The Hidden Cost of AI-Led Digital Asset Architecture
Why unverified AI advice can leave digital asset businesses with operating structures that are harder to repair than to build correctly from the beginning.
In recent advisory engagements, I have worked with digital asset projects and smaller investment firms that had already relied heavily on generative AI to make foundational business decisions.
Their use of AI had gone well beyond research. Recommendations generated by these systems had influenced operating models, infrastructure choices, internal procedures and, in some cases, decisions requiring considerable financial and regulatory expertise.
By the time we became involved, much of that work had already been implemented.
Our challenge was not simply to recommend a better structure. We first had to understand which earlier decisions were sound, how the different components depended on one another and what could be changed without creating further complications.
In practice, correcting these arrangements proved considerably more difficult than designing an appropriate operating model from the beginning.
Those experiences have made me think more carefully about how AI is being used in institutional digital assets.
I use AI myself. It is an effective research and analytical tool, particularly when exploring unfamiliar subjects, comparing alternatives or organizing complex information. Yet the quality of an AI-generated recommendation should not be confused with its suitability for a particular financial business.
That distinction becomes critical when the recommendation starts determining how an institution actually operates.
When a Convincing Answer Becomes a Business Decision
Generative AI can produce remarkably comprehensive responses. An answer may describe custody arrangements, regulatory considerations, technology providers and implementation steps in language that appears technically sophisticated.
Nothing in its presentation necessarily reveals that an important assumption is missing.
Consider a small investment firm preparing to enter digital asset markets.
Before selecting a custody provider or transaction infrastructure, the firm needs to establish the nature of its activity. Will it invest its own capital or manage assets for others? Which entity will legally own the assets? Who can authorize transfers? What controls govern those decisions? How will transactions be recorded and reconciled?
These are foundational operating questions. Their answers determine which technical arrangements may be appropriate.
An AI system can explain institutional custody models, compare service providers and propose authorization procedures. However, a recommendation built without a complete understanding of the firm’s circumstances may combine individually reasonable components into an unsuitable structure.
This is where industry experience becomes particularly valuable.
A professional familiar with financial operations may identify a fundamental weakness before the technical discussion progresses. Someone entering the sector for the first time may interpret the same detailed response as evidence that the necessary analysis has already been completed.
Better prompting certainly improves the quality of AI assistance. It does not guarantee that the underlying assumptions are correct, or that the person evaluating the response can recognize what has been overlooked.
Once these recommendations become operational decisions, the consequences extend well beyond the original AI interaction.
Why Repairing the Wrong Structure Is So Difficult
The most demanding part of the advisory situations I encountered was dealing with decisions that had already become embedded in the business.
When a project begins with a blank page, its legal structure, service boundaries, controls and infrastructure can be developed in a logical sequence.
An established system offers much less flexibility.
Technology providers may have been contracted. Internal procedures may depend on the original architecture. Employee responsibilities, transaction permissions and customer arrangements may all reflect assumptions that were never adequately tested.
Changing a foundational decision can affect several of these elements simultaneously.
Take custody as an example. Replacing an unsuitable custody arrangement may require more than migrating assets to another provider. Transaction approval rules, signing permissions, reconciliation procedures, recovery arrangements and contractual responsibilities may also need to be redesigned.
Tokenization presents a different set of possible complications.
A project might successfully issue and transfer tokens while leaving unresolved the legal rights represented by those tokens, the authoritative record of ownership or the relationship between onchain transfers and existing financial records.
The technology can function correctly even when the surrounding financial arrangement is incomplete.
These examples illustrate the types of structural dependencies an institutional review must examine. The underlying problem is that an early architectural mistake can spread into processes that were subsequently built around it.
Over time, this creates architectural debt.
In financial services, such debt is not necessarily confined to software or operational inefficiency. It can involve contractual commitments, external counterparties, regulatory obligations and client assets.
Some problems are manageable through targeted adjustments. Others require substantial reconstruction. Where irreversible transactions or established legal rights are involved, certain consequences may be impossible to eliminate completely.
The original implementation may have taken only a few months. Understanding how to unwind it safely can become a far more demanding exercise.
Digital Assets Still Have to Work Within Finance
Part of the difficulty comes from the relationship between emerging technology and established financial infrastructure.
Blockchain networks, smart contracts, tokenization and stablecoins introduce capabilities that can improve particular financial processes. Their technical functionality, however, does not automatically resolve the legal and operational requirements surrounding those processes.
A token transfer may be technically complete without settling every associated legal obligation. A stablecoin payment may reach the recipient while questions about redemption or access to bank money remain. A private key may be securely stored even though responsibility for authorizing transactions is poorly defined.
Financial institutions must account for these distinctions.
They need to establish where assets are held, which records are authoritative, when settlement becomes final and how transactions flow through treasury, accounting and risk management.
These requirements are sometimes treated as obstacles inherited from traditional finance. In reality, they are central to how institutions maintain control and accountability.
The Financial Stability Board’s October 2024 report on tokenisation examined vulnerabilities that could become more significant as tokenized financial activity expands, including operational fragilities and interconnectedness. It also recognized potential efficiency benefits, while noting that the scale of adoption remained limited at the time of its assessment.
The lesson for businesses designing digital asset operations is practical.
A new technical layer must be evaluated through the financial activity it supports.
Who initiates the transaction? Who approves it? Where does the money move? Which party assumes liability? How are records reconciled? What happens if the normal process fails?
Those questions reveal weaknesses that may remain invisible in a technically successful demonstration.
An operating model should therefore be established around the institution’s actual obligations and commercial requirements, rather than assembled from a collection of attractive technology features.
Giving AI the Right Responsibility
Excluding AI from digital asset development would be an unnecessary restriction.
Used by knowledgeable professionals, it can help compare infrastructure options, map dependencies, prepare preliminary requirements and identify potential failure scenarios. It is particularly useful for challenging an existing proposal and examining alternatives before resources are committed.
The problem begins when this supporting role becomes decision-making authority without adequate verification.
A company might ask AI to design its entire digital asset operation, receive a detailed implementation plan and proceed directly to execution. What may be missing is an independent assessment of whether the proposed business activity is commercially viable, legally permissible in the relevant jurisdiction and operationally sustainable.
That work requires more than checking whether the response contains obvious errors.
The NIST Generative Artificial Intelligence Profile, published in July 2024, provides a broader framework for identifying and managing risks associated with generative AI systems.
For institutional digital asset businesses, a useful application of this approach is to distinguish between research assistance, recommendations requiring independent validation and decisions that demand accountable professional judgment.
AI can help an experienced team ask better questions. It can also accelerate work that would otherwise consume substantial time.
Its limitations become more consequential when users lack the background to challenge a plausible answer.
Professional expertise includes recognizing missing information, testing a proposal against real financial workflows and understanding when a particular decision requires specialist legal, compliance or technical review.
A detailed response cannot establish those judgments on its own.
The Work That Should Happen Before Implementation
My experience with these advisory engagements has reinforced the importance of sequencing.
A business planning to enter institutional digital assets should first establish what financial activity it intends to conduct, which customers it expects to serve and what commercial problem the proposed operation will solve.
Infrastructure selection follows that assessment.
Starting with a preferred blockchain, custody platform or token model can reverse the process. The organization may then begin adapting its business around a technology choice rather than selecting technology to support a clearly defined business.
A sound operating assessment also needs to determine which established financial processes must remain in place.
Sometimes an existing service provider is more appropriate than a custom architecture. Conventional reconciliation procedures may still be necessary. A proposed tokenization model may introduce costs and dependencies that outweigh its benefits.
Recognizing these limitations is part of competent institutional design.
The required expertise may come from several disciplines. Market structure, custody, financial operations, compliance, legal interpretation and technology implementation overlap, but they are not interchangeable.
An experienced advisor should recognize where additional specialists are needed and where a simpler approach would serve the business better.
This matters because advisory value is not measured by how elaborate the recommended architecture becomes. A decision to avoid unnecessary infrastructure can be just as valuable as a successful implementation.
Building Has Become Easier. Evaluating What Should Be Built Has Not.
Generative AI has made sophisticated business plans, technical specifications and implementation proposals accessible to a much wider audience.
That is a meaningful advantage, particularly for smaller companies with limited research resources.
But producing a convincing plan and determining whether it should be implemented remain very different capabilities.
In institutional digital assets, the gap between them can be expensive.
An organization may believe its system is complete because the components have been selected, integrated and tested. Weaknesses may only emerge when a banking partner examines the arrangements, a compliance review exposes missing controls, real transactions require reconciliation or an operational failure demands recovery.
At that point, changing direction may involve significant costs and existing obligations that cannot simply be set aside.
The situations I encountered in advisory work demonstrated how difficult this becomes once earlier assumptions have shaped the entire operation.
I do not see that as an argument for reducing the use of AI. The technology can make experienced professionals more effective and help organizations examine possibilities that might otherwise remain unexplored.
But institutions still need people who understand how those possibilities interact with financial markets, regulatory responsibilities and operational reality.
The timing of that expertise matters.
Involving experienced professionals after an unsuitable system has been implemented is very different from involving them when its foundational decisions are still open.
The first approach may require extensive remediation. The second creates an opportunity to avoid the problem altogether.
And in the projects I have worked on, that difference has been substantial.