The Schema Remembers a Different World

A practitioner opens the tax research platform she has used since before her first audit defense. Some quarter in the past two years, a new pane appeared in it: an assistant that answers questions in prose, with citations into the publisher’s library. She asks it the question actually on her desk, which is half her client and half the law. What does the recent circular mean for the restructuring her client’s board approved in March, given the positions taken in last year’s filings?

The pane answers the half it can see. The law half comes back competent and cited. The client half has nowhere to live, because the platform was built to hold documents and queries, not a client’s accumulated situation. So she does what experienced users of any tool do. She reshapes the question into the shape the tool can take: what do the authorities say about restructurings of this type. That question it answers well.

Notice who adapted to whom.

The pattern is concrete and datable, and it is not unique to any vendor. Thomson Reuters’ Checkpoint Edge launched in 2019 as the successor to a research line decades older; its AI research capability arrived in 2024, built as retrieval over the publisher’s editorial library. Wolters Kluwer added generative answering to CCH AnswerConnect across 2024 and 2025, grounded, by its own description, entirely in its proprietary research library. Thomson Reuters likes to note that it has been working with AI and language models for over thirty years. None of this is criticism. The additions are competent, and the libraries underneath them are genuinely excellent. The question is a different one: did the AI get added to the workflow, or was the workflow built around the AI?

That sounds like a marketing question. I want to argue it is a structural one, and that the marketing has made the structure harder to see.

---

A Label Everyone Now Claims

Start with definitions that do not depend on what vendors call themselves.

In a native system, the model is a replaceable component, and the system around it is the product: the way knowledge is structured, the way retrieval works, the discipline that governs citations, the conditions under which the system declines to answer. In a bolted-on system, the model is the addition, and the architecture it was added to was fixed years earlier, by careful people designing for a world in which nothing could reason.

The labels themselves have detached from this structure entirely. Salesforce announced Einstein GPT in March 2023 as the first generative AI CRM; the metadata platform underneath dates to 2008, and one practitioner analysis described the result as having “bolted Einstein onto a 25-year-old database architecture.” The same company now markets its newest layer as natively embedded in the platform. Thomson Reuters describes its newest tax modules as AI-native applications. When the retrofit and the rebuild both call themselves native, the word carries no information. It has become a compliment vendors pay themselves.

What still carries information is a pair of questions. What in this system was fixed before the model existed? And what would survive the model being replaced tomorrow?

The second question is not hypothetical. Harvey, founded in 2022 with no pre-AI product to retrofit, runs each query through what the Vals Legal AI Report described in early 2025 as between 30 and 1,500 model calls, and in May 2025 it added a second and a third model provider underneath that machinery. Whatever else one thinks of the product, the model there is plumbing. Though the honest footnote matters: until May 2025 the same company depended on a single provider, which is a structural risk no founding date protects against. Being born native buys the option of treating the model as replaceable. It does not buy the discipline. The discipline has to be built either way.

---

The Assumptions Underneath

Here is the conceptual core, and it is about schemas, not models.

Every platform encodes the assumptions of the world it was founded in. How a document is stored. How an entity is modeled. How client context is held, if it is held at all. What an alert is, and who is expected to act on one. These choices were made by capable people solving the problems of their decade: search, storage, records, transactions, billing. Reasoning was not on the list, because nothing could reason. When a model is added on top of such a platform, it inherits every one of those choices as fixed input.

The inheritance is easiest to see where it failed in public.

In healthcare, the dominant record systems were built to store encounters and support billing. When a widely deployed sepsis prediction model was embedded into that world, external validation published in JAMA Internal Medicine in 2021 found its discrimination far below what had been reported, and at the operating threshold it missed roughly two thirds of septic patients while generating alerts on nearly one in five hospitalized ones. The failure was not only the model’s. It was the founding assumption that an alert is a task in a clinician’s worklist. Pour a weak signal into a workflow whose unit is the task and the result is alert fatigue, which is the architecture’s verdict, not the algorithm’s.

In financial crime monitoring, the assumption ran deeper. When FinCEN penalized U.S. Bank in 2018, the core finding was that the bank had capped the number of alerts its monitoring system would generate so that the volume matched its investigation headcount, and that when internal testing showed the cap was hiding suspicious activity, the bank ended the testing rather than the cap. The system’s deepest assumption, that detection should be sized to investigative capacity, was written into the platform itself. TD Bank’s record settlement in 2024 cited a failure to substantively update its monitoring system across eight years of growth. These cases predate the current generation of models, which is precisely the point: the outcomes were set by the architecture around the analytic core, not by the core.

In legal research, closest to my own work, the incumbents added retrieval layers over decades of curated law and made strong claims about the result; one stated its tool delivers “100% hallucination-free linked legal citations.” The Stanford study published in the Journal of Empirical Legal Studies, testing the tools as they stood in May 2024, found the major legal research products still hallucinating between 17 and 33 percent of the time. The vendors disputed the methodology, the tools have evolved since, and no equivalent independent measurement of the newer versions exists, so I am reporting a dated finding, not a verdict. But the structural reading survives the dispute. A database built to answer “find me documents about X” was handed to a model that answers “tell me the answer to X,” and the seam between those two questions, the place where verification has to live, belongs to neither inherited layer.

What lives in that seam was demonstrated in a courtroom this spring. In United States v. Farris, the Sixth Circuit sanctioned an attorney whose briefs, drafted with Thomson Reuters’ CoCounsel, cited real cases containing fabricated quotations, and the court held that quoting words that do not exist inside an authority that does is no less sanctionable than inventing the authority outright. A public database maintained by a researcher at HEC Paris counted more than 1,200 such decisions worldwide by April 2026. The instructive part is what grounding did and did not do. Anchoring generation to a verified database closed one failure mode, the invented case. It left another open, the invented words inside a real one. What stands between those two failure modes is not the model and not the database. It is what the system does at the moment a generated reference cannot be verified: substitute, smooth over, or stop and mark. That is an architectural decision. And it is exactly the kind of decision a schema designed for search never had to make.

---

Where Bolted-On Wins

If the previous section reads as a case against incumbents, the research does not support that reading, and I will not pretend it does.

The bolted-on side wins, first, on the integration surface. The most clear-eyed commentary on Microsoft’s Copilot holds that its deepest advantage “is not the model; it is the integration surface”: the assistant sits where work is created and inherits the organization’s existing permission model, seeing only what the individual user is already authorized to see. The same inheritance I described above as a constraint arrives here as trust. Decades of accumulated assumptions about who may read what are precisely what an enterprise buyer wants an AI to be bound by, and a native entrant has to rebuild that binding from nothing.

It wins, second, on content. The most defensible accuracy claims anywhere in the research are grounding claims, and practitioner commentary on legal research favors the incumbent-grounded tool for exactly that reason: it searches a verified, expert-curated corpus rather than generating citations from model memory. Libraries of that depth took the better part of a century to accumulate. A native architecture sitting on thin content can be honest only by declining often, and a system that declines often is honest and useless at the same time.

It wins, third, on procurement and institutional trust, which the research suggests decide more purchases than architecture does. A bolted-on capability is incremental spend on a contract that already exists; a native platform is a new budget line and a new risk review. In regulated finance, the platform a named analyst firm ranked a category leader in transaction monitoring in 2024 was founded in 1999. In medical imaging, the FDA had authorized over 1,300 AI-enabled devices by late 2025 while only two reimbursement codes existed for any of them, and the AI-native imaging companies that succeeded did so by integrating deeply with the legacy record systems, not by replacing them. Where liability and reimbursement gate adoption, the incumbent’s seat at the table is worth more than the entrant’s architecture.

And the scoreboard refuses to cooperate with a clean story. In the one independent benchmark the research surfaced, a top score was taken not only by the native flagship but by a tool that began life as a startup and was grafted onto an incumbent’s library, while human lawyers still outperformed every tool on two of the tasks. The native side has its own embarrassments: a legal AI unicorn was reported by Business Insider in late 2024 to rely heavily on human workers, with supervisors steering staff away from its own AI; and the most heavily marketed AI-first ambition in healthcare produced treatment recommendations that internal documents, reported by STAT in 2018, called “unsafe and incorrect,” before the division was wound down, though none of those recommendations reached a patient. Native branding guarantees nothing. Discipline does, and discipline has no founding date.

The classifier even blurs at its edges. Thomson Reuters did not retrofit its way to CoCounsel; it bought Casetext, a company founded in 2013 and native by any definition, and grafted it onto Westlaw. Visa bought Featurespace. The founding date of the parent tells you nothing about the architecture actually running.

So here is the honest synthesis. Bolted-on systems win wherever the question being asked is one the old world already knew how to ask. Where is the document. Who may read it. Which authority discusses this topic. Does this transaction match a known pattern. Those are real questions, and most professional days are full of them.

What a bolted-on system cannot do, without redesign, is answer a question its schema cannot represent. A question that lives half in a client’s accumulated situation and half in the law needs a system that holds both as structured knowledge, knows which is which, can show where its confidence comes from, and can stop when it should. None of that is a feature to be added to a platform whose schema was never designed to hold a client’s situation at all. It is a set of founding assumptions. And founding assumptions are the one thing a platform cannot retrofit.

---

What Native Buys, Concretely

ktAi is one attempt to make those assumptions deliberately, and rather than claim the label, I will describe what making them bought, because the properties are the argument.

The model is the most replaceable component in the system, by design. Which model handles which task is a configuration choice, and the architecture is built so that the answer could change without the system changing what it is. What the system is, is the discipline around the model. Knowledge lives in two physically separate graphs: the law, shared across all users, and the client’s situation, private per user, with the user’s identity carried on every node. The two are never joined in storage; they meet only at retrieval, in application code that can be read and reviewed. The client’s graph also remembers time twice over, when a fact was true in the world and when we learned of it, so that the audit trail can answer not just what the graph says today but what we knew on the date of the opinion.

Citations are not free text. Before drafting, the authorities actually retrieved are registered; after drafting, every citation the model produced is resolved against that registry, and a citation that cannot be resolved is preserved verbatim and marked unresolved, never replaced with a plausible neighbor. The quality gate fails closed: a section that cannot support its claims ships with the gap marked, not with fluent prose covering it. There are questions the system declines to answer, and the decline is structured rather than generic; it says why, and what evidence would change the answer. And at the end of the line there is a human gate. The reviewer who will sign cannot approve their own work, and cannot approve while a blocking comment stands unresolved, and those rules are enforced in the transition logic, not in a policy document.

The gaps belong in the same paragraph as the properties. The system does not yet bind each citation to the specific sentence it supports; faithfulness is checked at the level of atomic claims against the assembled sources, and the reviewer carries the residue. A revised section runs a narrower path of checks than a fully generated one, a divergence we accept because the reviewer is in the loop on both. And replaceability has an honest edge: swapping the model is configuration, while swapping the provider beneath it is a deeper commitment, and pretending those are the same line in a settings file would be exactly the kind of fluency this series argues against.

None of these properties is a virtue of being new. They are founding assumptions, available to us because nothing older was in the way.

---

The Question Each System Can Honestly Answer

The wrong question is which kind of system is better. The research will not crown a winner, and I have tried to show both columns of the ledger.

The right question is the one this essay has been circling: which questions can each system honestly answer, and at what cost. A bolted-on system answers, honestly and deeply, the questions its founding world anticipated, at the cost of quietly reshaping every new kind of question into an old shape. A native system can honestly answer, or honestly decline, the questions it was designed around, at the cost of starting without the content, the integrations, and the institutional trust that took the incumbents decades to earn, and that still decide most purchases.

For a professional, the test requires no architecture diagram. Ask the system a question that is genuinely yours, half facts and half law, and watch who adapts to whom. If you find yourself translating your question into the system’s shape, you have learned what its schema remembers.

I made the founding assumptions in the open, and this series is partly a record of them. Whether they were the right ones is not something an essay can settle, and not something a label can claim. It gets settled where everything in this profession gets settled, one signed report at a time.