figure is the one that tends to get overlooked, and it is often the easiest to realise first. Maintenance teams spend a substantial part of their week assembling information rather than acting on it. That cost is immediate, visible to the people doing the work, and does not depend on a model reaching production.
// Why fixing the data model costs less than rebuilding per site
Because the cost of a standard is paid once, and the cost of not having one is paid at every site, every time.
Think about what the second build actually repeats. Locating and interpreting tags. Agreeing what an event means. Reconciling units. Re-establishing the join to production records. Re-validating with the plant team. The model itself is a small part of that effort.
Siemens is a useful reference point. It now runs more than 600 projects on one governed platform, and provisions new data projects in minutes rather than weeks.
600+
projects on one governed platform.
The number that matters is not 600. It is that the six hundred and first project is cheap, because the standard already exists and nobody has to negotiate it again. New projects are provisioned in minutes rather than weeks.
Siemens, published case study
03
Chapter 03 · OTIF & compliance
THE NUMBERS YOU CANNOT DEFEND
Two situations, one underlying failure. A customer's scorecard that does not match yours, and a regulator asking for evidence you cannot assemble. In both cases the figure exists. What is missing is the ability to reconcile it.
// Your OTIF says 96%. Your customer says 88
Nobody in the room can reconcile the two numbers, so the conversation moves on to improvement plans based on a figure neither side fully trusts.
This is not dishonesty. You are measuring dispatch and they are measuring receipt, and nothing joins the two. On time and in full are assessed together with no partial credit, so an order that arrives complete but a day late fails, and an order that arrives on the hour with one line short fails. Between your loading bay and their goods-in sit the carrier, the delivery window, the booking slot and the receiving process. Any of them can turn a success into a failure without anything going wrong at your end.
The uncomfortable part is what follows. Their number governs the commercial relationship. Most large customers set the bar at around 95% and apply penalties below it. If you cannot reconcile line by line, you cannot dispute it, and you cannot fix what is causing it either.
And there is almost always a pattern to find. One depot with a tight booking window. One product that ships short because of a pick-face issue. One carrier on one lane. One order cut-off time that leaves no slack when a line runs late. Those are fixable. They stay invisible while the argument is about the headline percentage.
// Why the evidence cannot be assembled
Because it is spread across five systems and one of them belongs to somebody else. The order sits in the ERP. The dispatch record sits in the warehouse system. Proof of delivery sits with the carrier. The receipt sits in the customer's system. The scorecard arrives as a PDF or a portal download, monthly, in a format nobody can join to anything.
Reconciling those by hand for a single disputed month is a week of somebody's time. So it does not get done. The longer version of this argument, including how to run the reconciliation, is in why your OTIF and your customer's never agree.
The tooling explains a lot of this. In a March 2026 survey of 400 inventory and operations professionals, 84.8% said spreadsheets were their primary inventory tracking tool. That includes 53% of companies with 500 or more employees.
// OTIF and capacity are the same problem
Most OTIF failures were decided weeks earlier, when somebody promised a date.
A late delivery is usually not a logistics failure. It is a planning failure that reached the loading bay. The order was accepted against capacity that was already committed, or against stock that assumed a forecast which did not hold, or into a week that looked fine until a changeover ran long. That is why chasing OTIF through the warehouse rarely moves it far. The warehouse is where the failure becomes visible, not where it was created.
McKinsey puts the reduction in forecasting error from AI-driven planning on connected data at 20 to 50%. The honest version of that benefit is not a better forecast. It is fewer promises you cannot keep.
// The same failure, now with a statutory deadline
From 1st January 2027, UK importers of aluminium, cement, fertilisers, hydrogen, and iron and steel will pay a carbon charge on the emissions embodied in those goods. You become liable once imports of covered goods pass £50,000 over a rolling twelve months.
Live from
1 Jan 2027
Threshold
£50,000 rolling 12 months, or 30 days ahead
In scope
5 sectors aluminium, cement, fertilisers, hydrogen, iron & steel
The hard part
Evidence emissions per consignment, quarterly
The charge is not the difficult part. The evidence is. You can declare the actual emissions attached to each consignment, or fall back on default values based on global production-weighted averages. If your suppliers are cleaner than the global average, the default overstates what you owe, and you cannot recover the difference later without the records to prove it.
The unit of account is the consignment, not the company. You need emissions data by supplier, by product and by shipment, reconciled against your purchase and customs records. Most manufacturers hold those four things in four places that do not agree with each other.
Annual Scope 3 reporting tolerates estimates. A quarterly tax return does not. That is the shift, and it is the OTIF problem again with a legal deadline attached. What UK CBAM requires you to evidence sets out the five things to have in place before January.
Nor is CBAM alone. Three separate obligations now require product and supply chain data at a granularity most manufacturers cannot produce.
Table 2
Three obligations, one data problem
New UK reporting duties that require product and supply-chain data at a granularity most manufacturers cannot yet produce.
| From | Obligation | Data required |
|---|---|---|
| Apr 2026 | Packaging EPR, annual reporting | Material, weight, class, and which UK nation packaging is supplied and discarded in |
| Jan 2027 | UK CBAM | Direct and indirect emissions per consignment, quarterly |
| Jan 2027 | UK SRS S1 and S2, proposed for listed companies | Scope 1, 2 and 3 greenhouse gas reporting |
UK SRS timing is subject to FCA consultation, with final rules expected autumn 2026. Full sources listed at the end of the report.
// What good looks like
Valpak is a useful British example, and an unusual one, because compliance data is its product rather than its overhead.
Valpak is an environmental compliance and data services company, part of Reconomy Group, helping organisations navigate packaging and producer responsibility regulations across the UK, EU and US. It has run on Qlik for more than twelve years. Its Insight Platform ingests and transforms millions of rows daily, delivering single-tenant environments across regions so that data residency requirements are met and retailers can trust the figures.
The outcomes are commercial rather than technical. Onboarding a new customer's data went from weeks to days. The platform has become Valpak's primary shop window and has accelerated client growth by 40%. Confidence in it is such that Valpak has increased its own investment in the underlying platform by 200%.
40%
acceleration in client growth, with the data platform as the primary shop window.
Valpak, 2026
200%
increase in Valpak's own investment in the platform underneath it.
Valpak, 2026
12yrs
of continuous build on one governed platform, now extending into agentic AI.
Valpak, 2026
Underneath all of it, data quality is what protects the business. Valpak's customers use its figures to calculate packaging tax liabilities, so an error at scale is not a reporting inconvenience. It is a tax exposure running to millions of pounds. Auditability across large volumes is what prevents it.
Valpak is now extending into agentic AI with PackChat, built on Qlik Answers as an early adopter, which lets customers ask natural-language questions of their compliance and packaging data and get accuracy-validated answers. That is only possible because the underlying data model was already governed and consistent.
The lesson for a manufacturer is the order of operations rather than the tool. Valpak did not arrive at conversational AI over its compliance data by buying conversational AI. It arrived there after twelve years of keeping one data model clean enough to answer questions nobody had thought to ask yet.
Catalyst BI has supported Valpak's platform throughout.
04
Chapter 04 · The build order
WHAT TO FIX, IN WHAT ORDER
Each step makes the next one cheaper. That is the only reason the order matters.
- Standardise the asset hierarchy before the tags. Tags are easier to map once everyone agrees what an asset is and where it sits. Doing it the other way round means mapping twice.
- Agree event definitions with operations, not IT. If production and engineering do not accept the definition of a stoppage, no amount of governance will make the resulting number trusted.
- Apply the standard at ingest. Transforming on the way in means one correction rather than one per consumer.
- Land OT and IT data in the same place. A cloud data platform such as Snowflake, Databricks or BigQuery, fed by integration tooling such as Fivetran or Qlik Talend, with a transformation layer such as dbt modelling the joins so the logic is versioned, tested and reviewable. If your production and finance data sits in SAP, start with the diagnostic before the destination: are any of your pipelines dependent on ODP-RFC? That question decides your options and their timing: see how to check your SAP pipelines for an ODP-RFC dependency.
- Model the allocation rules explicitly. How you apportion base load between products, or whose timestamp counts on a delivery, is a judgement rather than a fact. Write it down, agree it with finance, and make it reviewable.
- Give the model an owner. Standards without an owner drift back to local variants within a year. This is a product with a roadmap, not a project with an end date.
// Where to start
Not with a new pilot. You almost certainly have one that already works.
Take the model running at your best-instrumented site and try to move it to the next one. Do not budget to rebuild it. Budget to find out precisely what stops it, and write that list down.
That list is your foundation backlog, and it will be more specific and more persuasive than any architecture review. It also gives you something rare in this kind of programme: a business case built on a failure you can demonstrate, rather than a benefit you have to promise.
The same logic applies to the other two chapters. For energy, take one line with decent metering and join its consumption to its production orders for a quarter, then look at the spread of cost per unit across the products it makes. That spread is the argument, and it is almost always wider than anyone expects. For OTIF, take your largest scorecard dispute and reconcile three months of order lines against the customer's view, one by one.
In each case the output is the same: a ranked list of specific, fixable causes, made out of your own data. That is a better business case than any maturity assessment, because nobody can argue with it.
05
Chapter 05 · The UK picture
THE CONTEXT, AND ONE FINDING THAT REFRAMES IT
The context this sits in, and one finding that should change how the investment case is framed.
Manufacturing accounts for 8.5% of UK economic output and around 48% of business research and development spend, across roughly 130,000 enterprises employing just under 2.5 million people.
The finding that ought to change the conversation is this: manufacturing labour productivity has grown at about 3.0% a year since 2000, ahead of the whole-economy average. Whatever is holding back UK productivity, manufacturing is not it. That makes the case for investment a growth case rather than a rescue case, and it should be argued that way in a board pack.
The pressures are nonetheless real and they are mostly about cost and capacity rather than demand.
Policy is moving in the same direction as this report's argument, which is worth knowing if you are building an investment case. The Government appointed an AI Champion for Advanced Manufacturing, and the adoption plan published in June 2026 proposes trusted data environments and SME adoption pathways among its five interventions. The Department for Business and Trade has earmarked £99 million for expanding Made Smarter Adoption from 2026, and £12 million for Data Sharing Infrastructure from April 2026. The direction of travel is towards funding the foundations rather than the applications.
The finding that ought to change the conversation is this: manufacturing labour productivity has grown at about 3.0% a year since 2000, ahead of the whole-economy average. Whatever is holding back UK productivity, manufacturing is not it. That makes the case for investment a growth case rather than a rescue case, and it should be argued that way in a board pack.
The pressures are nonetheless real and they are mostly about cost and capacity rather than demand.
Policy is moving in the same direction as this report's argument, which is worth knowing if you are building an investment case. The Government appointed an AI Champion for Advanced Manufacturing, and the adoption plan published in June 2026 proposes trusted data environments and SME adoption pathways among its five interventions. The Department for Business and Trade has earmarked £99 million for expanding Made Smarter Adoption from 2026, and £12 million for Data Sharing Infrastructure from April 2026. The direction of travel is towards funding the foundations rather than the applications.
Figure 2
AI readiness, two cuts of one survey
Composite readiness score out of 10. Manufacturing is last of five industries measured; the UK is last of four regions.
By industry
Healthcare
8.2
Retail
8.0
Technology
7.0
Finance
6.0
Manufacturing
5.8
By region
APAC
8.8
USA
8.2
EMEA
8.0
UK
6.0
Source: Fivetran, AI Readiness Index, May 2025.
// Energy
UK manufacturers paid 16.8 pence per kilowatt hour for electricity in the fourth quarter of 2025, down 7.9% year on year, and 3.5 pence for gas, down 11.6%. The relief is recent and the international gap is not closed: UK industrial electricity prices were the highest of all 24 International Energy Agency members in 2023, with UK firms paying around 50% more than German and French competitors.
The effect on energy-intensive sub-sectors has been severe. Between the first quarter of 2021 and the fourth quarter of 2024, gross value added fell 46.5% in basic metals and castings, 30.6% in inorganic non-metallic products and 30.2% in petrochemicals.
From April 2027 the British Industrial Competitiveness Scheme will cut electricity bills for over 10,000 manufacturers, worth around £600 million a year. Which means the organisations that can show where their energy actually goes will be best placed to model what that is worth to them.
// Skills
There were 47,000 manufacturing vacancies in June 2026. Skills shortages and gaps cost UK engineering and manufacturing an estimated £5.2 billion a year, around 2% of sector output. For the firms affected, the loss runs at roughly 10% of their own gross value added.
Over half of manufacturers name skills shortages as the main barrier to AI adoption, and half say staff have no time for training. This matters for how you sequence the work. A foundation programme that depends on hiring scarce data engineers will stall. One that reduces the manual effort those people currently spend reconciling figures pays for itself in capacity.
// Supply Chain
Supplier delivery times have lengthened for 31 consecutive months. In March 2026 they lengthened to the greatest extent in over four and a half years, with 25% of firms reporting longer lead times against 2% reporting shorter. 35% of UK and Ireland manufacturers now name supply chain disruption as a leading challenge, up from 31% the year before.
Add one more exposure that is easy to miss. 30% of manufacturers had a cyber incident in the past year, directly or through their supply chain. Among those affected, 31% lost production capacity and 31% had delivery delays.
06
The Connected Factory Benchmark
HOW CONNECTED IS YOUR FACTORY?
Ten questions about your own operation. You get a score out of ten and a sense of how you compare. Nothing is recorded. Most estates have ground to make up: only 1% of manufacturers say they use the data they already hold very effectively (Barclays, 2026).
0/10
Connected Factory Review
Find out what the disconnect is costing you
You are already using AI. The question is whether your data is connected enough across sites to show where the money leaks, and that is answerable in weeks rather than months.
The Catalyst BI Connected Factory Review maps your operational data against the foundations in this report, ranks the gaps by what they cost you, and gives you a costed plan you can take to a board. You would not be alone in looking for one: 35% of manufacturers say the single thing that would most accelerate their AI adoption is access to a proven, trusted partner (Barclays, 2026). We are the partner advanced manufacturers keep building with.
Common questions
The questions people ask
Short answers to the questions buyers, and AI search engines, ask about connecting manufacturing data.
QWhy do most manufacturing AI projects stall?
Because the operational data is not standardised, governed or connected across sites. A model built at one plant cannot travel to the next when every site names its assets, units and events differently. The technology usually works; the data underneath does not carry enough shared meaning to prove a return.
QWhere do UK manufacturers actually use AI today?
Mostly in the back office. In Make UK's June 2026 survey, 83% of manufacturers reported AI in HR, finance and admin, and only 11% in production, because back-office data was already structured and queryable while production data was not.
QWhat is a connected factory?
A manufacturing operation where machine, production, energy and ERP data are joined on one governed data platform, so figures reconcile across sites and AI can see the whole operation rather than a single line or plant.
QWhy is connecting the data the first step, not buying AI?
Because AI models, dashboards and reports all depend on the same connected, governed data underneath. Fix the foundation once and every use case, from energy cost per product to predictive maintenance to OTIF, has somewhere to stand. Skip it and each pilot becomes a bespoke build that does not scale.
QHow do we know how connected our factory is?
The Connected Factory Benchmark is ten questions about your own operation. You get a score out of ten and a sense of how you compare; UK manufacturing averages around six on the readiness measures in this report.
Sources & method
Every figure is attributed
Every figure in this report is attributed to a named organisation and a date. Where a source has a commercial interest in the answer, or where a sample is small or undisclosed, that is stated alongside the figure rather than buried.
- Make UK, AI, Skills and the Future of the UK Manufacturing Sector, June 2026. Sample size and fieldwork dates not disclosed by the publisher. Make UK is a trade body with a policy interest in this area.
- Omdia by Informa TechTarget, commissioned by Snowflake. Fieldwork 13th August to 17th September 2025, 2,050 qualified generative AI adopters at organisations of 500+ employees. The base is organisations with AI already in production, so return figures are subject to survivorship bias by design.
- Fivetran, AI and Data Readiness Report, May 2025. 401 data leaders, fieldwork by Redpoint. Composite score of the publisher's own construction; weighting not disclosed.
- Barclays, AI adoption in food manufacturing, promotional feature published by Food Manufacture, 2026. Survey of 100 UK food and beverage manufacturers. Sponsored content, so treated as a food and beverage source with a commercial interest and cited as such; used for the data-effectiveness and partner-demand figures only.
- MIT Technology Review Insights, sponsored by Databricks, February 2023. 600 CIOs; manufacturing sub-sample size not stated.
- Ipsos UK for the Department for Business and Trade, Made Smarter Adoption Programme Final Impact Evaluation, report dated June 2025, published 30th July 2026. Wave 1: 160 interviews, 18% response. Wave 2: 273 interviews, 15% response. Econometric analysis of 2,402 Made Smarter businesses, 2017 to 2023. Open Government Licence. Note that DBT commissioned an evaluation of its own programme; the null finding is reported here because it runs against that interest, not with it.
- HVM Catapult, AI Adoption Plan: Advanced Manufacturing, authored by the Government's AI Champion for Advanced Manufacturing and published by DSIT and DBT on GOV.UK, 8th June 2026. Qualitative; no survey data is claimed. HVM Catapult is publicly funded and the plan proposes interventions it would deliver.
- ONS, The impact of higher energy costs on UK businesses, May 2025; UK business: activity, size and location, September 2025; workforce jobs series JWR7 and vacancies series JP9I, 2026.
- DESNZ, Quarterly Energy Prices, March 2026, reporting Q4 2025 data.
- House of Commons Library, Manufacturing industries: economic indicators, August 2026. House of Lords Library, Electricity prices in Great Britain, June 2026, citing Deloitte analysis of 2023 data.
- Cambridge Industrial Innovation Policy, citing ONS, March 2026, for manufacturing productivity growth and R&D share.
- Enginuity, Mind the Gap, analysis by SQW, June 2026. Sample size not disclosed; Enginuity is a sector skills body.
- S&P Global / CIPS UK Manufacturing PMI, March to August 2026. A diffusion index: it reports the balance of firms seeing longer versus shorter lead times, not the length of the delay.
- MHA, Manufacturing Report 2026. Over 1,000 C-suite respondents across the UK and Republic of Ireland, skewed towards firms above £250m turnover.
- Make UK, Cyber Security in Manufacturing, August 2026.
- Carbon Trust on monitoring and targeting; Deloitte on predictive maintenance ranges; McKinsey on forecasting error reduction.
- inFlow, State of Inventory Management 2026. Survey of 400 inventory and operations professionals, March 2026. Vendor-published; the sample skews towards smaller and non-manufacturing businesses, so the 500-plus employee cut is quoted rather than the headline.
- GOV.UK and ICAP on UK CBAM scope, threshold and coverage; GOV.UK on Packaging EPR and the British Industrial Competitiveness Scheme; DBT on UK Sustainability Reporting Standards.
- Valpak, figures supplied by Valpak and cleared for publication, September 2026. Supporting public material: the AWS, Qlik and Valpak video case study.
- Case examples: tesa, Bonfiglioli and Siemens, from published customer case studies.