Vertical AI and the operator-knowledge moat.

This is the working operating thesis on Vertical AI as it has been built from inside the verticals at NXT Companies, rather than from a deck. It is the position that informs how we enter new industries, how we sequence software against operations, and how we deploy AI inside our portfolio companies. It is written from the perspective of an operator first and an investor second, because that is the sequence in which the conclusions were earned.

The thesis, in brief: the durable moat in Vertical AI is the operator knowledge captured inside the customer's organization — structured, permissioned, and made queryable — not the model, the orchestration layer, or the workflow embedment. Everything else commoditizes on a twelve-month cycle. That asset compounds.

1. The vSaaS foundation: operator first, software second

True Vertical SaaS is not a clever feature set. It is a feature set built on real-time insight into how an industry actually behaves on a Tuesday afternoon when something is going sideways — lead generation, sales cycles, vendor relationships, payment dynamics, regulatory edge cases, fulfillment failures. No engineering team, regardless of pedigree, can produce that fluency from observation. It has to be earned from participation.

At NXT, when we enter a new vertical we become an operator first. A retailer, a distributor, an air charter operator, a building-products supplier, a metals service center. We build software for our own teams before we build it for the market. That sequence — operator first, software second — is non-negotiable. The result is software that carries the weight of authentic operational understanding, and customers who stay because they recognize their own world inside the product.

The reference frame here is Mark Leonard's letters at Constellation, where customer attrition runs roughly four percent and average customer tenure approaches twenty-six years. That number transforms a software outcome into a relationship outcome. The software is just the instrument through which the relationship is monetized. Constellation's portfolio is broad and shallow on technology and deep on customer fluency. Ours is narrower and AI-native, but the structural logic is identical.

The most durable software companies are starting to look less like technology companies and more like firms — where institutional knowledge is the primary asset, and technology is the mechanism through which that knowledge is delivered at scale.

2. The wedge: unstructured inbound and judgment-intensive workflows

The clearest expression of the thesis in production is the SmartInbox™ module across the NXT portfolio. The initial deployment was inside Levo, our private air charter platform, where the sales office was processing thousands of inbound emails per day, the vast majority of which were broker requests with no real revenue path. The signal was buried under correlation problems — fleet availability, repositioning economics, traffic patterns, broker quality, time-of-day pricing — that no human triage layer could solve at speed without losing accuracy.

SmartInbox™ is an AI-native triage and prioritization layer that learns the operator's actual judgment patterns and surfaces the requests that deserve attention. In Levo, it sits in front of LevoAI™ Matching, which scores aircraft against requests using a proprietary Probability of Confirmation model. The pair has moved Levo from a charter brokerage running on human triage to the first AI-native operating system in private aviation. It is now in production, driving measurable revenue capture, and — critically — the underlying module has been abstracted and adapted across other verticals where the same pattern recurs: unstructured inbound, multi-variable correlation, human-in-the-loop prioritization.

Reusability is the part that matters to the model. The same triage architecture now operates against window installation requests at Every Window and RFQ packets at Equipio. One module is being used in different industries, each with vertical-specific training data on top of a shared core. That is the modular scalability that makes the wedge defensible at a portfolio level rather than at a single-company level — and it is why we treat NXTOrder™ as the underlying asset rather than any single vertical.

3. The Source of Truth: captured organizational context as the durable asset

Most of the market is currently focused on the model layer and the orchestration layer. Both will commoditize. Frontier models are already converging in capability; orchestration vendors will be unbundled within two product cycles. The variable that actually determines whether two companies deploying the same agentic stack get radically different outputs is the depth and structure of the captured organizational context the technology has access to.

We treat this as a strategic asset rather than a side project. Before we deploy a single autonomous agent into a portfolio company, we stand up what we call the Source of Truth: a hierarchical, permissioned data repository that mirrors the org chart, mirrors the functions, and is treated as the foundational layer of an internal LLM. HR identifies contributors. Department heads own cadence. Every function — sales, fulfillment, vendor management, compliance, finance, logistics — captures context as a normal operating rhythm rather than a documentation sprint. It is unglamorous, deeply human work involving change management, role definition, and a lot of patient reinforcement. That is precisely why it is the moat.

Three things follow from this discipline.

First, the AI gets fluency rather than just intelligence — industry-specific, company-specific judgment instead of generic output. The behavior of agents inside NXTOrder™ is shaped by years of resolved edge cases, not by prompt engineering.

Second, the product roadmap writes itself. Once context is mapped to the org chart, the high-leverage automation gaps become visible directly off the structure of the repository. We do not need to guess where to deploy agents next; the gaps in the Source of Truth identify themselves.

Third — and most important from a defensibility standpoint — competitors can reverse-engineer features and license a similar AI stack, but they cannot reconstruct years of cadenced, function-specific operational knowledge that has been structured, permissioned, and trained on. That asset is non-portable and non-purchasable. It exists only inside the firm that built it.

This is the part of the thesis the market is still catching up to. The conversation has been about which model wins. The conversation that matters is about which firm has captured the operating knowledge worth pointing a model at.

4. The three eras: from operator-driven systems to fully agentic execution

Inside NXT we describe the evolution arc in three eras, and we use it to position every portfolio company along a single trajectory.

Era 1 — Operator-driven systems. Core workflows executed manually by experienced operators. Critical business logic lived in people, not systems. Spreadsheets used for reconciliation, QA, and edge cases. Quoting, fulfillment, and exception handling required human judgment. High dependency on tribal knowledge inside each vertical. This is the state every underserved industry is still in, and it is the state our verticals were in when we entered them.

Era 2 — Platform plus AI-assisted workflows. NXTOrder™ centralizes CRM, workflows, pricing, and communications. Core operational logic is systemized through rules engines and workflow automation. SmartInbox™ structures and acts on inbound demand. Eight AI features sit in production across the portfolio today — email parsing, document parsing, AI search, message generation, CMS generation, and the matching engines inside Levo. This is the era NXT is currently operating in, and the proprietary data being generated inside Era 2 is what makes Era 3 possible.

Era 3 — Agentic systems. AI evolves from assisting workflows to executing them end-to-end within defined guardrails. Agents ingest demand, interpret context, make decisions, and execute actions. Routine workflows run continuously and automatically. Humans shift to exception handling and approval. The first full instantiation is the four-agent Agentic Mesh inside Levo — Perception, Sourcing, Negotiation, and Transaction agents collaborating to resolve charter requests with no human in the loop. Once proven in private aviation, the same architecture deploys across every other vertical on NXTOrder™.

The three eras are not a marketing structure. They are the actual operating cadence of how Vertical AI gets built when it is built from inside the business rather than from a research lab. Companies entering this space without sitting in Era 1 and Era 2 first will produce thin Era 3 deployments, because they will not have the captured context to point the agents at.

5. Reframing the labor question

The dominant Vertical AI discourse has collapsed into a TAM-expansion story: software spend versus labor spend, per-seat versus per-outcome, IT budget versus payroll budget. The numbers are real and we do not dispute them. The U.S. labor budget is north of four trillion dollars and a meaningful share of it is addressable by agentic systems. But a Vertical AI deployment positioned purely as headcount substitution is a thin product. It will be unbundled, repriced, and outcompeted within an iteration cycle.

The deployments that compound do something different. They take the repeatable execution layer of an employee's day — the workflows that consume time without producing leverage — and hand that time back. What employees do with the recovered time is where the durable value lives. They surface the pricing test that had been sitting on a shelf, the vendor relationship they had privately known was structured wrong, the fulfillment workflow the industry treats as normal but they have always considered broken. The operator knowledge those employees carry was previously trapped underneath a workload that did not let it surface. Vertical AI, done right, is the unlock.

The framing matters because it changes what gets measured, what gets priced, and how the buyer's organization receives the deployment. Productivity software loses to the next iteration. Creativity unlocks compound. Companies that frame Vertical AI as a labor reframe — internally, externally, in pricing, in change management — build the durable moat. Companies that frame it as labor replacement build a thin one. We have built our entire portfolio against the first framing and we have priced accordingly.

6. The compounding portfolio: why this only works at platform scale

A final point that does not show up in single-company Vertical AI theses but is central to ours. The captured-context moat compounds inside a single firm, but it compounds faster across a portfolio of firms running on the same platform. Every operator correction inside Axe & Kindle improves the agentic layer that runs inside Equipio. Every resolved edge case in Levo's fleet pricing engine feeds the rules architecture that Get Metals licenses to a metals service center in Ohio. Every new vertical we launch on NXTOrder™ starts with the cumulative learning of every other vertical that came before it.

Single-vertical AI companies cannot do this. They are building one firm's worth of context and competing against six firms' worth of it. The structural cost advantage of running shared engineering, finance, marketing, QA, customer operations, and logistics across the portfolio is real, but the more important compounding is in the intelligence layer. New verticals launch on proven infrastructure, with proven AI modules, against six verticals' worth of resolved exceptions. That is the part of the thesis that is the hardest to copy and the easiest to underestimate from the outside.

Competitors can reverse-engineer features and license a similar AI stack, but they cannot reconstruct years of cadenced, function-specific operational knowledge that has been structured, permissioned, and trained on across a portfolio of operating businesses.

7. What this implies

The conclusions we draw from this thesis are operational rather than rhetorical. We will not enter a vertical we are not willing to operate inside first. We will not deploy agents on top of a Source of Truth we have not yet built. We will not position our portfolio companies as labor substitutes, and we will not price them that way. We will keep the platform shared and the verticals specialized, because that is what creates the compounding the model depends on.

Most of the Vertical AI market will spend the next twenty-four months optimizing the wrong variable — chasing model performance, orchestration elegance, and per-seat pricing. The companies that come out of that period with durable businesses will be the ones that spent the same twenty-four months capturing the operating knowledge worth pointing a model at. NXT is built for the second path. The thesis is the same as the operating plan.