An innovation leader at a Big 4 audit firm told me a number of years ago that he’d offered discounts to clients who could make their data audit ready. Audit by that measure meant 3+ on a 1 to 5 data maturity scale. Focus then was just on transactional data.
The vast majority of clients (remember, most were major public corporations) ranked only 1 or 2 on a 5 scale data maturity wise. So almost all clients just paid extra to have the audit teams assemble the necessary data. Some clients still provided data via spreadsheet.
Organizations haven’t been forced to focus on the data layer until recently. Leadership for the most part continues to be data layer averse.
That’s one reason why so much talk about “AI” and what to do about it merely focuses on agents, algorithms and what statistical machine learning teams think should happen with probabilistic-only AI and the pipelines of data they’ve selected and worked with.
Agents fail without the right information. For businesses to succeed with AI, they need to fix a deeply rooted information problem. This post details how a self-describing digital twin of each business can help.
How digital twins function optimally at the data layer
To meet this information challenge, enterprises each need to build a digital twin of their own business to guide their agents with. That digital twin can express the specifics of the business that are obvious to your human workforce, but need to be spelled out for agents in machine-readable, disambiguated language at the data layer.
That digital twin is the challenge. Not the agents. Agents in general aren’t novel. Organizations can rent agents, but they need to be hands on when it comes to knowledge engineering a digital twin.
Language model-enabled agents were novel when they appeared a number of years ago. Today’s LLMs are ubiquitous and therefore won’t really deliver competitive advantage on their own.
However, company-specific digital twins can provide competitive advantage by providing an agent-accessible knowledge foundation for LLMs. If there is no digital twin of the business to work with, giving language models agency will not help solve your business problems, because agents won’t know specifically how to act safely on your company’s behalf.
Building an AI-ready, audit-ready digital twin of the business
A company creates a semantic backbone in RDF (Resource Description Framework) to build a universal, machine-readable knowledge graph that represents how the business operates. By structuring enterprise facts as subject-predicate-object relationships, RDF allows companies to unify siloed data, establish shared business definitions, and feed layers of interrelated context directly to AI and analytics.
In other words, RDF can clear contextualized paths for your agents to find and use the information they need to complete the tasks expected of them.
A company starts building a digital twin of its business by representing the relatedness of real-world things—customers interacting with products, contracts, or employees, for instance—in machine-readable graph data as RDF triples. Each triple is a self-contained, self-describing fact: a subject, predicate, and object.
By the way, this isn’t the RDF of decades past, but a contemporary, more developer-friendly approach. Traditionally, creating an RDF knowledge graph required developers to learn complex semantic languages and rewrite their data from scratch.
YAML Ain’t Markup Language for Linked Data (YAML-LD) removes this hurdle by letting teams use standard YAML—the clean, human-readable format they already use every day for configuration and data management. By simply adding a small mapping shortcut (called a context) to a regular YAML file, the data automatically translates into a structured, audit-ready RDF graph behind the scenes.
Ultimately, YAML-LD makes RDFization simpler because it requires no new coding languages, cuts out visual mess like brackets and commas (such as in JSON or JSON-LD) and turns a complex data-engineering problem into a familiar, everyday task.
How AI ready becomes audit ready
It turns out that, when it comes to RDF, AI-ready also means audit ready. An RDF knowledge graph prepares your data for both audits and AI by turning messy, scattered information into clear, verifiable facts.
Using an RDF graph based on standardized structures and logic removes mystery from the process, helping algorithms run accurately and auditors to trace every data point back to its source.
If your audit data is self-describing in RDF, audit teams won’t have to sample that data. Sampling during an audit is costly and time consuming because it relies on human audit teams manually checking small, random samples for errors.
RDF knowledge graphs turn often cryptic, ambiguous data into highly structured, machine-readable facts using subject-predicate-object triples. Because machines can easily and reliably read this format, automated scripts can scan 100% of the dataset in minutes, replacing slow human sampling with instant, total analysis.
How to move safely between contexts with the help of knowledge engineering
Once the data is RDFized or triplified, a knowledge engineering team creates abstractions and necessary governance with the help of semantic standards that are each a part of the RDF stack. This is a critical, hand-on part of the process organizations should take charge of internally. The ideal result allows machines to work with twins across contexts. The process involves these standards:
- Schemas and n-dimensional models: RDFS and OWL define the classes and relationships between them. For example, when a customer buys a product from a merchant, that transaction exemplifies relationships between the buyer, the seller and the product that has brought the two together .
- A shapes and constraints language: SHACL enforces rules, allowing graphs to function as smart contracts that, for example, require a link to a valid customer. This builds one connected graph that mirrors the actual business, not a copy locked inside one application.
- A graph-traversing query language: SPARQL then queries that graph directly, answering questions like which customers touch which products. It can do this across departments or even company boundaries without extra integration work. The graph stays synced with reality as data updates.
The organizational stumbling blocks to digital twins at the data layer
Not only has leadership generally been data averse, the data they focus on managing is only their transactional data.
Roughly 80% of data does not fit the structure of traditional relational databases. AI requires access to this unstructured data—often referred to as “content,” or “knowledge” when managed by a Knowledge Management department—to operate effectively.
Organizations continue to struggle with fundamental data challenges. They are failing to manage their data comprehensively by including bringing unstructured data into the mix. They’re not dismantling silos, or otherwise creating the unified landscape of business information that agents need for their workflows.
However, the situation is not hopeless. There’s a clear opportunity: Build out the most essential parts of the data landscape that agents need. Craft a solution with broad utility applicable to all data types.. Organizations can achieve two goals simultaneously: they can prepare their data for AI readiness while simultaneously ensuring it is audit-ready–if they use the right method.
Other needs digital twins may need to deliver
Depending on the use case, digital twins may need to offer other functionality besides the semantic backbone and graph I’ve outlined. Some of those needs and requirements include these:
- Real-Time Telemetry (IoT)
- The Need: Text focuses on static/transactional data; lacks live feeds.
- Requirement: Integrate time-series data and high-velocity IoT sensor streams (e.g., emissions, energy use) to turn a static graph into a living model.
- Geospatial & Physical Layer
- The Need: Text views assets conceptually; lacks physical geometry.
- Requirement: Map assets to 3D models (BIM/CAD) and geographic space (GIS) for physical operations (ports, supply chains).
- Closed-Loop Actuation
- The Need: One-way data flow.
- Requirement: Enable AI agents to write back commands to physical infrastructure (e.g., autonomously adjusting grid traffic or battery storage).
- Predictive Simulation
- The Need: Text relies strictly on deterministic, historical facts.
- Requirement: Overlay probabilistic physics/simulation engines onto the semantic graph to forecast trends, equipment failures, or bottlenecks.
A broader view of AI + audit-ready data
Neither AI- nor audit-ready data is what most people think it is. It is not about fixing spreadsheets or collecting more training data. Instead, it requires building a machine-readable digital twin of your business to guide your AI safely.
The biggest problem with AI today is that it guesses, which leads to errors and erodes trust. You can solve this by grounding your models in verifiable facts using modern RDF knowledge graphs. When your data is structured this way, it serves two purposes: it gives your AI agents the specific context they need to act reliably, and it provides auditors with a complete, automated trail to verify every conclusion.
The successful organizations will not be those with the most intelligent agents, but those with the most precise, reliable and useful data foundations. The future of enterprise AI is not merely probabilistic; it is also deterministic, structured so it can be verifiable and trusted.
For more information:
Cheryl Dunn: Unlocking Agent-Based Accounting with a Semantic Approach
Leave a Reply