# The Coordination Tax: How AI Ontologies and Cross-Process Intelligence Can Lower the Cost of Running a Business
Canonical HTML: [https://collty.com/insights/the-coordination-tax-how-ai-ontologies-and-cross-process-intelligence-can-lower-the-cost-of-running-a-business](https://collty.com/insights/the-coordination-tax-how-ai-ontologies-and-cross-process-intelligence-can-lower-the-cost-of-running-a-business)
Markdown URL: [https://collty.com/insights/the-coordination-tax-how-ai-ontologies-and-cross-process-intelligence-can-lower-the-cost-of-running-a-business/index.md](https://collty.com/insights/the-coordination-tax-how-ai-ontologies-and-cross-process-intelligence-can-lower-the-cost-of-running-a-business/index.md)
Author: Collty
Published: 2026-09-25T14:01:00.000Z
Updated: 2026-09-25T16:26:22.637Z
Category: Strategy
Tags: #AI #Ontology #Coordination-Tax #ERP #HR
The economic case for enterprise AI is often framed around automating individual tasks. The larger opportunity may lie elsewhere: reducing the enormous coordination costs created when processes, teams, data, decisions and software operate as separate systems.
![The Coordination Tax: How AI Ontologies and Cross-Process Intelligence Can Lower the Cost of Running a Business](https://collty.com/insights-media/covers/2026/09/7db06a86-507a-4d7b-a800-47124934b91c.jpg)

Over the past two decades, companies have invested heavily in digitization, but digitization has not necessarily produced operational unity. Sales adopted CRM platforms, finance moved to ERP and accounting systems, HR deployed dedicated workforce software, project teams adopted task-management platforms, marketing accumulated analytics and automation products, and individual departments continued building their own spreadsheets, databases and dashboards around all of them. Each investment could improve a particular function, yet the cumulative effect has often been an enterprise composed of dozens or hundreds of partially connected systems, each representing only one fragment of how the company actually works.

This fragmentation has an important economic consequence because business outcomes are rarely produced by individual departments or applications. Revenue may begin with marketing, move through sales and contracting, require delivery by several teams, depend on procurement or external specialists, generate invoices through finance and later return as customer retention or additional projects. From the perspective of a customer or shareholder, this is one economic process. From the perspective of the company's software architecture, however, it can be fragmented across six, ten or twenty systems, with different owners, data models, metrics and definitions at every stage.

The result is a form of operating cost that rarely appears explicitly on a profit-and-loss statement. It appears instead as time spent finding information, reconciling conflicting figures, recreating work that already exists, coordinating handovers, preparing management reports, correcting process errors, attending status meetings and manually moving information from one system to another. It also appears in less visible forms: unused capacity in one team while another hires external contractors, delayed invoicing because project information has not reached finance, duplicated software subscriptions, inconsistent customer data, missed dependencies between projects, and management decisions made from data that describes separate functions rather than the business as a whole.

This can be described as a coordination tax: the economic cost a company pays because its operating reality is more interconnected than the systems through which it manages that reality. Artificial intelligence can reduce this tax, but only if AI is introduced as more than another application sitting on top of the existing fragmentation. The more consequential opportunity is to combine AI with a shared business ontology, centralized visibility across processes, and analytics that connects operational events with teams, resources and financial outcomes. Under this model, AI does not simply make individual tasks faster; it becomes part of an architecture capable of understanding how the business operates as a system.

The hidden cost of time, energy, and money an organization pays just to keep people aligned, organized, and communicating. As teams grow, the effort required to collaborate and sync often increases exponentially, reducing the time actually spent on productive work.

### Why task automation alone does not solve the problem

The first wave of generative AI in business has largely followed the pattern of earlier enterprise software adoption. Organizations introduced copilots for developers, assistants for customer-support teams, AI tools for marketing, document processing for finance and conversational interfaces for corporate knowledge. Many of these applications are useful and some generate significant productivity gains, but improvements at the task level do not automatically translate into comparable improvements in company-level economics.
Research increasingly supports this distinction. In McKinsey's 2025 global AI survey, workflow redesign showed the strongest relationship with reported EBIT impact among the organizational practices the researchers tested. Only 21 percent of respondents whose organizations were using generative AI said that at least some workflows had been fundamentally redesigned, suggesting that most companies were still adding AI to existing operating structures rather than redesigning those structures around the new technology. The implication is important: organizations can make dozens of tasks individually more efficient while preserving much of the coordination overhead that exists between those tasks.

There is already strong evidence that AI can improve productivity within a clearly defined activity. A well-known field study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined 5,179 customer-support agents and found that access to a generative AI assistant increased issues resolved per hour by an average of 14 percent, with substantially larger improvements among novice and lower-skilled workers. That is a meaningful result, but it illustrates only one layer of the economics. If a customer-support process still depends on inaccurate CRM records, several disconnected approval systems, manual billing reconciliation and information distributed across multiple teams, improving the productivity of the support interaction itself leaves much of the system-level cost untouched.

The distinction therefore is not between automation and no automation, but between local optimization and systemic optimization. Local AI asks how a person can complete a particular activity faster. Systemic AI asks why the activity exists in its present form, what information it depends on, which process precedes it, which process follows it, what other teams are involved, what outcome it produces, what it costs, and whether software or an AI agent could coordinate part of that flow automatically. The financial opportunity becomes considerably larger when the unit of optimization changes from the task to the process and eventually from the process to the operating system of the company.

Local optimization focuses on making a single part of a system as efficient as possible, while systemic optimization focuses on making the entire system work together to achieve the best overall result. Optimizing locally can often accidentally harm the broader system a phenomenon known as the "local optimization trap."

### The enterprise has a data model, but usually not a model of itself

Most companies already possess enormous quantities of operational data. What they frequently lack is a common model that explains what those data represent in business terms and how the different parts of the organization are related. A row in a database might describe a project, a customer, an invoice or an employee, but the relationships among those objects are often reconstructed separately by every application, analyst and integration.

An AI Ontology addresses this problem by creating a machine-readable representation of the important entities within a business, their properties, relationships, states, rules and available actions. Instead of exposing an AI system to thousands of unrelated tables and documents, an ontology can describe a connected operating environment in which a customer has contracts, contracts generate projects, projects use teams, teams consist of people and agents, tasks consume resources, milestones trigger invoices, invoices produce payments, and project outcomes affect future planning. The important change is not simply better data organization. The ontology makes the relationships that constitute the business explicit and reusable.

This architecture is moving rapidly into mainstream enterprise technology. Palantir describes its Ontology as a digital representation of an organization built from objects, properties, links and actions, connecting underlying data and models to real-world concepts and operational workflows. Microsoft has introduced Ontology in Fabric as a shared, machine-understandable business vocabulary designed to represent entities, properties and relationships and provide consistent context to analytics and AI agents; Microsoft explicitly identifies cross-domain reasoning and consistency as target use cases. SAP's 2026 AI-native architecture similarly describes a transition from systems of record toward systems of context in which enterprise data, process knowledge and decision history become connected so that AI agents can reason across them rather than operate on fragments contained within individual applications.

The significance of this development is economic as much as technical. Every time two systems contain different representations of the same customer, project, employee or financial event, someone eventually has to resolve the difference. Every time a new application needs its own mapping between corporate concepts and raw data, the company incurs another integration cost. Every time an AI agent requires a separate description of how a project, approval, team or invoice should be interpreted, the organization recreates context that should already be part of its operating architecture. A shared ontology cannot eliminate these costs completely, but it can move the company from repeatedly reconstructing context toward maintaining that context as reusable infrastructure.

Applied to business, an ontology is a structured, machine-readable model of how a company understands its own operating reality. It defines the principal business objects — such as customers, contracts, projects, teams, employees, AI agents, tasks, products, invoices and payments — together with their properties, relationships, states, rules and, increasingly, the actions that can be performed on them. A database may tell a system that a project has an ID, a budget and a start date; an ontology explains that the project belongs to a client, is governed by a contract, is executed by particular teams, consumes resources, contains tasks and milestones, generates costs and invoices, and contributes to financial and operational outcomes. In this sense, a business ontology is not merely another data structure: it is a common semantic model through which people, analytics systems and AI agents can interpret the company in the same way and reason across processes that would otherwise remain separated by applications and departmental boundaries.

### The economics of finding, reconciling and recreating information

One of the easiest parts of the coordination tax to observe is the amount of employee time consumed by information friction. APQC research involving knowledge workers found that the average employee spent 8.2 hours per week looking for, recreating or duplicating information and expertise. In another APQC study of 982 full-time knowledge workers, employees reported spending 2.8 hours per week simply looking for or requesting information, alongside additional time consumed by unnecessary communication, meetings and workarounds for broken systems and processes.

These figures illustrate why seemingly minor information problems become financially material at scale. Consider, purely as an illustrative model, an organization employing 1,000 knowledge workers at an average fully loaded labor cost of $60 per hour. Applying the 8.2-hour APQC benchmark across 48 working weeks represents approximately 394,000 hours of employee capacity, equivalent to roughly $23.6 million in labor cost. A company would not realistically eliminate all of this activity, nor would every hour recovered become a cash saving, but recovering even 15 to 25 percent of that capacity would represent approximately $3.5 million to $5.9 million a year that could potentially be redirected toward productive work.

This distinction between capacity value and cash savings is essential when evaluating AI business cases. Saving an employee ten hours does not automatically reduce expenses if the employee remains on payroll and the recovered time is not used productively. Financial impact appears when better information and automation allow the company to avoid additional hiring, reduce contractor expenditure, increase output with the same workforce, remove redundant administrative roles, shorten revenue cycles or redirect scarce expertise toward higher-value activities. For that reason, the most credible AI economics should connect productivity metrics directly to operational and financial outcomes rather than simply convert every saved hour into hypothetical dollars.

An ontology combined with cross-process analytics creates the infrastructure for making that connection. Instead of measuring only how much time an AI tool saved, the organization can observe whether reduced administrative effort allowed a team to run more projects, whether faster access to knowledge shortened delivery cycles, whether improved capacity visibility reduced external hiring, and whether better coordination improved project margin. Productivity stops being an isolated software metric and becomes part of the economic model of the business.

American Productivity & Quality Center is a U.S.-based nonprofit organization specializing in benchmarking, process and performance improvement, best practices, and knowledge management. It was founded in Houston in 1977 by business leader and educator C. Jackson “Jack” Grayson, initially as the American Productivity Center, with the goal of improving organizational productivity and competitiveness. Over the following decades, APQC became one of the most established international sources of comparative operational data and developed the Process Classification Framework (PCF), a widely used reference model for describing and benchmarking business processes.

### Process visibility can reveal costs that departmental analytics cannot

Traditional business intelligence usually mirrors organizational structure. Management receives sales analytics, finance analytics, HR analytics, marketing analytics and project dashboards, each describing a particular part of the company. This can produce accurate reporting while still failing to explain the causes of important business outcomes because many causes exist in the relationships between those functions.

A project that becomes unprofitable provides a simple example. Finance can see the margin deterioration, project management can see delays, HR can see utilization and sales can see the original contract value, yet none of these perspectives alone explains why the project failed economically. The underlying cause might have been an unrealistic estimate during sales, an inappropriate team composition, repeated client approvals, excessive use of senior specialists, poor task dependencies, external contractor costs, rework caused by quality problems or simply a delay between completed work and invoicing. When process, workforce and financial information are connected, management can begin asking questions about cause and effect that are extremely difficult to answer from functional dashboards.

Process-mining evidence illustrates how financially significant this visibility can become. McKinsey describes a distribution company that combined process and task mining to examine its order-to-cash process and discovered extensive manual intervention, including manual work affecting up to one-third of invoices. Changes based on the analysis generated approximately $30 million in efficiency savings, while improvements associated with write-offs and credit information contributed $18 million in additional revenue and changes to payment discipline reduced working-capital requirements by $5 million. These figures are a case study rather than a universal benchmark, but they show why the economic value of process intelligence can extend beyond labor savings into revenue leakage, working capital and customer service.

The same principle can be applied to project-based and knowledge-intensive companies. If the operating model connects projects, tasks, teams, skills, costs, contracts, timelines and outcomes, analytics can examine not only whether a project succeeded but which combinations of conditions repeatedly produce successful or unsuccessful outcomes. Management may discover that a particular handover consistently adds five days, that projects involving a certain approval path generate more rework, that certain team structures produce higher margins, or that expensive external specialists are regularly hired while relevant internal capacity remains unused elsewhere in the organization. None of these discoveries requires replacing the company's existing ERP, CRM or accounting software; what is required is an analytical layer capable of seeing across them.

### From workforce utilization to workforce orchestration

Labor is one of the largest cost categories in most knowledge-based businesses, but conventional workforce analytics often describes people independently of the processes in which they create value. Companies monitor headcount, salary, utilization and perhaps skills, while project systems monitor tasks and deadlines and finance measures revenue and margin. As long as these layers remain separate, management has only a partial understanding of workforce economics.

A connected operating model can treat workforce allocation as a dynamic optimization problem. The system can understand which projects are beginning, what capabilities they require, which people or teams possess those capabilities, what their current workloads are, which external resources are being purchased, how different team configurations have performed historically and what financial constraints apply. AI can then assist not merely with finding a person who possesses a particular skill but with determining how internal employees, external specialists and AI agents should be combined across a portfolio of work.

The potential saving is not primarily produced by paying every employee less or attempting to maximize utilization to 100 percent. In fact, excessive utilization can reduce resilience and create new bottlenecks. The more meaningful opportunity comes from reducing mismatches: unnecessary contractor spending, duplicated expertise, teams waiting for unavailable specialists, expensive specialists performing low-value administrative work, managers manually coordinating capacity, and hiring decisions made because existing capabilities are not visible across organizational boundaries. Cross-process workforce analytics therefore links labor cost to actual demand instead of treating headcount and project execution as separate planning exercises.

AI agents add another dimension to this model because they become resources that can be assigned to work alongside people. If agents remain isolated inside individual applications, organizations risk recreating the fragmentation of the SaaS era in agentic form: dozens or hundreds of specialized agents, each with its own memory, permissions, integrations and representation of company context. An ontology offers an alternative in which agents operate against shared definitions of customers, teams, projects, tasks and business rules. The same context used by analytics and human managers can therefore become the context through which agents understand what they are allowed to do and how their work affects the rest of the organization.

### Centralization does not have to mean more bureaucracy

The word centralization can be misleading in this context because the economic objective is not to transfer every operational decision to a central management office. That approach could easily increase overhead rather than reduce it. What needs to become centralized is the representation of the business, the visibility of its processes, the governance of shared definitions and the ability to observe performance across organizational boundaries, while execution itself can remain highly distributed.

The economics of this distinction can already be seen in shared-services and global-business-services research. Deloitte's 2025 Global Business Services Survey found that approximately half of responding organizations reported savings above 20 percent from their GBS operations, while around 55 percent of organizations with a global GBS leader reported average savings above 20 percent. Deloitte associates the results with factors including governance, digital technology, process standardization, end-to-end ownership and more unified decision-making. These results should not be interpreted as evidence that an AI ontology by itself produces 20 percent savings; GBS transformations can involve organizational restructuring, shared services, labor-location strategies and many other changes. They do, however, provide evidence for a broader economic principle: fragmentation has a cost, and integrating governance and processes can release material value.

Historical research on process standardization points in the same direction. One empirical case study of a global company's recruiting operation found that standardization reduced time-to-hire from 92 to 69 days and cut overall recruiting-process costs by approximately 30 percent while improving the quality and transparency of applicant data. Again, the important lesson is not the precise percentage, which cannot simply be transferred to another process or organization. The lesson is that variability, duplication and inconsistent process execution are themselves cost structures that can be measured and reduced.

An AI-native operating layer extends this logic because standardization no longer needs to mean forcing every situation through an inflexible workflow. The organization can standardize the definitions, states, permissions, data relationships and expected outcomes of a process while allowing AI and human managers to adapt execution to context. This makes it possible to combine the economic benefits of common infrastructure with the flexibility required for projects, knowledge work and other environments where rigid workflow automation is often unsuitable.

### The cost of fragmented technology is also becoming an AI problem

Software itself represents another area in which unified process visibility can create savings. When applications are purchased function by function, it becomes difficult to determine whether the company is paying several times for overlapping capabilities. Different teams may independently procure analytics, task management, workflow automation, knowledge management, collaboration or AI services without a clear view of the business capabilities already available elsewhere.

The problem becomes larger when data infrastructure is included. McKinsey has described a global bank that accumulated more than 600 data repositories at an annual management cost of approximately $2 billion. By reorganizing the environment around 40 data domains, defining standardized authoritative sources and retiring redundant repositories, the bank reduced annual data costs by more than $400 million. McKinsey notes more generally that data storage and management can account for 15 to 20 percent of the IT budget in mature organizations. The bank's scale makes this an extreme example, but it demonstrates how architectural complexity eventually becomes a significant operating expense rather than merely a technical inconvenience.

Agentic AI can intensify this problem if every function independently creates agents connected to its own systems and data. Each agent then requires integration, context, permissions, monitoring and maintenance, and multiple agents may recreate the same business logic in different forms. A common ontology reduces part of this duplication by establishing reusable business entities, relationships and actions that can be consumed by different agents and applications. This does not eliminate integration work, but it changes the architecture from repeatedly integrating every AI application with every underlying system toward connecting AI to a common representation of the operating environment.

The financial implication is similar to the move from individually built infrastructure toward shared platforms in earlier generations of enterprise IT. When context, permissions, business entities and process states can be reused, the marginal cost of introducing another analytical application or agent can decrease. More importantly, the company can begin evaluating technology according to the business capabilities and processes it supports rather than simply by application category, making redundant software and duplicated functionality easier to identify.

### The largest saving may come from shortening the distance between problem and decision

Direct labor savings are relatively easy to model, but some of the most valuable effects of integrated operating intelligence appear through time. A problem discovered after a quarterly review is more expensive than the same problem detected after a day. A project margin issue recognized after completion cannot be corrected, an invoice delayed for a month ties up working capital, a procurement anomaly repeated across thousands of transactions becomes material expenditure, and a capacity problem noticed only after external hiring has already occurred creates avoidable cost.

Cross-process analytics can shorten this feedback loop because operational events, team behavior and financial outcomes are observed in the same model. Instead of waiting for a manager to combine reports from several departments, analytics can continuously identify deviations from expected process behavior and AI can investigate the surrounding context. The organization moves gradually from retrospective reporting toward operational decision support, and eventually some decisions can be executed automatically within defined limits.

This is where an ontology becomes particularly important for agentic AI. An agent that can read information but does not understand the relationships, rules and permissions of the business remains primarily an assistant. An agent that understands that a delayed milestone is connected to a client commitment, which is connected to an invoice, which affects cash flow, while also knowing which team owns the milestone and which actions are authorized, has the context required to participate in an operational process. Current enterprise architecture directions from Microsoft, Palantir, SAP and AWS all increasingly emphasize semantic or ontological context as a foundation for AI agents that must reason across enterprise data rather than simply retrieve text.

The long-term economic consequence could therefore be larger than the automation of individual tasks. When a business creates a continuous loop in which operational activity produces structured data, analytics identifies patterns, AI proposes or executes interventions, outcomes are measured, and the resulting knowledge updates future decisions, process improvement itself becomes partially continuous. The company is no longer periodically analyzing how it operates; its operating infrastructure is continually generating evidence about how it could operate better.

### How the savings compound

It is tempting to evaluate each of these effects independently: several percent from automation, another amount from reducing information search, another from process optimization and another from software consolidation. In practice, the effects interact. Better process visibility improves automation because AI receives better context; better ontology increases analytics quality because metrics operate on consistent definitions; better workforce data improves project planning; better project execution improves financial forecasting; and better outcome data makes future AI recommendations more accurate.

This compounding effect explains why the business case for an AI ontology is fundamentally different from the business case for purchasing another productivity tool. The ontology itself does not directly save money in the same way that eliminating a software subscription does. Its value lies in creating the common operating context through which many forms of inefficiency become visible and addressable simultaneously. It is therefore better understood as infrastructure for continuous cost optimization rather than as a single cost-reduction feature.

The same reasoning also explains why ROI calculations should remain conservative. The credible claim is not that every enterprise implementing ontology and cross-process AI will reduce operating expenses by a predetermined percentage. Existing evidence comes from different technologies, sectors and organizational models, and the financial outcome will depend heavily on the level of fragmentation, process maturity, labor intensity and existing technology landscape. What the evidence does show consistently is that information friction consumes material employee capacity, process standardization can lower administrative cost, end-to-end process analysis can identify substantial operational and working-capital opportunities, unified governance can generate significant efficiencies, and AI creates more value when workflows themselves are redesigned rather than when AI is merely attached to existing tasks.

### From digital transformation to operating-model transformation

For most of the digital era, companies improved themselves by digitizing individual functions. The next stage is likely to be more structural. AI makes it technically possible for organizations to connect processes, data, teams and decisions at a level that would previously have required enormous amounts of manual analysis and integration, while ontologies give those AI systems a common representation of the environment in which they operate.

The resulting architecture does not require one giant application to replace every CRM, ERP, accounting platform or specialist system. A more realistic model is a common operating layer above those systems: an ontology that defines important business objects and relationships, a process layer that represents how work moves through the organization, an analytical layer that connects operational activity with economic results, and an AI layer that can reason and act within that context. Existing systems remain systems of record where appropriate, while the new layer becomes a system of context and increasingly a system of coordinated action.

For executives, the economic question therefore changes. Instead of asking only how many tasks AI can automate or how many employees a particular tool can make more productive, management can begin asking how much the organization spends because information, processes and resources are fragmented in the first place. How much capacity is consumed by searching and reconciliation? How much margin disappears through rework and poor handovers? How much external labor is purchased because internal capacity is invisible? How much cash is trapped by process delays? How much software and integration work exists because every function maintains a separate representation of the company? How many problems are discovered only after they have already become expensive?

These costs have always existed, but until recently they have been difficult to observe as a single economic category. AI Ontologies, process intelligence and cross-team analytics make it possible to treat them as parts of the same problem. The most important promise of enterprise AI may therefore not be simply that machines can perform more work. It may be that companies can finally understand, coordinate and continuously improve the complex system through which all of their work is performed.

The word ontology comes from the Greek ontos (“being” or “that which exists”) and logos (“study” or “account”). Although questions about the nature and categories of existence go back to ancient philosophy, the term ontologia itself appeared much later: one of its earliest known published uses was in Jacob Lorhard’s Ogdoas Scholastica in 1606, and it was subsequently popularized as a philosophical concept in the eighteenth century by Christian Wolff. In philosophy, ontology is concerned with identifying what exists and how different kinds of entities relate to one another. During the development of artificial intelligence and knowledge engineering in the late twentieth century, the concept was adapted for computing. In 1993, computer scientist Thomas Gruber gave what became one of the field’s most influential definitions, describing an ontology as an “explicit specification of a conceptualization”: in practical terms, a formal description of the concepts that exist within a particular domain and the relationships between them.