AI Harness, Data Orchestration Layer and Unified Data Model: The Architecture Behind Trustworthy AI

Last week, I wrote about vendor lock-in and the "single pane of glass."

This week, I want to go underneath that glass.

A unified dashboard is useful. But once AI starts analysing information, updating records and triggering workflows, visibility alone is not enough.

Three pieces become essential:

• An AI Harness.
• A Data Orchestration Layer.
• A Unified Data Model.

They sound technical. The questions behind them are surprisingly simple:

• Does the information mean the same thing across our systems?
• Can we move it safely to where it needs to go?
• Who decides what AI is allowed to do?

Start with the operating model, not the chatbot

A client recently described their long-term target in refreshingly practical terms:

• A data warehouse owned by the firm.
• Best-of-breed tools for individual functions.
• An orchestration layer that moves data and workflows in both directions.
• The ability to replace individual technology vendors over time.

Notice what is missing.

No ambition to put everything inside one application. No requirement to commit permanently to one AI model.

The objective is to make the components work together without making any one component indispensable.

That is where these three architectural layers connect.

1. Unified Data Model: agree on meaning before generating answers

Putting data in one warehouse does not automatically make it consistent.

Consider a seemingly simple question:

"What is this family's total exposure to private investments?"

Before answering, the firm needs to agree on:

• Which accounts and entities belong to the family.
• How ownership percentages are applied.
• Whether exposure means invested capital, current valuations or commitments.
• How unfunded commitments are presented.
• Which valuation dates apply.
• Whether the same investment appears in multiple sources.

A Unified Data Model provides shared definitions and relationships so different systems can describe the business consistently.

It should not force every family office into the same rigid template.

The model needs to accommodate the firm's own households, entities, assets and reporting requirements.

The distinction is important:

A warehouse brings information together. A unified model gives it shared meaning.

Without that meaning, AI may produce a coherent answer to the wrong interpretation of the question.

2. Data Orchestration Layer: connect the work, not just the screens

A reporting layer helps people see information.

A Data Orchestration Layer coordinates how information and tasks move between systems.

The flow should work in both directions:

• Pull positions from a portfolio system.
• Match them to the correct household and entity.
• Check for missing or inconsistent records.
• Prepare an analysis.
• Route exceptions to the appropriate person.
• Write approved updates back to the CRM.
• Record whether the update succeeded.

This is more than having an integration available.

It means managing sequence, dependencies, permissions and failures.

If a source is unavailable, does the workflow stop or quietly use yesterday's data?

If an update fails, can it retry without creating duplicate records?

If two systems disagree, who resolves the exception?

These questions turn connected applications into a functioning operation.

APIs and protocols such as MCP can support tool connections. But connecting a tool is not the same as defining how its use should be governed.

3. AI Harness: give the model tools without handing it the keys

An AI Harness is the software environment surrounding the model. It manages how the model receives context, uses tools and moves through a task, alongside execution controls and monitoring.

Think of it as the difference between:

"Here is a goal. Do whatever you can."

And:

"Here is the objective, the approved information, the permitted tools and the boundaries within which you may operate."

For a wealth manager, the harness should help answer:

• Which household's information can this user access?
• Is this task read-only, or can it change records?
• Does the proposed action require approval?
• What happens when information is incomplete?
• Which model and policy versions were used?
• Can the workflow be paused or stopped?
• Can the firm reconstruct what happened?

A prompt saying "do not make unauthorised changes" is not the same as enforcing permissions.

The harness also should not be the only line of defence. Identity controls, databases and destination systems need their own enforceable boundaries.

Put the three together

Consider an illustrative request:

"Prepare a household concentration-risk review and record the approved follow-up actions."

Each layer has a different job:

• The Unified Data Model defines the household, ownership relationships and exposure calculations.
• The Data Orchestration Layer retrieves records, runs checks, coordinates the analysis and routes approved updates.
• The AI Harness governs how the model uses the information and which actions it may request.

AI can help interpret an unfamiliar document, configure a new analysis or draft an explanation.

But an established calculation does not need to be reinvented by a language model every morning.

My preference is straightforward:

• Use AI where interpretation or adaptation adds value.
• Turn approved recurring logic into tested, version-controlled code.
• Keep human review where judgment or material consequences require it.

Code is not automatically correct. It still needs testing, validation and monitoring.

The advantage is that the execution logic can be inspected, reproduced and governed.

Singapore is making accountability explicit

This is no longer only an architectural discussion.

On 7 October 2026, the Monetary Authority of Singapore issued AI risk-management guidelines addressing oversight, lifecycle controls and third-party AI risks.

The central message: financial institutions remain accountable for AI used in their services, including AI supplied by external providers.

The guidelines take effect in phases from October 2027, with further requirements taking effect in October 2028. They are Singapore-specific supervisory expectations—not a universal rulebook for every RIA or family office.

But the business implication travels well:

Buying an AI service does not mean outsourcing responsibility for its behaviour.

Inventories, testing, change management, monitoring and accessible evidence therefore belong in the operating architecture—not in paperwork added after deployment.

The uncomfortable link to vendor lock-in

Here is my concern:

A firm can become less dependent on its CRM and more dependent on the layer connecting everything else.

If one provider controls the shared data definitions, workflow logic, AI permissions and audit trail, the dependency has not disappeared.

It has moved.

So independence needs to be tested in practical terms:

• Can we retain our data mappings and business definitions?
• Can we export and maintain our workflow logic?
• Can we access the logs without the vendor's intervention?
• Can we change models without rebuilding every process?
• Can another qualified team operate the environment?

Hosting a warehouse in the firm's cloud can support control.

It does not automatically establish ownership of the code, portability of the workflows or independence from the provider.

Those depend on architecture, documentation, operating capabilities and contractual rights.

The real strategy

The next competitive advantage in wealth management may not be choosing the smartest model.

It may be building an environment in which different models can work safely on the firm's terms.

The three layers have distinct responsibilities:

• Unified Data Model: shared meaning.
• Data Orchestration Layer: coordinated movement and execution.
• AI Harness: bounded model behaviour.

Together, they offer something more valuable than another impressive demo: the ability to make AI useful without surrendering operational control.

The model can change. The applications can change.

The firm should retain control over the knowledge, workflows and evidence that make its business work.