The Financial Statement
Is a Projection

The x-ray

A financial statement is one of civilization’s most successful compression technologies.

A company may contain thousands of employees, millions of inventory movements, years of contracts, hundreds of machines, informal promises, supplier dependencies, customer relationships, software systems, operating routines and unfinished decisions. The balance sheet reduces this moving institution to assets, liabilities and equity at a moment in time. The income statement reduces a period of activity to revenue, expense and profit. The cash-flow statement explains the change in one especially important resource.

This is low-resolution information in the same sense that an X-ray is a low-resolution image of a body. It discards color, tissue chemistry, motion, memory and intention. It still reveals the skeleton.

The loss of detail is not a defect accidentally left behind by accountants. It is what makes the statement portable. An owner, lender, investor, regulator or distant manager can receive a compact document, compare it with prior periods, reconcile its totals and use the same conceptual language across companies. Double-entry bookkeeping was practiced before Luca Pacioli, but his 1494 Summa contained the first printed exposition of the Venetian method and helped make the system transmissible.1

Financial reporting standards remain explicit about the boundary. The IFRS Conceptual Framework describes general-purpose reports as information for investors, lenders and other creditors. It also says those reports cannot provide all the information those users need, are not designed to show the value of the entity, and need not be management’s only source because management can obtain internal information directly.2

The financial statement is not a failed model of the company. It is a successful projection of the company.

The next question is therefore not how to make the income statement infinitely longer. It is whether the paper-era projection should remain the canonical computational object inside the firm.

Compression was the feature

For most of commercial history, preserving the full structure of a firm was prohibitively expensive.

Every additional dimension required clerks to observe it, choose a category, write it down, move it between books, reconcile discrepancies and produce a report before the information became stale. The organization therefore evolved a hierarchy of compression:

This hierarchy created the modern corporate bureaucracy. Local operators generated documents. Accountants classified and reconciled them. Controllers closed periods. Executives received summaries. Boards and outside parties received an even narrower layer. The organization could be managed across distance because the information became progressively smaller.

The cost was that causality disappeared early. A payroll expense tells us money was consumed. It does not preserve which customer created the work, which process absorbed it, whether the work was productive or corrective, whether it created a reusable capability, or which future commitments now depend on the people who performed it.

Once the aggregation has occurred, no amount of intelligence can reliably reconstruct what was never retained.

The computer kept the form

Computers removed many physical limits. Storage became cheap. Arithmetic became automatic. A transaction could be copied without a clerk rewriting it. A database could preserve details that would never fit inside a paper ledger.

Yet enterprise software largely computerized the inherited bureaucracy. The screen became an invoice. The database table became a journal. The module became a department. The monthly report became a dashboard. The general ledger remained the point at which the company’s different activities became comparable and auditable.

This was rational. The ledger had legal meaning, a mature control system, trained users and a clear test of closure. A new database architecture could not ask a company to surrender the one representation known to balance.

But the result is a peculiar enterprise stack. The firm now possesses more raw data than any prior bureaucracy could imagine, while management still spends enormous effort reconstructing context around financial aggregates. Data warehouses copy records out of operating systems. Analysts join tables. Finance teams ask departments why a number moved. Managers produce slide decks that manually reconnect the monetary symptom to the operational cause.

The bottleneck moved. It is no longer the cost of writing another number. It is the cost of maintaining a coherent meaning across all the numbers.

The alternative has a history

The idea that computers should preserve economic events rather than only their accounting summaries is not new.

In 1969, George Sorter’s “events” approach argued that accounting could provide information about relevant economic events and let users apply their own decision models, rather than having accountants pre-aggregate every fact into a small set of values. Sorter explicitly treated aggregation as an information loss whose benefit had to justify its cost.3

In 1982, William McCarthy proposed the REA model—Resources, Events and Agents—for accounting systems in a shared data environment. Instead of storing the debit and credit as the deepest representation, REA modeled the economic resources, economic events and participating agents that the accounting records described. The paper’s database framing was direct: an enterprise-wide model could support many local views.4

XBRL Global Ledger later pursued a related goal at the reporting and interchange layer. Its official description notes that summary reports ordinarily lose their connection to supporting detail; XBRL GL is intended to retain transaction-level financial and non-financial information as it flows toward reports.5

Process mining developed on another branch. It uses event logs to discover, monitor and improve the processes that actually occurred, including conformance checking, bottleneck analysis, organizational relationships and simulation. The Process Mining Manifesto begins from a simple observation: modern systems already leave event trails, and those trails can generate living process models rather than static diagrams.6

Research on digital twins of organizations and business processes extends the ambition again: connect relatively static enterprise models to operational flows, update them as work occurs, and use the result for simulation and adaptation.78

The pieces already exist. The conceptual step is to make the high-resolution model—not the report, dashboard or departmental database—the canonical internal object.

The enterprise geometry

“Geometry” should not be understood as a decorative three-dimensional visualization. The full company graph would be unreadable on a screen. Geometry describes the structure preserved by the computational object.

A practical enterprise geometry would be a temporal, typed economic hypergraph with quantities, states, constraints and provenance.

A graph edge ordinarily joins two objects. A commercial event often joins many at once. A goods receipt connects a supplier, purchase order, contract, product, batch, warehouse, employee, inspection result, payable, currency exposure and future production capacity. Treating the event as a hyperedge preserves that joint structure.

The model needs at least six primitives:

The geometry creates operations that ordinary statements cannot perform naturally.

Paths trace how demand becomes cash: customer request → order → procurement → production → delivery → invoice → payment. The system can measure duration, working capital, handoffs, rework and uncertainty along the path.

Topology reveals what remains connected when something fails. A supplier, employee, database or license can be an articulation point whose removal disconnects a large share of revenue from the resources needed to produce it.

Distance allows objects to be compared by actual economic behavior rather than a single category. Two customers may be close because they buy similar products, consume similar support, use similar routes and pay with similar delay.

Deformation through time shows how the firm changes shape: concentration rises, workflows lengthen, knowledge accumulates in one person, commitments outrun capacity, or a formerly modular operation becomes dependent on a central hub.

The model is therefore not a data lake. A lake stores observations. A geometry states what the observations are observations of.

Accounting as conservation

A rich enterprise graph without accounting closure would become another management dashboard: visually persuasive, semantically loose and impossible to audit.

The ledger should remain authoritative for statutory accounting, tax, reconciliation and external reporting. But it should sit beside the richer event model, linked through common identifiers. Whenever possible, both should derive from the same normalized business event.

Accounting supplies conservation laws:

This distinction creates two levels of truth. Book truth is controlled, reconciled and recognized under accounting policy. Model truth includes operational facts, inferred relationships, forecasts and counterfactuals. They can inhabit the same geometry only if the interface refuses to present them as equally certain.

Agents supply semantic labor

The reason this architecture has remained uncommon is not database storage. It is semantic labor.

Someone must connect a supplier’s email to the correct purchase order, notice that two systems refer to the same customer, interpret a contract clause as a future capacity claim, assign an exception to a workflow, preserve the history when a department changes its name, and decide whether an apparent relationship is causal or merely correlated.

Traditional systems handle the clean portion through forms, mandatory fields and integration projects. The unclean portion remains in email, spreadsheets, meeting notes, chat, contracts, attachments and human memory. Maintaining a high-resolution model by hand would create a bureaucracy larger than the one it replaced.

Agents change the economics because they can perform bounded semantic work continuously:

This should not be confused with making the language model the ledger. Agents are probabilistic interpreters. The historical record, accounting engine, identity system, permissions and reconciliation controls should remain deterministic wherever the underlying fact permits it.

Agent capabilities are improving, but real-world reliability remains uneven. METR’s task-horizon work is useful precisely because it measures whether an agent can complete a multi-step task at a stated reliability, while warning that its current suite is mostly self-contained software, machine-learning and cybersecurity work. That is evidence for a falling cost of bounded agentic workflows—not proof that a model should autonomously close the books.9

AI should maintain the map around the ledger. It should not be allowed to redraw the territory without evidence.

Reports become compiled views

Once the firm exists as a coherent computational object, a report is no longer a manually assembled destination. It is a query-defined projection.

The same model can generate a statutory balance sheet, a causal profit map, a commitment statement, a dependency report, a process-flow view or a counterfactual simulation. Each view uses different dimensions, but all should retain drill-through to the same events and reconcile wherever they touch book truth.

The most important views may not resemble existing financial statements.

A causal profit report could follow revenue through products, workflows, rework, delivery and support rather than stopping at functional expense categories. A commitment statement could show promises against future cash, inventory and human capacity whether or not they meet the recognition threshold for a liability. A dependency statement could reveal that a profitable product relies on one supplier, one employee and one database. A process-flow statement could show where working capital and calendar time accumulate. A counterfactual report could change a price, supplier, approval rule or capacity constraint and then reconcile the simulated consequence back to cash and accounting position.

These reports do not compete to become the new universal statement. Their value comes from being different projections of the same controlled object.

The model-maintaining bureaucracy

The institutional change is larger than a new reporting feature.

Today, much corporate bureaucracy exists to convert activity into documents and documents into reports. People request files, copy fields, chase explanations, rebuild context, reconcile versions and prepare recurring packs. A model-first organization would move part of that labor upstream. Agents would continuously ingest and connect activity; humans would maintain definitions, controls, exceptions and authority.

This could alter the meaning of closing. External books would still close under controlled accounting policy. Internal understanding would become more continuous. A monthly close would certify a boundary inside a living model rather than awaken the organization from a month of informational sleep.

The model would also change management interfaces. Instead of receiving every possible metric, a manager could ask a question that defines a slice:

The answer might be a PDF. It might be a graph, table, alert, narrative or simulation. The document becomes a temporary interface. The model remains.

What not to build

A high-resolution model can fail more dangerously than a low-resolution statement because it can create the impression of total knowledge.

There is also a political risk. A firm model can become an instrument of total managerial surveillance, allowing every deviation to be scored while hiding the uncertainty and value of local judgment. The architecture should preserve access controls, purpose limitation, contestability and the ability to distinguish an exception from misconduct.

The model must remain subordinate to the institution it describes. It should make the company more legible without pretending the company is only what the model can see.