One of the things that the Clairvoyint team is tired of hearing me say is “What does the AI do?” That usually comes up in the demo of a new vertical-oriented bundle or of a new platform capability. We’re an AI company, after all, but I don’t ask that question as a means of forcing AI into the solution for every problem Clairvoyint addresses. I ask it because I want to keep our allocation model at the forefront of everything we do.
A couple of months ago, on my personal blog, I wrote a post called “Interpretation and Ownership” that examined how modern generative AI may relate to and affect the Fitts allocation model commonly known as HABA-MABA. Since then, we’ve been using that topic as a guide for how we approach building Clairvoyint.
AI has a trust problem. This is especially true of LLMs. The topics of auditability and defensibility are increasingly written about in tech media. The problem of variance is baked into the inherently probabilistic nature of LLMs. The vast majority of users have come of age in a world where deterministic software has set the expectation that the same inputs should produce the same outputs. The fact that LLMs don’t behave that way is in their very nature. While this is not the “fault” of the LLM as much as it is a question of appropriate use, it does create a trust issue.
Which brings me back to Clairvoyint and the allocation model. We are building Clairvoyint on the principles of neuro-symbolic AI. I’ll dig into that more in a minute, but we’ve found that the allocation model is key to that approach. If the AI has a role in adjudicating ambiguity in the middle of analysis and managing the workflow accordingly, then the answer to “What does the AI do?” depends on how much ambiguity is in that workflow. This can obviously vary.
Neuro-symbolic AI is an incredibly academic term for a concept that pairs neural networks, which are what we encounter most commonly in the market today, with symbolic AI, which represents knowledge explicitly through rules, relationships, logic, constraints, or structured representations. In neuro-symbolic AI, the two concepts support each other.
One of the key ways that we use it in Clairvoyint is to recognize when parts of a workflow are inherently deterministic. The system decomposes those parts into traditional software and the allocation model is the guiding principle to ensure those parts are executed appropriately. This approach simultaneously helps us with token economics and variance and we’ve already seen benefits in both areas. So the first answer to “What does the AI do?” is “It helps figure out how much the AI needs to do.”
Symbolic methods help us deal with the trust issue in other ways as well. Core to Clairvoyint’s approach is an analytical framework based on the extensive experience of our Co-Founder, Nada Bakos. The framework was developed and refined through her experience applying it in high-value, mission-critical situations, and we have baked it in throughout the Clairvoyint architecture. Rules, relationships, logic, and constraints make up the framework and guide how Clairvoyint analyzes and interprets information. While neural methods help with interpretation and the reduction of ambiguity, symbolic methods guide how that happens, up to and including guidance such as “don’t do this part yourself, run this Python script instead.”
Additionally, we’ve implemented a source validation framework, also based on Nada’s experience. Clairvoyint will often need to reach out to external sources during the course of an analysis. In the best case, the user can specify sources that the system can treat as canonical for their use. Often, though, good analysis requires additional or more up-to-date context, which can mean reaching out to external sources such as the web. It is important that externally sourced information be classified and ranked for validity and suitability. Clairvoyint layers this approach across the system as well to ensure that results are produced using information that has been vetted for quality, accuracy, and provenance.
Our neuro-symbolic approach enables this three-legged stool of function allocation, analytical framework, and source validation to address increasing trust concerns around AI. It also has advantages when applying Clairvoyint to specific verticals. Geospatial analysis, which is core to Clairvoyint, is particularly compatible with symbolic methods because geospatial science is full of explicit rules, relationships, and constraints. Concepts such as DE-9IM, coordinate systems, topology, and many others give the system deterministic structure that doesn’t need to be relearned or approximated by a neural model.
By contrast, assessing geopolitical risk inherently involves subjectivity and nuance, especially in the field of gray zone analytics, where intent and motivations often need to be inferred from information that may be incomplete, ambiguous, or intentionally obscured. This is a case where “What does the AI do?” means that the AI may be carrying a heavier interpretive load to surface signals and developments that may be of interest to the user. The allocation model is a gradient, and Clairvoyint’s goal is to establish the right spot on that gradient for a given use case.
So when I ask “What does the AI do?” I’m really asking what parts of this problem benefit from probabilistic interpretation, and what parts are better handled by rules, software, or known analytical methods? For Clairvoyint, the goal isn’t to maximize the amount of AI in the system. It is to put AI where ambiguity requires it, constrain it where methodology matters, and get it out of the way where deterministic approaches are better suited to the task. We think that kind of deliberate allocation is one of the ways AI systems become more useful, more defensible, and ultimately more trustworthy.