The Industrial Revolution of the Information Factory

What Huddersfield’s textile heritage teaches us about artificial intelligence and the organisation of work

What changed at Armitage Bridge?

At Armitage Bridge, just outside Huddersfield, the buildings of a former textile mill tell a more interesting story about industrialisation than the familiar account of machines replacing hand work. The Brooke family’s mill complex developed in stages. Its early nineteenth-century buildings accommodated mechanised preparation and finishing alongside workshops for hand spinning and weaving. A further building, probably intended for machine spinning, followed in 1828–29, and a shed for power-loom weaving was added in 1838. Production was being brought together on one site before every operation was mechanised; the organisation of work and the technology used to perform it evolved together.

This sequence matters. It would be tempting to tell the story as a succession of inventions, each making a particular operation faster. The physical arrangement of Armitage Bridge suggests a more complicated process. Space, power, materials, workers and different forms of production had to be coordinated as the business changed. Mechanisation affected the work around the machines as well as the work performed by them. The surviving buildings provide evidence of that changing organisation; they do not, by themselves, tell us which decisions succeeded or how much time was lost between operations. Those are different historical questions, and we should resist inventing answers to make a contemporary analogy work.

The local history is distinctive in another respect. In the Huddersfield district, textile production had long been distributed among households, workshops, clothiers and merchants. The move towards larger mills in the valleys was gradual and uneven, rather than a clean transition from an unorganised past to an efficient factory system. The district’s engineering, dyeing and transport arrangements developed alongside textiles. Its industrial transformation involved markets, capital, labour and skills as well as machinery. [1–3]

There are obvious limits to comparing nineteenth-century mechanisation with artificial intelligence. AI does not depend on steam power or require the same capital investment, and there is no basis for assuming equivalent social or economic consequences. What interests me is a narrower comparison: when a new productive capability becomes available, the organisation must decide whether to insert it into existing work or reconsider how the work is organised. We have learned a great deal about that problem since the first mills were built.

The factory that exists alongside the factory

In manufacturing, we are accustomed to following material through operations. We look at where it waits, what changes it, what can go wrong and what determines whether a finished product reaches the customer. We are often less systematic about the work that turns a customer’s requirement into something the factory can manufacture.

An enquiry becomes a quotation; a quotation becomes an order; an order becomes an approved specification and production instructions. Information is checked, interpreted, copied, challenged and sometimes corrected. It passes between commercial staff, technical specialists, planners and production teams. A mistake in that sequence can emerge as a manufacturing problem much later, when correcting it is more expensive.

I find it useful to think about this as an information factory. I do not mean a separate department, or that every conversation and management decision can be reduced to a production line. The term is an analytical perspective: it directs attention to the work involved in producing information that other people need in order to act. The concept is not new. Ballou and colleagues modelled information-manufacturing systems in the 1990s, examining the quality, timeliness and cost of information products. Wang and Strong showed why information quality must be judged by its usefulness to those who consume it, not merely by whether individual data items are accurate. [4–5]

The analogy is useful precisely because it has limits. Information can be copied and reused without being consumed like material. A specification may change because a customer is learning what is possible, not because somebody made an error. A technical approval may involve a difficult judgement, rather than a routine step whose speed should be maximised. If we treat every return to an earlier stage as waste, we risk eliminating necessary inquiry. The objective is to understand which activities create value, which protect the business and which exist because the organisation has become unnecessarily complicated.

Information work also has a history. Paper files, telephone calls and handwritten records gave way, in different places and at different times, to spreadsheets, databases, enterprise software, shared documents and rules-based automation. Each generation changed what could be recorded and transmitted, but often left older practices in place. A business may have a sophisticated production system and still depend on someone searching an email thread to discover which version of a customer requirement was agreed. AI enters this accumulated landscape; it did not create the information factory.

A quotation that takes four days

Consider a hypothetical manufacturer of specialist woven fabric. A customer sends an enquiry specifying composition, weight, colour and finishing requirements, with a reference to an earlier order. Sales wants to issue a quotation promptly. Technical staff must establish whether the requirements are feasible, and production planning needs to understand material availability and capacity. The example is constructed to examine the reasoning; it is not a claim about a named textile business.

Suppose management believes that the technical department is responsible for slow quotations. The proposal is to use AI to read incoming enquiries and prepare technical assessments, reducing the time specialists spend drafting them. That might be useful, but before selecting it I would want to know what actually happens between receiving the enquiry and sending the quotation.

Imagine that a sample of completed enquiries produces the following illustrative findings. The median time from receipt to quotation is four working days. Across the sample, the time spent actively preparing technical assessments is less than an hour per enquiry, while many cases wait for customers to clarify missing performance or finishing requirements. Some are then returned for approval because the commercial and technical teams are working from different versions of the specification. These figures are invented for illustration; they are not evidence about textile manufacturers or the expected performance of AI.

The investigation has changed the question. Instead of asking how quickly AI can draft a technical assessment, we would ask why the organisation so often begins that assessment without sufficiently clear requirements. We would inspect a representative set of enquiries, compare the original requests with the approved specifications, check timestamps and ask employees to explain exceptions. We would also look for cases that progressed smoothly, because those may reveal practices worth retaining. If the records cannot distinguish customer waiting time from internal waiting time, we should acknowledge that limitation rather than manufacture a precise process map.

Several interventions become possible. The simplest might be to agree a clearer set of questions for incoming enquiries, with an explicit route for genuinely novel requirements. If the problem is inconsistent versions of specifications, better document ownership and a reliable approved record may matter more than faster drafting. A rules-based form could capture standard requirements without AI. AI might be helpful where customer correspondence is varied and extracting the relevant details is genuinely difficult, provided somebody verifies the output before it becomes a manufacturing instruction.

A fair comparison would include the cost of implementing and maintaining each option, the effort required to check exceptions, and the effects on quotation lead time, specification accuracy and customer experience. We might compare a short baseline sample with a subsequent sample of similar enquiries, looking at the proportion requiring clarification, time to an approved specification, avoidable revisions and the number of problems discovered after release. Changes in order mix and customer behaviour would complicate the comparison; a faster result alone would not prove that a particular intervention caused it. It would also consider whether faster quotations actually help the business: if production capacity or material availability prevents acceptance of more orders, speed alone may have little commercial value. This is how a view of the whole information process can change the choice of technology.

There is an equally important counterexample. If a stable, high-volume activity consists of transferring well-defined order details between systems, and its purpose and controls are already clear, straightforward automation may be justified immediately. A large discovery exercise could cost more than the problem warrants. Understanding work is not a ritual that must precede every experiment; the depth of investigation should reflect uncertainty, risk and the consequences of being wrong.

What AI changes, and what it does not

Conventional software has been automating information work for decades. Rules can validate required fields, route approvals, calculate prices and move records between systems. For predictable tasks with clear inputs and outputs, those approaches may remain cheaper, more reliable and easier to audit than a language model.

The newer opportunity is that AI can help interpret material that does not arrive in a neat structure: correspondence, specifications, meeting notes and technical documents. It may assist with finding relevant previous decisions, drafting preliminary assessments or comparing alternative interpretations of a requirement. It may also make some activities feasible that a small manufacturer would previously have considered too costly, such as examining a large collection of historical quality reports for recurring themes. Whether such analysis is useful depends on the completeness of the records and whether the proposed findings survive examination by people who understand the work.

Research helps clarify why task automation and organisational improvement must be distinguished. Amaya and Holweg studied algorithmic adoption in four areas of a multinational energy company. They identified different pathways: replacing work within a task and redesigning the process around it. The latter generally offered greater scope for improvement, but also depended on changes to established knowledge practices. That is evidence about the cases studied, not proof that every AI project should involve process redesign. [6]

A further complication is that AI creates work of its own. Somebody may have to maintain trustworthy source documents, check generated assessments, resolve ambiguous cases, investigate errors and decide who is accountable when a recommendation is wrong. If AI saves ten minutes in drafting but adds fifteen minutes of verification, the apparent improvement disappears. The relevant comparison is the complete process, including the work that is newly required and the risks that remain.

Changes to knowledge and responsibility also matter. Making previous technical decisions easier to retrieve could help a less experienced employee prepare an enquiry, but the organisation must decide which judgements still belong to a specialist. Experienced employees may be concerned, reasonably, about the reliability of the records or about being made accountable for work they did not perform. An effective design needs their involvement, not merely their acceptance after the system has been installed.

Can AI make it cheaper to understand work?

There is a second use of AI that interests me as much as its operational applications. A manufacturer deciding what to improve first needs a sufficiently reliable account of how the relevant work happens. That may involve speaking to employees, observing activities, reading documents, examining system records and testing explanations. For a small business, this investigation competes with urgent operational demands and can be difficult to resource.

AI may reduce some of the effort. It can organise interview notes, compare accounts, identify questions that have not been answered and help an investigator consider explanations other than the one they initially favoured. In the quotation example, it might point out that employees describe technical approval differently, or suggest examining whether missing customer information explains the delay better than specialist capacity. I describe this use as AI on the business: applying AI to the investigation and improvement of organisational work, rather than only using it in the business to perform that work.

The distinction is useful, but it is not a new theory of organisational improvement. It is an application of existing investigative practices using newer tools. Its potential advantage is economic: AI might lower the effort needed to organise evidence and explore hypotheses at the same time as it lowers the cost of experimenting with changes to information work. That would be valuable for smaller manufacturers, provided the understanding produced is dependable.

The qualification is important. A coherent summary of interviews is not a validated account of a process. Two employees may describe different procedures because their work genuinely differs, or because the process changes when an exception occurs. AI can conceal that disagreement by writing a smooth composite account. It may mistake frequent complaints for significant causes, invent links between events or reflect the investigator’s assumptions back to them. A process map can look convincing long before anyone has established that it is correct.

There is also an uncomfortable possibility: AI may reduce the cost of producing an attractive description without reducing the cost of finding out whether it is true. Observation, reconciliation of conflicting accounts and obtaining trustworthy records may still take most of the time. AI could even make an investigation slower if it generates numerous plausible but unhelpful questions that people then have to check. For that reason I would compare an AI-assisted investigation with a sensible non-AI alternative, including the hours spent gathering evidence, checking interpretations and correcting the final account, rather than comparing only the time needed to draft a process map. Research on AI-supported validation in process mining is emerging, but results from a particular prototype or research setting cannot establish a general productivity benefit for consultancy-style discovery in manufacturing SMEs. [7]

The right test is therefore the performance of the investigation itself. Did AI help us reach a defensible understanding with less total effort, including checking and correction? Did it expose an important alternative explanation that we might otherwise have missed? Did the eventual intervention improve the outcome the organisation cares about? If the answer is unknown, we should say so. The information factory used to investigate the information factory deserves the same scrutiny as any other process.

Keeping evidence separate from interpretation

In practical terms, I would begin with the evidence rather than a generated diagram. An employee’s account should remain identifiable as that person’s account; a timestamp from a business system is a different kind of evidence; and an AI-suggested explanation is a hypothesis until checked. Disagreements should be retained where they matter, and missing evidence should be recorded rather than silently supplied.

AI can then be given a deliberately critical role: identify assumptions, propose plausible alternatives and specify what observations would distinguish them. A suggestion that technical approvals cause delay is useful only if we can compare it with evidence about when enquiries became complete, when they entered the approval queue and when decisions were made. If an employee disputes the resulting account, the disagreement should lead us back to the work and the records, not to a prompt asking AI to make the description more consistent. Where the records are inadequate, further observation may be more valuable than another round of prompting.

This does not mean every small improvement requires a formal research exercise. The discipline can be proportionate. Sometimes a short conversation with the people doing the work and a review of a few real cases will be sufficient to reject an unnecessary automation. In other circumstances, particularly where a change affects production instructions, customer commitments or regulatory obligations, much stronger evidence and controls will be needed.

The next transformation of information work

The Armitage Bridge story offers a useful perspective because it shows that the organisation of production and its machinery did not change in a single step. Work was brought together, skills and arrangements evolved, and new equipment altered what the mill could do. We should not use that history to predict the scale or social consequences of AI, but it is a reminder that the productive capability of a technology is only part of the organisational question.

For manufacturers today, the information that connects a customer enquiry to a production instruction is as consequential as many of the physical operations that follow. Its quality, timing and interpretation affect cost, delivery, quality and trust. AI can help with parts of that work, and it may enable entirely different ways of organising it. Some improvements will be simple changes to information management; others will involve conventional automation or carefully controlled AI. The appropriate choice depends on what the organisation is trying to achieve and what the evidence reveals.

I am especially interested in the possibility that AI can support the investigation as well as the intervention. If it makes reliable organisational understanding more affordable, that could broaden access to systematic improvement for smaller businesses. We do not yet have sufficient evidence to treat that benefit as established, and there are good reasons to test it critically. Nevertheless, the question seems worth pursuing because it applies the same standard to AI that manufacturing has learned to apply to machinery: judge the change by what happens to the system, rather than by how impressive an individual operation appears.

Sources and further reading

1. Historic England, Armitage Bridge Mills, heritage record (site development, machinery and buildings). https://historicengland.org.uk/listing/heritage-gateway/results/record/0d636742-8028-4a8c-8b7c-e1f9aaf2d338/eb347450-eb3b-4114-8182-f31d0a50aa7e
2. David Griffiths (2014), Industrial Villages South of Huddersfield, Huddersfield Local History Society (local context for Armitage Bridge and the Brooke family). Read the article.
3. Historic England (2024), Textile Mills: Introduction to Heritage Assets, by Ian Miller (context for the evolution of textile-mill buildings, integration of spinning and weaving, and changing production arrangements; not a source for the site-specific Armitage Bridge dates). https://historicengland.org.uk/images-books/publications/iha-textile-mills/heag320-textile-mills/
4. Ballou, D. P., Wang, R. Y., Pazer, H. and Tayi, G. K. (1998), Modeling Information Manufacturing Systems to Determine Information Product Quality, Management Science, 44(4), 462–484. https://doi.org/10.1287/mnsc.44.4.462
5. Wang, R. Y. and Strong, D. M. (1996), Beyond Accuracy: What Data Quality Means to Data Consumers, Journal of Management Information Systems, 12(4), 5–33. https://doi.org/10.1080/07421222.1996.11518099
6. Amaya, J. and Holweg, M. (2024), Using Algorithms to Improve Knowledge Work, Journal of Operations Management, 70(3), 482–513. https://doi.org/10.1002/joom.1296
7. Dormehl, J. A., Andrews, R., Kratsch, W., Röglinger, M., Wynn, M. T. and Zetzsche, F. (2026), Domain experts in the loop: Leveraging generative artificial intelligence for interactive data validation in process mining, Information Systems, 140, 102715. DOI. This study concerns event-log validation in process mining; it does not establish general productivity gains for organisational discovery in manufacturing SMEs.