High AI Agency and Low Time to Value are Not Mutually Exclusive
The “best practice” architecture options for enterprise agentic systems are starting to take shape, but our understanding of their ramifications and tradeoffs are still a work in progress. A recent analyst report got me thinking specifically about whether better context building and data integration equates to longer “time-to-value” for AI projects. They argue it does. I don’t think so, and I believe that Kamiwaza is a great demonstration of why.
What Is “Time-to-Value"?
I should begin by defining what I mean by “time-to-value.” In this context, time-to-value represents how long it takes to see real business value once a project has started. The project may be a production application providing key business value, but in this case the time-to-value I really care about is the time it takes to go from agreeing to evaluate a technology to seeing it deliver something the business deems valuable. That “something” may just be proof that a product or service can deliver on its promises.
In this context, time-to-value is affected by many variables. For example, the quality of the product or service, its applicability to the problem at hand, the skill set of the people working on the project, and so on. Perhaps the most important factor for the enterprise is that users trust the results.
For any evaluation, one key variable is setup time: the time it takes to get the product or service installed, configured, pre-loaded with data, and so on. For agentic orchestration, this setup time varies greatly, in part due to the complexity of the environment, and in part due to the sheer time it takes for set up processes to execute.
For example, workflow based systems require not only installation and configuration, but the time it takes to define, implement, and test those workflows. Meanwhile, online frameworks, like those offered by major AI providers, might be much faster to get running, but with the tradeoff that they may be more limited and expensive to use.
The argument made against semantic- and ontology-based approaches is that the set up time required to define ontologies, ingest data, and “tune” the indexes is significant, and can add up to much more time than what other options require.
But Kamiwaza thinks differently about this.
AI as an Answer to an AI Problem
One of the things that our founders thought of from day one is how AI itself can be part of the solution of building ontology-aware AI orchestration systems. There are several parts to this approach, but key to this is that with some basic guidance, AI is actually quite good at building indexes for itself. Models can be used to quickly evaluate existing data stores and build both semantic and ontological indexes to be used in context building for agents. Basic embedding and augmentation approaches can be used to not only retrieve data, but to first validate what data is relevant to the problem at hand and tune retrieval to just that data.
Kamiwaza takes full advantage of the automation available through agentic AI to build a comprehensive picture of an organization’s data portfolio, and to use that picture to fine tune how context is built when inference takes place. Data ingestion requires putting the right data connectors in place, pointing them to the right data sources, and initiating the ingestion process. And when inference is initiated, Kamiwaza validates that agents and users have appropriate access to the data, and then uses metadata, semantic, and ontological indexes to build the right context for the response.
There are two key consequences of this approach. First, AI can run ingestion processes 24 hours a day, seven days a week, 365 days a year. Quickly. At scale. Second, with the right prompts in the right loop, AI has the wonderful ability to be self critical. And through that self-evaluation, it can learn. It can adapt to every change in the business, because it adapts to every change in the way data represents the business.
The Human in the Loop
It must be said, however, that this isn’t just magic. There is no level of artificial intelligence capable of getting your business modelled correctly from scratch (yet). Kamiwaza has discovered that the quality of AI-built ontologies increases greatly if it is given a “seed” ontology that defines a basic framework on which the AI can build. And that ontology should reflect the critical relationships between data concepts that are most important to the business.
How is that any different from other products that require you to build ontologies to enable their key data features? Mostly this: that framework can be surprisingly minimal, and is probably the “easy part” of ontology building for any enterprise. Take a few minutes to agree on some core concepts (maybe you already have!) and the AI can do the rest in hours. It completely changes the time commitment required for creating business ontologies.
Humans are also necessary to provide feedback to the AI models and agents whenever work is done. Did the right action get taken? Was the analysis in that report correct and useful? Was access control defined and applied correctly?
(This last one is actually quite tricky in multi-agent environments, as the question becomes not only whether or not data should be accessed, but whether results containing that data should be allowed outside of a given jurisdiction, whether the human user can see the specific data used in calculating an aggregate response to a prompt, and so on. This is why we applied Relationship-based Access Control (ReBAC), to inferencing.
The key point is that humans must still play a role in the definition, management, and oversight of agentic systems, but we believe AI can automate most if not all of the toil required to make semantic and ontological systems trustworthy and effective.
Time-to-Value with Kamiwaza
The result of Kamiwaza’s approach is that we have been able to spin up prototypes of enterprise applications in a matter of hours or a few days, and to turn those applications into production systems with just a bit more testing and tweaking. This is no different than the time it would take to make a completely goal-oriented approach production ready.
Furthermore, once you get the semantic and ontology index flywheels going—assuming you have those adaptive feedback loops—the effort required to make the next system use your data efficiently and effectively goes down dramatically. So the “next time-to-value” is significantly less than the initial work. As your ontologies get stronger and stronger, your agentic systems find it easier to use existing data and integrate new data, which leads to a more comprehensive ontology, and so on.
With goal-oriented systems, the AI relies on memory to improve its knowledge of the business, which is generally quite unstructured and random. Each new agent has to relearn all necessary data relationships for its goal from this loose set of notes. This works great for some greenfield applications, but in the enterprise where data is shared across divisions and even between companies, this is a huge time sink, and a major use of precious tokens. Kamiwaza avoids all of that as your business scales.
To find out more about how Kamiwaza brings structured knowledge of your business to AI without an excessive amount of preparation and design, contact us. We’d love to hear from you.
And, as always, I write to learn, so tell me your thoughts in the comments below. I’d love to hear from you.
