Enterprises are investing billions of {dollars} in AI brokers and infrastructure to remodel enterprise processes. Nonetheless, we’re seeing restricted success in real-world purposes, typically because of the incapacity of brokers to actually perceive enterprise information, insurance policies and processes.
Whereas we handle the integrations properly with applied sciences like API administration, mannequin context protocol (MCP) and others, having brokers really perceive the “that means” of knowledge within the context of a given businesis a special story. Enterprise information is generally siloed into disparate programs in structured and unstructured varieties and must be analyzed with a domain-specific enterprise lens.s
For example, the time period “buyer” might discuss with a special group of individuals in a Gross sales CRM system, in comparison with a finance system which can use this tag for paying shoppers. One division may outline “product” as a SKU; one other might signify as a "product" household; a 3rd as a advertising and marketing bundle.
Knowledge about “product gross sales” thus varies in that means with out agreed upon relationships and definitions. For brokers to mix information from a number of programs, they need to perceive completely different representations. Brokers have to know what the info means in context and methods to discover the correct information for the correct course of. Furthermore, schema adjustments in programs and information high quality points throughout assortment can result in extra ambiguity and incapacity of brokers to know methods to act when such conditions are encountered.
Moreover, classification of knowledge into classes like PII (personally identifiable data) must be rigorously adopted to keep up compliance with requirements like GDPR and CCPA. This requires the info to be labelled accurately and brokers to have the ability to perceive and respect this classification. Therefore we see that constructing a cool demo utilizing brokers may be very a lot doable – however placing into manufacturing engaged on actual enterprise information is a special story altogether.
The ontology-based supply of fact
Constructing efficient agentic options requries an ontology-based single supply of fact. Ontology is a enterprise definition of ideas, their hierarchy and relationships. It defines phrases with respect to enterprise domains, may help set up a single-source of fact for information and seize uniform subject names and apply classifications to fields.
An ontology could also be domain-specific (healthcare or finance), or organization-specific based mostly on inside constructions. Defining an ontology upfront is time consuming, however may help standardize enterprise processes and lay a robust basis for agentic AI.
Ontology could also be realized utilizing frequent queryable codecs like triplestore. Extra complicated enterprise guidelines with multi-hop relations may use a labelled property graphs like Neo4j. These graphs also can assist enterprises uncover new relationships and reply complicated questions. Ontologies like FIBO (Finance Trade Enterprise Ontology) and UMLS (Unified Medical Language System) can be found within the public area and is usually a excellent place to begin. Nonetheless, these normally should be custom-made to seize particular particulars of an enterprise.
Getting began with ontology
As soon as applied, an ontology might be the driving power for enterprise brokers. We will now immediate AI to comply with the ontology and use it to find information and relationships. If wanted, we are able to have an agentic layer serve key particulars of the ontology itself and uncover information. Enterprise guidelines and insurance policies might be applied on this ontology for brokers to stick to. This is a wonderful option to floor your brokers and set up guardrails based mostly on actual enterprise context.
Brokers designed on this method and tuned to comply with an ontology can stick with guardrails and keep away from hallucinations that may be brought on by the massive language fashions (LLM) powering them. For instance, a enterprise coverage might outline that until all paperwork related to a mortgage do not need verified flags set to "true," the mortgage standing needs to be stored in “pending” state. Brokers can work round this coverage and decide what paperwork are wanted and question the information base.
Right here's an instance implementation:
(Unique determine by Creator)
As illustrated, we have now structured and unstructured information processed by a doc intelligence (DocIntel) agent which populates a Neo4j database based mostly on an ontology of the enterprise area. A knowledge discovery agent in Neo4j finds and queries the correct information and passes it to different brokers dealing with enterprise course of execution. The inter-agent communication occurs with a preferred protocol like A2A (agent to agent). A brand new protocol referred to as AG-UI (Agent Consumer Interplay) may help construct extra generic UI screens to seize the workings and responses from these brokers.
With this methodology, we are able to keep away from hallucinations by imposing brokers to comply with ontology-driven paths and keep information classifications and relationships. Furthermore, we are able to scale simply by including new property, relationships and insurance policies that brokers can mechanically comply to, and management hallucinations by defining guidelines for the entire system slightly than particular person entities. For instance, if an agent hallucinates a person 'buyer,' as a result of the linked information for the hallucinated 'buyer' is not going to be verifiable within the information discovery, we are able to simply detect this anomaly and plan to eradicate it. This helps the agentic system scale with the enterprise and handle its dynamic nature.
Certainly, a reference structure like this provides some overhead in information discovery and graph databases. However for a big enterprise, it provides the correct guardrails and provides brokers instructions to orchestrate complicated enterprise processes.
Dattaraj Rao is innovation and R&D architect at Persistent Methods.
Learn extra from our visitor writers. Or, take into account submitting a put up of your individual! See our pointers right here.
