Data preparation costs should be attributed to AI initiatives based on how directly those activities serve the AI project. Costs tied exclusively to a single AI use case, such as labeling training data or building a dedicated data pipeline, belong entirely to that initiative. Shared data infrastructure costs require allocation models that reflect actual consumption or benefit. Below, we unpack the full picture: what these costs include, why attribution is difficult, and how IT financial management practices can bring consistency to AI investment accounting.
What types of costs make up data preparation for AI?
Data preparation costs for AI initiatives include every expense involved in making raw data ready for model training, validation, and deployment. These costs typically span labor, tooling, storage, compute, and governance activities. Understanding the full scope is the first step toward accurate AI cost attribution.
The main cost categories include:
- Data ingestion and integration: Extracting data from source systems, building pipelines, and transforming formats so data is usable by AI workflows
- Data cleansing and validation: Identifying duplicates, correcting errors, and enforcing quality standards, often the most labor-intensive phase
- Data labeling and annotation: Manually or semi-automatically tagging data so supervised learning models can train against it
- Storage and compute during preparation: The cloud or on-premises resources consumed while running transformation jobs, running quality checks, or hosting staging environments
- Data governance and compliance activities: Privacy reviews, consent management, and lineage tracking required before data can legally or ethically feed an AI model
- Tooling and platform licenses: Data engineering platforms, ETL tools, and orchestration software used specifically during preparation phases
Many organizations underestimate data preparation costs because they treat them as general IT overhead rather than as direct inputs to an AI investment. In practice, data preparation routinely represents the largest share of total AI project spend, which makes accurate tracking important for understanding true AI investment value.
Why is attributing data preparation costs to AI projects so difficult?
Attributing data preparation costs to AI initiatives is difficult primarily because data assets and infrastructure are shared across multiple consumers. A single data pipeline may feed a business intelligence dashboard, a regulatory reporting process, and an AI model simultaneously. When there is no clear boundary between these consumers, assigning costs to any one of them requires judgment rather than direct measurement.
Several structural factors compound this difficulty:
- Shared infrastructure with no metering: Most data platforms were not built to track consumption per downstream use case, so there is no native way to separate AI-related usage from other usage
- Preparation work that predates the AI project: Historical data cleansing or pipeline work done years earlier still benefits the AI model, but charging that sunk cost to the current initiative distorts the project’s economics
- Cross-functional ownership: Data engineering, IT, and business teams each own parts of the preparation process, and costs sit in different budget lines across departments
- Iterative and non-linear workflows: AI projects cycle through data preparation repeatedly as models improve, making it hard to define when preparation ends and model development begins
- Lack of agreed allocation policy: Without a formal governance framework, each team applies its own logic, producing inconsistent numbers that finance and leadership cannot rely on
The result is that AI cost allocation decisions are often made informally, late in the project lifecycle, and with limited visibility into the full cost base. This is exactly the kind of problem that FinOps practices are designed to address by building structured accountability into cost management from the start.
What attribution models are used for shared data infrastructure costs?
Organizations use several attribution models to allocate shared data infrastructure costs to AI initiatives. The right model depends on the quality of available consumption data, the maturity of the IT financial management practice, and how precisely leadership needs to understand AI investment costs.
Direct consumption-based allocation
Where platforms support usage metering, costs are allocated based on actual compute hours, storage consumed, or pipeline runs triggered by the AI project. This is the most accurate model and the one FinOps practitioners recommend as a default because it ties spending directly to decisions made by the AI team. It requires tagging or labeling infrastructure resources at the workload level.
Proportional or activity-based allocation
When metering is unavailable, organizations estimate the proportion of shared infrastructure consumed by the AI initiative relative to all consumers. For example, if an AI project accounts for roughly 40% of data pipeline runs, it absorbs 40% of the pipeline’s operating costs. Activity-based costing applies a similar logic but uses cost drivers such as number of data records processed or number of transformation jobs executed rather than simple percentages.
Fixed allocation or cost pool distribution
Some organizations pool shared data platform costs and distribute them equally across all consuming projects or business units. This approach is administratively simple but the least accurate. It works as a starting point when consumption data is unavailable, but it can distort AI investment analysis by over- or under-charging individual initiatives depending on their actual usage patterns.
Whichever model you choose, the important thing is to document it, apply it consistently, and revisit it as AI workloads scale. Inconsistent attribution across projects makes it impossible to benchmark AI investment performance over time.
How does ITFM practice handle AI cost transparency?
IT Financial Management (ITFM) handles AI cost transparency by creating a structured framework that maps every cost element, including data preparation, to the AI initiative it supports. Rather than leaving AI costs scattered across departmental budgets, ITFM practice consolidates them into a coherent view that finance, IT leadership, and business sponsors can all trust and act on.
In practice, an ITFM approach to AI cost transparency involves several connected activities:
- Taxonomy alignment: Defining a consistent cost taxonomy that includes data preparation as a recognized cost category within AI initiatives, rather than treating it as generic IT overhead
- Cost layer mapping: Tracing costs from underlying infrastructure (compute, storage, licenses) through shared services to the specific AI products or use cases they enable, using frameworks such as the TBM taxonomy
- Showback and chargeback: Making AI teams aware of the full cost of data preparation they consume, whether through informational showback reports or formal chargeback mechanisms that create real financial accountability
- Forecasting and planning integration: Incorporating data preparation cost estimates into AI project business cases before work begins, so leadership approves investments with realistic total cost of ownership in view
- Reporting cadence: Establishing regular reporting cycles that track actual data preparation spend against budget, enabling early course correction when costs deviate from plan
Technology Business Management (TBM) frameworks extend this further by connecting AI investment costs to the business outcomes those initiatives are expected to deliver, making it possible to evaluate not just what AI costs but what it returns.
Should data preparation costs be capitalized or expensed?
Whether data preparation costs should be capitalized or expensed depends on whether the work creates a long-lived asset with future economic benefit, or whether it supports current-period operations. Under most accounting standards, costs that directly enable the development of an internally developed AI model may qualify for capitalization during the application development stage, while costs associated with preliminary research or post-deployment maintenance are expensed.
The practical distinction works as follows:
- Capitalize data preparation costs when they are directly tied to building a specific AI model that will be used beyond the current period, the project has passed feasibility assessment, and the costs can be reliably measured. Examples include labeling training datasets for a proprietary model or building a dedicated ingestion pipeline that only serves that model.
- Expense data preparation costs when they relate to exploratory or experimental work, when the data infrastructure serves multiple purposes beyond the AI project, or when the AI initiative does not meet the threshold for recognizing an internal asset.
Shared data infrastructure costs that serve multiple consumers present a gray area. Most organizations expense the shared portion and only capitalize the incremental costs that are directly and exclusively attributable to the AI development project. Your finance and accounting teams should align on a policy that reflects both your accounting standards and the specific nature of each AI initiative, because inconsistent treatment across projects creates audit risk and distorts investment comparisons.
What governance steps ensure consistent AI cost attribution over time?
Consistent AI cost attribution over time requires a governance structure that defines who owns attribution decisions, how those decisions are made, and how they are reviewed as projects and infrastructure evolve. Without governance, attribution approaches drift as teams change, tools are replaced, and AI initiatives scale in ways that were not anticipated at project start.
The following governance steps build lasting consistency:
- Define an attribution policy: Document the allocation models your organization uses for shared data infrastructure, the conditions under which each model applies, and who has authority to approve exceptions. Make this policy accessible to finance, IT, and AI project teams.
- Assign clear ownership: Designate a cost owner for each AI initiative who is accountable for tracking data preparation spend, validating allocations, and reporting actuals to finance. Without named ownership, attribution decisions default to whoever happens to be available.
- Establish a regular review cadence: Review AI cost allocations at defined intervals, typically monthly or quarterly, to catch drift early. As AI workloads grow, consumption patterns change, and allocation percentages that were accurate at project start may no longer reflect reality.
- Integrate attribution into project governance: Require that AI business cases include a data preparation cost estimate and a defined attribution approach before project approval. This creates a baseline against which actuals can be measured.
- Use consistent tooling and tagging: Enforce resource tagging standards across cloud and on-premises environments so that data preparation workloads are identifiable in cost reports without manual reconciliation.
- Connect to broader ITFM and TBM reporting: Embed AI cost attribution into your wider strategic portfolio management process so that AI investments are evaluated alongside other technology initiatives using the same financial framework.
Governance is not a one-time setup. As your AI portfolio grows and data infrastructure becomes more complex, your attribution framework needs to evolve alongside it. Organizations that treat this as an ongoing management capability, rather than a project-specific task, consistently achieve better financial visibility and more reliable AI investment decisions.
How we help you attribute and manage AI costs
At It’s Value, we help organizations build the financial management capability needed to track, attribute, and optimize AI investment costs, including data preparation, with the precision that leadership and finance teams require. Our approach combines ITFM and TBM frameworks with hands-on implementation support so you move from scattered cost data to structured, decision-ready insight.
Specifically, we support you with:
- Cost taxonomy design: Defining how data preparation and other AI cost categories fit within your broader IT cost structure, so every expense is visible and traceable
- Attribution model selection and implementation: Helping you choose and apply the right allocation model for shared infrastructure costs, and building the tooling and tagging standards that make it repeatable
- FinOps practices for cloud-based AI workloads: Applying FinOps disciplines to the cloud compute and storage costs that AI and data preparation workloads generate, so optimization decisions are made continuously rather than retrospectively
- Governance framework setup: Establishing the ownership structures, review cadences, and reporting processes that keep AI cost attribution consistent as your portfolio scales
- TBM integration: Connecting AI investment costs to business outcomes within a TBM framework, so you can demonstrate the value of AI spend to the board and to business stakeholders
If you want to bring structure and transparency to your AI cost management, get in touch with us and we will show you how to build an attribution approach that works for your organization.