AI is changing the practice of tax law. This series examines the ethical, legal, and practical implications of AI across key areas of tax practice.
How should enterprises allocate the value resulting from AI enablement when people, data, and workflows are interdependent and reside globally?
In the last 16 months, companies have been scaling the use of AI to design and deploy AI within their business operations. This type of Enterprise AI is reshaping how companies generate productivity, insight, and competitive advantage, and unlike prior “ages” of digital transformation, Enterprise AI cuts across virtually all functions, can be deployed by non-technical staff, and scales rapidly. The off-the-shelf models offered by the major AI vendors can easily be adapted to company specific know-how through ordinary use. As a result, companies may end up with valuable internal AI capabilities without any formal product management or development processes, creating uncertain transfer pricing profit allocations.
Internal AI Capability Value
Companies are realizing AI value less through standalone tools and more by embedding AI directly into end-to-end business processes where work is already happening. The biggest gains come from integrating models with enterprise data and systems, then wrapping them in clear workflow steps including drafting, routing, approval, and exception handling so outputs translate into decisions and execution. When these AI-enabled steps are standardized, governed, and distributed improvements in speed, quality, and capacity compound across teams and geographies. This generates additional profit for the company that are difficult to attribute to particular jurisdiction or set of resources.
Workforce augmentation and capacity expansion. Automating repetitive analysis and judgment-supported tasks can enable professionals to focus on higher-value work. In practice, this automation often includes drafting first-pass analyses, summarizing large document sets, classifying or tagging provided data, and preparing recommendations that a human then validates and refines. The result is faster output with more consistent baseline quality.
Embedding productivity gains into daily operations means integrating AI into the tools and routines people already use such as email, spreadsheets, and knowledge bases, so the time savings accrue continuously rather than as one-off pilots. Over time, teams standardize prompts, templates, and review steps so AI becomes part of “how work gets done” and not a separate activity. This creates compounding benefits as practices spread across teams and geographies.
AI holds the promise of effectively allowing all members of a business’s workforce to exploit the know-how of the top performers in a given technical area, with a much lower investment in time and effort relative to creating extensive training manuals or process guidelines.
Given this potential to exploit the top performer’s know-how, companies should start to rationalize how to attribute intercompany profit to this activity to ensure accurate remuneration.
Innovation and discovery applications. Accelerating insight generation in areas such as drug discovery and risk modeling involves using AI to search broader solution spaces, identify non-obvious patterns, and prioritize hypotheses for human and experimental follow-up. For example, models can uncover candidate molecules, flag emerging risk indicators, or run rapid scenario analyses that would be infeasible manually at scale. Other applications of AI enable much more comprehensive and quicker searches for existing research or patents; AI can quickly compare a product idea to millions of filed patents and assist in determining where a company should direct its efforts to ensure its product is sufficiently differentiated. This type of effort would have previously required extensive human input. AI shifts research and analytics teams from spending time on exhaustive enumeration to spending time on validation, interpretation, and decision execution.
Given the geographic diversity of where people reside in contributing to these AI processes, proper steps should be taken to document the decision-making process such that profit attribution can be properly made.
Content and process generation. Marketing, invoicing, reporting, and recruitment workflows increasingly use AI to generate first drafts and standardized outputs focused on speed, consistency, and scalability. These systems can create tailored messages, populate structured documents, and assemble status reporting from multiple inputs, with humans providing review and approval where needed. When governed well, this reduces cycle time and variability while allowing teams to redeploy effort to higher-judgment work.
As shown in these examples, valuable internal process improvements are being realized by a global workforce: how should companies think about the allocation of such profits?
AI Similar to ERPs
As companies have been embedding AI into their processes, these same companies have been creating value when integrating within workflows, connecting to enterprise data, and governing/administrating it as scale. Rather than treat AI as a licensed platform that it uses for discrete tasks, companies that truly see value are embedding AI into its core processes that results in a quicker return on investment.
A helpful analogy is how organizations have deployed enterprise resource planning, or ERP, systems. Historically, companies procured third‑party ERP platforms and then invested heavily in integration, configuration, and controlled rollout to standardize core processes into a single source of truth, reducing costs, errors, and operational friction while improving decision‑making and controls.
Enterprise AI is increasingly being adopted in a similar way—as infrastructure accessed via third-party licensed models and managed platforms, where durable value is realized primarily through:
- integration with enterprise systems and data (connectors, retrieval, permissions, and data quality);
- configuration of how work is performed (prompts, guardrails, agents, routing, and human review steps); and
- operating decisions about who can change what (access controls, auditability, escalation paths, and model selection).
Where the analogy to ERP becomes incomplete is in how quickly AI creation moves outside traditional IT and how “configuration” can embed judgment and know-how. Unlike ERP customization, typically controlled by central system owners, formal change control, and specialized technical teams, many AI capabilities are created directly by finance, human resources, legal, or operations staff through prompts and agents. These activities resemble self-built tools (such as macros or low-code apps), but with significantly greater reach, adaptability, and reuse because they can be shared globally instantly and applied across many processes. As a result, the boundary between routine end-user operation and higher-value process design can be harder to define for AI than it was for ERP.
AI capabilities also tend to be used faster than ERP enhancements. A workflow or agent may originate as a local efficiency tool, but once it proves effective, it can be reused across countries and legal entities, often without the same level of centralized implementation discipline that accompanies ERP rollouts. Enhancements in one location can therefore generate benefits globally, creating transfer pricing issues as it relates to remuneration of these activities.
Factors Affecting Intercompany Pricing
Given the difficulty in measuring and tracking the costs and benefits of AI implementation relative to ERP, taxpayers will face practical challenges around deciding when AI-powered capabilities have become an intercompany transaction that requires transfer pricing to be applied, or when an AI-capability represents an intangible asset that requires a return commensurate with intangible exploitation.
Some considerations in setting up intercompany pricing are:
When is AI capability development treated as IP development versus service activities? The US transfer pricing regulations regard an intangible asset as existing when it provides value beyond routine services. Treas. Reg. §1.482-1(d)(3)(i); OECD Guidelines, Chapter VI, ¶¶6.54-6.55. For internal AI, the practical line is whether the activity produces an identifiable, reusable capability with ongoing usefulness—something the group can repeatedly deploy to generate benefits beyond the immediate task. When that threshold is met, the work begins to look like IP/intangible development; when it isn’t met, the activity is more naturally characterized as routine services.
Internal AI capabilities are often created through day-to-day use by operations personnel rather than through traditional, developer-led software projects. A finance team, for example, may standardize AI-enabled workflows so a smaller team can do the same job that a much larger team once did.
The US transfer pricing regulations look to the economic substance of the capability in determining whether it’s an intangible asset, rather than whether development costs were tracked or formal IP registration occurred. Treas. Reg. §1.482-1(d)(2). A related question is how to characterize the affiliates’ benefit from that capability once it exists: in most enterprise fact patterns, using an internal AI capability is closer to receiving a service (a functional output delivered through ongoing operation of the capability) than “access” to an intangible asset that can be independently exploited by the recipient. This is consistent with the definition of a service under both the US transfer pricing regulations and the OECD Guidelines. Treas. Reg. §1.482-9(b); OECD Guidelines, Chapter VII, ¶¶7.6–7.10.
In general, AI work starts to look like IP/intangible development when teams build something reusable and scalable that captures proprietary know-how and can be rolled out across the group, for example:
- a standard agent design or workflow that can be deployed in multiple countries;
- durable workflow logic/toolchains and evaluation artifacts that materially improve output quality; and
- curated, labeled, or otherwise enhanced datasets that improve performance.
By contrast, the work is more naturally a service when it’s focused on operating, supporting, and administering the AI-enabled process without creating a transferable capability, for example: day-to-day use within an established playbook, human review and exception handling, access administration and training, monitoring for issues, and routine prompt/model/workflow updates.
How should value be attributed to all the stakeholders? Value attribution in AI model building will be where much scrutiny is focused because the “asset” is almost never just the model. From a transfer pricing perspective, the question is which entity controls the key development, enhancement, maintenance, protection, and exploitation, or DEMPE, decisions and puts capital at risk, such as build vs. buy, which use cases matter, what “good” looks like, how to define the data, how models are tested and constrained, and who has release authority—if engineering work is spread across countries. Treas. Reg. §1.482-1(d)(3)(iii)(B); OECD Guidelines, Chapter VI, ¶¶6.32-6.38. Companies may follow a similar approach to the transfer pricing of other intangibles and separate the value created by AI across the contributing functions, assets, and risks, documenting who owns and controls each one. This also helps separate returns tied to third parties (licensed foundation models, cloud) from any residual return tied to group-created capability.
While measuring value for this new technology may be uncertain, when it comes to allocating value across jurisdictions, the traditional transfer pricing methods are still applicable. In scenarios where it’s extremely difficult to determine where AI value was generated, a profit split model may be most appropriate. When a central entity maintains strict control over AI development and offers use to its affiliates, that entity might charge its related parties based on the number of tokens used.
Depending on the facts, the outcome can range from cost-based service charges for routine build/run work and a shared-service allocation for central enablement to a residual profit split to reward the entities performing the non‑routine DEMPE functions.
As is the case with any business changes, documentation matters—information such as governance materials, decision logs, funding approvals, evaluation artifacts/model documentation, and change-control records that show who actually made the economically significant calls.
Path Forward
Enterprises need to treat AI as an operational capability that is actively governed, not just adopted. That means establishing clear governance over how AI is created, deployed, and reused across the organization so that productivity and quality gains are traceable to use of specific capabilities.
Enterprises must also define who owns key responsibilities: decision rights over design and change, accountability for risk, and stewardship of ongoing enhancements as workflows evolve. From a transfer pricing perspective, DEMPE can still be a useful lens, but it should not automatically convert ordinary AI usage into “IP development.” Many activities will be better characterized as routine while only some activities rise to the level of creating or enhancing an AI-capability intangible.
Finally, enterprises need disciplined documentation that explains how AI-enabled value is generated and scaled so benefits can be governed, audited, and attributed consistently across teams and legal entities. In practice, that documentation should evidence who performs and controls the day-to-day AI-enabled services, and where any non-routine developments or enhancements occur. This helps support the appropriate characterization (services versus intangible-related returns), clarifies decision rights and risk ownership, and strengthens the basis for any intercompany charging model.
An immaterial amount of this content was drafted by generative artificial intelligence.
This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.
Read More on BLaw:
In Focus: Artificial Intelligence (AI)
Author Information
Samit Shah is National Transfer Pricing Leader and Mat Knudson is a Transfer Pricing Manager in Grant Thornton’s Transfer Pricing Group.
Grant Thornton LLP and GT Advisors (and their respective subsidiary entities) practice as an alternative practice structure in accordance with the American Institute of Certified Public Accountants Code of Professional Conduct and applicable law, regulations, and professional standards. Grant Thornton LLP is a licensed independent CPA firm that provides attest services to its clients, and GT Advisors and its subsidiary entities provide tax and business consulting services to their clients. GT Advisors and its subsidiary entities are not licensed CPA firms.
Interested in writing? Review our author guidelines, and submit pitches to Insights@bloombergindustry.com.
