From Corporate Archives to Digital Twins: How AI Is Becoming an Operating Layer

Artificial intelligence is moving from a tool to an information layer

The first wave of generative artificial intelligence inside many companies looked familiar: a chat window, a drafting assistant or a faster way to summarise a document. Those uses can save time, but they do not yet change how an organisation understands itself. A more consequential shift begins when a system can work across the firm's own knowledge, connect evidence from different periods and reveal how decisions, people, processes and assets influence one another. At that point artificial intelligence is no longer an isolated productivity feature. It starts to become an operating layer between corporate data and managerial judgment.

An October 31, 2024 Economist Education article offered a compact view of that transition. The course excerpt drew on a podcast led by Tom Chatfield and presented ideas from four members of The Economist's editorial team. Their examples ranged from searching an archive of financial crises to modelling the structure of a company and building a more intuitive interface to information. The discussion did not promise a precise future. It described directions that business leaders can translate into practical design choices.

The business opportunity is not simply to generate more text. It is to reduce the distance between a question and the evidence needed to answer it, while preserving permissions, provenance and human responsibility. The challenge is equally important: a system that appears to know the organisation can be trusted too easily. Companies need to build a useful knowledge layer without turning confident software output into an unexamined version of corporate truth.

A financial archive shows the value of organisational memory

Arjun Ramani, then a global business and economics correspondent, described a leading hedge fund with a collection of historical documents about past financial crises. An analyst could ask a question and use artificial intelligence to search the archive for relevant cases. The source did not name the fund, disclose its location or publish a measured return from the system. The importance of the example lies in the pattern: valuable knowledge already existed, but finding the right precedent at the right moment had been difficult.

Most established organisations have a similar memory problem. Reports sit in shared drives, project spaces, email, contract systems and personal folders. Experienced employees remember why a decision was made, while a newer colleague sees only the final document. Search can locate matching words, but a business question often uses different language from the record that contains the answer. A useful knowledge system must connect concepts, dates, entities and decisions, then show the analyst where each claim came from.

Historical context is particularly valuable during an unfamiliar event. A current disruption will never perfectly reproduce an earlier crisis, but past cases can expose recurring questions: which assumptions failed, which indicators changed first, which interventions created unintended consequences and how quickly operations recovered. Artificial intelligence can shorten the work of assembling those cases. It cannot decide whether the analogy is valid. That remains a task for professionals who understand the market, the quality of the archive and the differences between periods.

Retrieval must preserve evidence rather than manufacture certainty

A corporate knowledge assistant should be designed as a research system, not an oracle. Its strongest answer is often a structured set of evidence: the relevant passages, the documents that contain them, their dates, authorship, status and any conflict between sources. A concise summary is useful only when the user can move back to the underlying material. Without that path, speed comes at the cost of auditability.

This distinction matters because internal information is not automatically correct. A draft may have been superseded. A sales presentation may express an ambition rather than a result. A policy may apply to one subsidiary and not another. A project review can reflect the incentives of its author. Connecting a language model to more files increases coverage, but it also increases the number of contradictions, outdated statements and uncertain definitions that the system must surface.

What a dependable corporate answer should reveal

• Source identity: the file, system, owner and version supporting each material claim.
• Time context: when the information was created, approved and last reviewed.
• Access context: why the user is permitted to see the evidence and which details remain restricted.
• Confidence limits: missing records, conflicting definitions and conclusions that require human interpretation.
• Decision boundary: whether the output is research, a recommendation or an authorised action.

These controls make the experience slightly less magical, but much more valuable. Managers do not need theatrical certainty. They need a faster route to defensible judgment.

The digital twin idea expands the scope beyond documents

Ludwig Siegele, senior editor for artificial-intelligence initiatives at the time of the discussion, suggested that models would increasingly be pointed at forms of corporate data beyond text and images. He anticipated systems that could extract the structure of a firm and create a kind of digital twin of the organisation. This was a forward-looking proposition rather than a claim that a complete model was already operating everywhere.

In a factory, the phrase digital twin often refers to a virtual representation of a machine or production process. An organisational twin would be more abstract. It might connect customer demand, staffing, approvals, suppliers, applications, inventory, cash and service levels. The purpose would not be to produce an impressive three-dimensional diagram. It would be to represent dependencies well enough to ask how a change in one part of the business could affect another.

Consider a product launch. The visible plan may include a date, a budget and a list of teams. The underlying system includes component availability, regulatory approval, training, marketing assets, support capacity, regional pricing and technical integrations. A model that connects those elements could expose a bottleneck earlier or help leaders explore alternatives. Yet the model would always be incomplete. Informal relationships, local workarounds and rapidly changing conditions are difficult to capture, and a diagram can create false confidence if its omissions are not visible.

Data integration is an organisational negotiation

Building an enterprise information layer is often described as a technical integration project. The harder work is deciding what information means and who owns the definition. Finance, sales and operations may calculate the same apparently simple measure in different ways. Customer names may not match across systems. A project can have one identifier in budgeting and another in delivery. Artificial intelligence can propose connections, but leaders must decide which distinctions are legitimate and which inconsistencies need correction.

The safest starting point is a bounded domain with a clear owner and a recurring decision. A company might begin with product incident reviews, commercial proposals or maintenance records rather than opening every repository at once. The team can map the information lifecycle, remove obsolete material, define access rules and observe how users test the answers. Each mistake becomes evidence for improving retrieval and governance before the scope expands.

This sequence also protects trust. If an early assistant exposes restricted information or repeatedly cites the wrong version, employees will learn to avoid it. If the system performs well in a meaningful but contained workflow, users can develop an accurate sense of its strengths and limitations. Adoption then grows from demonstrated usefulness rather than an executive announcement.

A new interface to information changes managerial behaviour

Abby Bertics, then a science and technology correspondent, focused on the difficulty of navigating abundant data. Her argument pointed toward a more intuitive interface: people would spend less effort locating a needle in a haystack and more effort working with organised information. For businesses, the interface is not a cosmetic layer. It influences which questions are asked, whose knowledge becomes visible and how quickly a tentative idea turns into a decision.

A conversational interface lowers the cost of exploration. A regional manager who cannot write a database query can ask why delivery times changed, request a comparison with previous periods and identify records that deserve review. A product leader can ask which customer objections recur across interviews. A new employee can trace the history of a policy. These capabilities can distribute analytical access more widely, provided the system respects the same permissions and context that governed the original information.

Convenience can also encourage shallow reasoning. A polished paragraph may end investigation too early, especially when the answer confirms what a manager already believes. Interface design should therefore invite challenge. It can show alternative explanations, disclose the period covered, highlight missing data and offer a direct path to the original record. The goal is not to remove productive friction from judgment; it is to remove clerical friction from gathering evidence.

Permissions must travel with the knowledge

Traditional access controls are often attached to applications or folders. An artificial-intelligence system cuts across those boundaries, so a single response could combine information from several sources. That creates a basic governance rule: the generated answer must never reveal more than the user could legitimately retrieve from the underlying systems. Security cannot be added after the model has indexed everything.

Identity, role, geography, confidentiality and purpose may all affect access. A manager might be allowed to see an aggregate trend but not individual employee records. A legal team may review privileged material that must not appear in a general search. A merger project can require a temporary information boundary. The knowledge layer needs to enforce these distinctions during retrieval and generation, record what was accessed and support investigation when an answer is challenged.

Data minimisation is equally important. A system does not become better simply because it ingests every available file. Duplicate, obsolete and irrelevant material can reduce quality while increasing exposure. Information owners should decide which collections serve the defined use case, how long derived indexes remain current and when deleted source data must disappear from the knowledge layer.

Human responsibility cannot be delegated to a fluent answer

Tom Standage, course adviser and deputy editor, placed the debate between extreme optimism and extreme pessimism. His practical expectation was that artificial intelligence would become a useful tool employed in ways that are not yet fully predictable, with the central challenge being to make benefits outweigh harms. That middle position is especially relevant inside companies, where ordinary decisions can still have serious effects on employees, customers and suppliers.

Human review should not be a ceremonial approval at the end of an automated process. The responsible person needs enough context, authority and time to disagree with the system. High-impact uses require a clear record of the question, evidence, model output, edits and final decision. If an employee is evaluated, a supplier rejected or a customer denied service, the organisation must be able to explain the business rule and correct an error.

Responsibility also means assigning an owner to the system itself. Technology teams can operate infrastructure, but the business function must own the decision process and quality standard. Legal, security and risk specialists define constraints. Information owners maintain source collections. Users report failures. Senior leadership decides which uses are unacceptable even when they appear efficient.

Measurement should begin with the decision, not the model

Artificial-intelligence programmes often report technical activity: number of users, queries, documents indexed or responses generated. These measures describe adoption, not business value. A credible evaluation begins with the decision or workflow the system is intended to improve. The baseline may be research time, error rate, duplicated effort, delayed approval, missed precedent or the cost of assembling evidence.

Quality needs several dimensions. An answer can be factually supported but incomplete. It can retrieve the right documents but present them too slowly. It can save analyst time while shifting verification work to a reviewer. A programme should therefore combine task outcomes with sampling of citations, permission tests, user corrections and incidents. Different uses deserve different thresholds: a brainstorming assistant can tolerate more uncertainty than a system influencing a contract or regulatory filing.

A balanced scorecard for the knowledge layer

1. Measure the time from a defined business question to reviewed evidence.
2. Test whether material claims point to the correct, current and permitted source.
3. Track corrections, unresolved conflicts and cases where users abandon the answer.
4. Compare the quality of the final decision with the previous workflow, not just the first draft.
5. Monitor security, privacy and policy exceptions as operating outcomes rather than technical footnotes.
6. Review whether benefits reach different roles and locations or remain concentrated among specialists.

These measures create an honest feedback loop. They can show when a narrow deployment is ready to expand and when the underlying data or process needs repair first.

The workforce effect is a redesign of expertise

If artificial intelligence reduces the time required to search, summarise and organise records, professional work will not simply become the same work performed faster. Roles will change around the new constraint. Analysts may spend less time assembling material and more time testing assumptions, interpreting conflicts and explaining consequences. Managers may ask more questions because exploration is cheaper, creating a greater need for prioritisation rather than less.

Expertise remains important because a system cannot reliably determine which analogy is useful without understanding the domain. The hedge-fund archive described by Ramani can surface earlier crises, but an experienced analyst must judge whether market structure, policy and incentives make the comparison meaningful. The organisational model anticipated by Siegele can expose a dependency, but operating leaders must know whether the recorded process reflects how work actually happens.

Training should therefore combine tool use with source criticism, data literacy, privacy, business judgment and clear writing. Employees need to know how to frame a question, inspect evidence, recognise uncertainty and document a decision. Organisations should reward people who improve the shared knowledge base, not only those who privately accumulate expertise. The quality of the system will depend on the quality of the records and corrections contributed by its users.

A phased path turns experimentation into infrastructure

The distance between a chat pilot and a dependable operating layer is large, but companies do not need to cross it in one programme. A disciplined sequence can produce value while containing risk. First, select a recurring knowledge task with clear pain, available evidence and an accountable business owner. Define what a better result means before choosing a model. Establish a baseline using the current process.

Second, prepare a limited collection. Resolve obvious duplication, label authoritative versions and preserve access controls. Build retrieval that returns evidence with dates and ownership. Ask representative users to perform real tasks, and retain difficult questions as an evaluation set. Record not only successful answers but also where the system refuses, misses context or exposes disagreement.

Third, integrate the tool into the decision workflow. Define when human review is required, what must be documented and how users report a problem. Measure the final outcome over a meaningful period. Only then should the organisation add new collections, roles or actions. A digital twin should emerge from validated connections between systems and processes, not from an attempt to map the whole company before any use has proved valuable.

The strategic advantage is better institutional learning

Tom Chatfield's discussion with Arjun Ramani, Ludwig Siegele, Abby Bertics and Tom Standage captured four connected ideas: corporate archives can become more usable, models can work across broader operational data, interfaces can make information easier to navigate and the most likely future sits between technological utopia and disaster. Together they describe a practical business agenda rather than a single product.

Companies already possess much of the raw material: decisions, contracts, incident reviews, customer conversations, process records and performance data. Their weakness is often the inability to connect that material at the moment a decision is made. Artificial intelligence can reduce this fragmentation, but only if evidence, permissions and accountability remain visible. Otherwise the organisation merely replaces scattered documents with a fluent new source of confusion.

The enduring advantage will not come from generating the largest volume of content. It will come from learning more reliably: remembering why choices were made, comparing outcomes with expectations, reusing lessons across teams and updating the model when reality contradicts it. An artificial-intelligence knowledge layer becomes strategic when it improves that cycle without pretending to remove uncertainty. The company still needs people to choose. It simply gives them a better map of what the organisation knows, how its parts connect and where the map remains incomplete.