If you visit the Clairvoyint website, it says that we turn expert analytical methodology into reusable agents that can perform analysis and produce reports, briefs, and assessments while preserving the evidence behind their findings. That’s a reasonably concise description of the product. But what does it actually mean?
The answer goes back a couple of years, to a consulting engagement focused on the demographic hourglass. The Baby Boomer retirement wave was approaching its peak, Generation X was comparatively small, and younger workers were beginning to inherit responsibilities from people with decades of accumulated experience. The purpose of the engagement was specifically to understand how AI might support knowledge work through that transition.
I wasn’t examining whether AI could simply automate some of the work those experienced people performed. It was more about whether AI could augment the people coming behind them. Subtle knowledge was walking out the door at an increasing rate and no amount of PDFs or Word documents was going to effectively transfer that knowledge. The real question was whether a runnable, probabilistic technology could help preserve some of the nuanced understanding so that it could be assumed by incoming workers. That experience is something I keep returning to.
It helps explain the team we’ve assembled at Clairvoyint. The company was founded by me, Tee Barr, our Chief Product Officer, and Nada Bakos, our Chief Analytics Officer. Between us, we bring backgrounds in geospatial and analytical systems, product development, strategic operations, and intelligence work. Those perspectives are important because the problem we’re working on isn’t fundamentally a software problem. It’s about how people approach difficult analytical work, and how more of that capability can survive beyond the individual expert.
The demographic problem is fairly straightforward. As experienced workers retire, organizations don’t simply lose headcount. They lose accumulated experience. An engineer, intelligence professional, geospatial practitioner, underwriter, or consultant carries around a lot of knowledge that probably isn’t in a process manual. They can determine which source to trust when two sources disagree. They recognize the exceptions that invalidate the usual rule. They understand how the distinction between signal and noise varies by use case. They often possess knowledge that doesn’t neatly externalize into a document, script, or database and that accumulated experience is the key constriction in the hourglass.
Organizations are entering a substantial retirement wave while succession planning and formal knowledge transfer remain uneven. The organization that retained me for that engagement recognized that. It’s not simply about filling open seats, but also about transferring capability across generations quickly enough.
AI introduces interesting possibilities. What if, instead of focusing only on automating an expert’s work, we used AI to help preserve and extend the value of their expertise? Not by pretending that thirty years of experience can be poured into a model, but by helping experts make more of what they know explicit, accessible, repeatable, and useful to the people who follow them.
Working through the problem in public
That question has been underneath a lot of my writing on geoMusings this year. I’ve been using this blog to work through ideas we were already exploring at Clairvoyint, but the traffic runs both ways. Product work gives me concepts to examine, and writing about them forces me to push on the assumptions and sharpen them before some find their way back into the roadmap.
In AI Still Requires You to Understand Your Business, I argued that AI doesn’t eliminate the need for domain expertise. Someone still has to understand what the process is supposed to do and, critically, what correct looks like.
In Post GIS Revisited and Shortening Translation Distance, I explored what happens when implementation becomes easier and experts can work closer to the language of their actual problem. Value shifts from knowing how to make the software perform an operation toward articulating the business logic and expected outcomes behind it.
That’s particularly apparent in geospatial work. Spatial joins and distance calculations can tell you how two things relate spatially, but they can’t quantify the significance of those relationships.
Plausibility Is Not Provenance worked through the concept that AI can produce answers that are polished, coherent, and wrong. A map can do the same thing. If the source data is wrong, the scale or temporal context is mismatched, or even if the wrong datum transformation is used, an authoritative-looking, but misleading, map can be the result. The evidence trail has to remain inspectable all the way to the conclusion.
Then, in Does Your Workflow Need a Model? I argued that AI can help us understand and formalize a process without necessarily belonging to every part of its execution. Once something can be represented reliably as a rule, threshold, validation, spatial operation, calculation, or transform, deterministic software may be the better tool. That concept is the basis of Clairvoyint’s “deterministic first” approach. Known things should become known things. Models should be reserved for the places where inference actually adds value.
Interpretation and Ownership gets closest to what we’re building. One of the core questions I repeatedly ask the Clairvoyint team is “What does the AI do?” That answer can vary widely depending on the use case. Our bet is that most analysis relies on a mix of determinism, probabilism, and judgement calls. AI can increasingly interpret information, synthesize evidence, and execute significant portions of analytical workflows, but someone still has to decide what problem matters, what evidence should count, and who owns the consequences of acting on the result. It’s less about human work versus machine work than about where different kinds of judgment belong. AI can interpret output and make a judgment call on tool selection, but people still need to apply expertise to set the rules of the road and assess the outcomes.
All of those posts explored different aspects of the question at the core of Clairvoyint. Rather than asking how much of an expert’s work can be automated, we ask how much of an expert’s methods can be made lasting?
The methodology is the valuable part
One of the foundational things we’re doing at Clairvoyint is baking Nada’s and Tee’s knowledge about and approach to complex problem solving into the system.
How do you approach an ambiguous problem? Which evidence matters? Which sources outrank others and why? What causes you to question an initial conclusion? Which parts can become rules? When does geography change the interpretation? Where does judgment remain necessary? What should trigger escalation?
If we captured only conclusions, we’d have a knowledge base. If we automated only the sequence of steps, we’d have a workflow system. We’re after something deeper, which is the methodology by which an expert moves from evidence to a defensible conclusion. Some of that methodology can become explicit rules, thresholds, validations, schemas, calculations, or transformations. Some consists of source preferences, evidence standards, exceptions, and assumptions. Some remains judgment and Clairvoyint is being designed around those distinctions.
Clairvoyint’s Concierge is an agent that works with experts and the materials they use to pull structure out of what they know. Requirements, interpretation rules, thresholds, exceptions, source preferences, assumptions, output expectations, and human-review points can become explicit rather than remaining part of institutional memory.
Conversation in the classic AI chat sense is important because it shortens the translation distance. Experts can recognize their own method as it takes shape and decide what can safely be formalized. Clairvoyint doesn’t force everything into a rule. If something can be represented reliably as a deterministic operation, it should be. If it genuinely requires judgment, pretending otherwise makes the system less trustworthy. This line is still fuzzy, but comes into better focus with each iteration.
Geospatial fits naturally into that approach. Sometimes geography is data. Sometimes it’s a deterministic operation like a coordinate transformation. Sometimes it provides the context that connects otherwise unrelated evidence, such as an implicit spatial relationship versus an explicit primary/foreign key relationship. Often the spatial pattern itself requires interpretation. This is especially true in the field of human geography, which I’ve also explored a lot this year.
Nada and Tee bring a wealth of knowledge in the practical application of these concepts across a range of topics from intelligence to climate risk. They are helping to make our approach real, shaping how we can take the way experienced people approach difficult analytical problems, separate what can be externalized from what still requires judgment, and preserve the transferable parts without flattening the nuance that made the method valuable. Our ultimate goal isn’t software that thinks like Nada or Tee, but a platform that lets organizations do this with their own experts.
The conversation isn’t the artifact
At Clairvoyint, conversation is an authoring interface, but it isn’t the primary output. What comes out of that process is what we call a Bundle. That’s a somewhat unglamorous name for an inspectable, versioned representation of an analytical agent containing its methodology, evidence references, rules, schemas, scripts, templates, assumptions, and other material needed to perform a particular kind of analysis.
The Bundle can evolve as the methodology evolves. Rules have provenance. Evidence can be superseded without erasing history. Repeated use can expose gaps and exceptions that feed back into the method. The Bundle also makes clear that the agent is more than just the model/LLM. It is the combination of method, rules, evidence, tools, state, constraints, and model reasoning needed to perform the work. Depending on the problem, those tools may include databases, documents, statistical models, GIS, imagery, APIs, or other computational systems and AI is simply one component inside that system.
Clairvoyint’s goal isn’t to create an AI that gives an impressive answer once. It is to turn a good analytical method into reusable infrastructure. If an expert helps build that method inside Clairvoyint, the capability no longer exists only because that person happens to be available. Other analysts can use it, inspect it, challenge it, improve it, and trace its outputs back through the method to the underlying evidence. Over time, the analytical capability can survive the people who originally built it.
The hourglass meets AI
None of this means experience itself can be automated. Some expertise may never externalize cleanly. Conditions, sources, and geography change and methods that worked five years ago may not apply tomorrow. Ownership is still important. Someone has to decide what the analysis is for and what to do about it. That’s why provenance, memory, versioning, deterministic execution, and human review are part of the architecture rather than features wrapped around an AI model.
Much of the anxiety around AI has focused on machines acquiring skills that previously belonged to people, but another transition is happening at the same time. Organizations are losing a large amount of accumulated human experience just as machines are becoming dramatically better at absorbing information, executing procedures, connecting different forms of evidence, and assisting interpretation.
At Clairvoyint, we see that as an opportunity. If AI primarily replaces tasks, we may get a substantial productivity gain, but if it can also help experienced people expose how they approach difficult problems, we may get something more valuable. We may get a mechanism for carrying more institutional capability and supporting the people who execute it.
That’s what we mean when we say Clairvoyint turns expert analytical methodology into reusable agents and that’s the problem we’re solving at Clairvoyint.
Header image: André Jacob Roubo, Public domain, via Wikimedia Commons
