On Harvey, ritualized oversight, and the quiet erosion of institutional accountability

I was talking recently with a lawyer I have known for years. His firm had been integrating AI into contract review for most of the past year. Some firms in his orbit had adopted Harvey. Others were running general-purpose models embedded into their own workflows. The mechanics differed, but the underlying posture was the same. Everyone was attempting to "embed AI into daily practice." Review. Research. Drafting. Memos. Even the explanatory cover notes that go out to clients. Each step that used to consume an associate's afternoon now passed first through a model. A partner reviewed what came back. Work moved.

None of this is controversial when described in the abstract.

What stayed with me was the offhand sentence he closed with.

"Honestly, lately it feels like I am mostly just pressing the confirm button."

He is not, I suspect, the only one.

The quiet collapse of "we review everything"

Conversations about AI review tend to land in the same comfortable place. We will have a human check everything. That sentence is supposed to reassure. It probably reassures the speaker more than anyone else.

But if you sit with the actual artifacts these systems produce, the reassurance gets thinner. Modern legal AI is exhaustive. It will surface every flagged clause, every deviation from the firm's standard playbook, every plausibly relevant precedent, every counterargument worth pre-empting. The output is color-coded, structured, and annotated. It arrives looking like a finished work product rather than a draft.

Now apply the old norm — the partner reviews everything — to that volume of output.

The total reviewable surface from a single AI-augmented matter often exceeds what a junior associate would have produced on the same matter unaided. If you genuinely read every annotation at the same depth your training tells you to apply, AI-assisted review takes longer than the workflow it replaced. So in practice, nobody does that. Reviewers triage. They follow the flags that look serious. They skim the summaries. If nothing trips a sense of wrongness, the work clears.

This is not laziness. It is physics.

But the institutional vocabulary has not caught up. Time entries still say reviewed. Engagement letters still promise partner-level oversight. Internal QC checklists still assume the work was inspected end to end. A thin membrane forms between the formal narrative and the actual one. Early on, the membrane is almost invisible. Matter by matter, it thickens.

Confirm as ceremony

Human-in-the-Loop is the phrase that appears in every governance document on the subject. Japan's regulators have moved decisively in this direction. The government has begun requiring developers of autonomous AI agents to build in mechanisms that keep human judgment in the loop, citing concerns about malfunction and privacy violations. The most recent revision of the AI Business Guidelines (v1.2), jointly issued by Japan's Ministry of Internal Affairs and Communications and METI, explicitly names Human-in-the-Loop, continuous monitoring, and log management as necessary components of responsible AI deployment.

The direction of policy is right. Arguably it arrived later than it should have.

What policy has not yet caught is the distance between the phrase human in the loop and what actually happens when a human is, technically, in the loop.

The Human-in-the-Loop picture imagines something like this. The model produces output. The human examines it. The human accepts, modifies, or rejects. Judgment lives on the human side. The model does the legwork.

What actually happens, once the model's output quality reaches a certain threshold, is more elusive. The reviewer begins reading. Nothing triggers an alarm. The reviewer continues. Something feels slightly off in a clause, but the reviewer is no longer sure whether the unease comes from a genuine defect in the model's reasoning or from their own misreading. The cost of investigating is non-trivial. The cost of approving is one click. After a few weeks of this rhythm, the reviewer's default response is probably fine.

Is that judgment? Is it inspection? Is it the procedural performance of approval?

From the outside, all three look identical. The confirm button is pressed. The audit log records a human in the loop. The governance requirement is, formally, satisfied.

But when the person doing the pressing describes themselves as "mostly just pressing the confirm button," something has shifted. Whatever that act is, it is no longer the substantive judgment the loop was designed to preserve.

Where junior lawyers used to learn the thing nobody can teach

There is a second, slower shift happening underneath the first.

For most of the modern history of the legal profession, lawyer quality has been measured — implicitly, never explicitly — by something close to a sense of wrongness. The trained ability to read a contract clause and feel that it does not quite sit right. To follow a chain of reasoning in a judgment and notice the silent leap. To listen to a client describe a transaction and detect the unstated fact whose absence shapes the whole story.

That sense is not taught from a textbook. It is accumulated. A junior reads thousands of documents, gets corrected, misses things they should have caught, watches a senior partner catch what they missed, and slowly, over years, builds an inner radar that is reliable enough to bet a client's interests on. The work is unglamorous and the payoff is delayed. It is also the foundation of the profession.

Insert AI into the pipeline and the shape of this apprenticeship changes.

The first-pass review now comes from a model. The flags are pre-set. The issues are pre-organized. The junior's job is to inspect a structured artifact rather than to construct one. Inside that frame, developing an independent sense of what should have been flagged becomes structurally harder. To override what the model has presented, the junior needs evidence beyond what the model has presented — and the discipline to even ask is precisely what they have not yet built.

Senior partners are insulated, for now. They carry an embodied catalogue of friction from the pre-AI years. When the model output looks clean, the senior partner can still feel the absent flag — the thing the model is silently confident about that warrants a second look. That instinct is the residue of fifteen years of catching things the hard way.

The harder question is what happens to the next generation. The associates joining now will not accumulate the same catalogue. The friction that produced it has been smoothed out of their daily work. By the time they are senior, the question will not be whether they are capable lawyers in some general sense. The question will be whether the sense of wrongness, as the profession historically understood it, can be reconstituted in people who never had to develop it.

The model has not replaced the apprenticeship. It has stepped in front of it.

Where responsibility goes when nobody is quite holding it

Follow this far enough and a familiar question surfaces.

When something goes wrong, who is accountable?

The formal answer is fixed. The signing lawyer is accountable. The model is a tool, not a principal. No serious bar regulator anywhere is preparing to change this. Good.

The phenomenological answer is messier.

There is a difference, inside the person doing the work, between I drafted this and missed it and the model drafted this, I reviewed it, and we both missed it. Both produce the same legal liability. They do not produce the same felt sense of ownership. In the second case, some portion of the responsibility silently relocates — not to the vendor, not to anyone identifiable, but somewhere outside the reviewer's chest.

Responsibility, as an internal experience, has always been bonded to authorship. I decided, so I answer for it. When the act of deciding becomes the act of confirming, that bond loosens. The looser version is not absent. It is just lighter than it used to be.

Where does the displaced weight go?

Not to the AI vendor. Their terms of service are unambiguous on this point — they make no warranty regarding the substantive accuracy of model output. Not to the regulators. Every governance framework, from the EU AI Act to Japan's own guidelines, locates final operational responsibility with the deploying organization.

So the weight does not actually go anywhere. It hangs.

Legally accountable parties remain identifiable. What dissipates is something subtler: the institutional sense that someone is carrying this matter. On any given file, you can still name who signed. What you increasingly cannot do is point to anyone for whom that signature was the visible surface of a fully-owned act of judgment.

The discomfort that does not yet have a name

None of this is an argument against AI in legal practice. That argument is no longer available. The productivity gains are too clear, the client expectations are too far along, and the firms that try to stand outside the trend will not have clients to stand outside of.

The interesting problem lives in what remains after the adoption decision has been made.

The reviewer who presses confirm and is unsure whether they decided anything. The associate who is doing the work and learning a different thing than the previous generation learned. The matter where the partner of record carries less than they used to, and the difference goes nowhere identifiable. The firm whose internal narrative still says we review everything while the actual work is being triaged in ways nobody has explicitly authorized.

None of these are well-described by the vocabularies the profession has on hand.

Governance language describes rules and their enforcement. The problem here is upstream of any rule. Digital transformation language describes operating model change. The problem here is what happens to authority once the operating model has changed. Automation language describes the substitution of machine work for human work. The problem here is what happens to the human work that remains. AI ethics language describes principles a system should satisfy. The problem here is structural, not principled.

Each of these vocabularies is useful. None of them name what is actually shifting.

The people closest to the work can feel it. They have not yet been given the words.

What follows is an attempt to give the discomfort a structure — to name the confirm-button reflex as a specific failure mode, to describe the apprenticeship erosion as an architectural problem rather than a cultural one, to locate the dispersing accountability inside a framework that can be deliberately designed rather than passively inherited.

Governance is not sufficient. Digital transformation is not sufficient. Automation is not sufficient. AI ethics is not sufficient.

The work the next decade requires sits in the gap between all four. Organizations that do not address that gap will, several years from now, find themselves staffed by people who cannot fully account for the judgments their own organization has issued in its name.

The problem is, at this stage, still early. That is the only good news in it.

The paid section begins here

Decision Design: making judgment itself the object of architecture

The conceptual move this article wants to make is straightforward to state and slow to absorb.

It is this. The category we have been missing is not AI governance, not digital transformation maturity, not human-AI collaboration design. It is the category of judgment itself, treated as something an institution architects on purpose.

Call it Decision Design.

Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate.

At its center sits a concept I have been calling the Decision Boundary. The Decision Boundary is the line — usually invisible, almost never explicit — between what the institution has delegated and what it has retained. Who decides. What gets delegated to a model, to a workflow, to a junior, to a partner. What is reclaimed at the threshold. Where the institution stops outsourcing and starts answering.

Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority.

The provocation here is not that this concept is new. The provocation is that almost no organization, including most that consider themselves mature on AI, has anyone whose explicit responsibility is to design it. Vendor selection sits with procurement. Workflow integration sits with operations. Compliance sits with legal. Talent development sits with HR. The judgment architecture that connects all four sits nowhere. It emerges as a byproduct of decisions made for other reasons.

When the AI layer arrives, the byproduct stops being benign.

What Decision Design actually designs

To make this concrete, consider the elements that fall inside the scope of judgment architecture once you take it seriously as a design discipline.

The first is authority allocation. Which categories of judgment can legitimately be delegated to the model, which to the workflow, which to a junior, which to a reviewing partner, and which must reach a named principal. This allocation cannot be uniform across matter types. Risk tier, client profile, deal size, regulatory exposure, and reversibility of the underlying decision all bend the line in different directions. A flat policy — AI handles drafts, humans handle finals — fails on contact with practice. The allocation has to be matrixed, written down, and revisited.

The second is escalation logic. Under what observable conditions does an AI-handled item surface to a human, and under what conditions does it not. This is the single largest determinant of whether Human-in-the-Loop is real or ritual. Set the escalation threshold too low and the human queue fills with noise; reviewers respond by learning to clear it quickly, which is exactly the failure mode Human-in-the-Loop was supposed to prevent. Set it too high and material exposures never reach a human at all. The escalation rules need to be specific enough to operate — "low model confidence on a clause whose monetary exposure exceeds X, in a matter classified above tier Y" — not aspirational, like important matters get a second look.

The third is override governance. What does it cost, structurally, for a reviewer to disagree with the model? Cheap overrides erode the productivity case for the system. Expensive overrides train reviewers to suppress their own unease and pass things through. Somewhere between those poles lives a friction calibration that institutions need to set deliberately — and document — rather than letting it set itself through fatigue.

The fourth is delegation explicitness. Most organizations have not actually said, out loud, what they have delegated to the model. They have allowed delegation to drift in through the workflow. Where exactly the delegation begins, what its bounds are, who authorized it, and who carries the residual when the model errs — these are governance questions, and most institutions cannot currently answer them in writing.

The fifth is decision logging infrastructure. Not output logging. Decision logging. The structured preservation of what was decided, by whom, on what evidence, against what alternative, and at what point in the chain the institution committed. Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes. This is the substrate that makes the rest of the architecture auditable rather than aspirational.

The sixth, and the one that ties the others together, is the visibility of the Decision Boundary itself. Every party operating inside the system — the model, the reviewing associate, the supervising partner, the client-facing relationship lead, the general counsel of the client — needs to be working from a shared map of where authority sits. If that map exists only in the head of the managing partner, or worse, exists nowhere, it is not an architecture. It is folklore.

These six are not six independent topics. They are six faces of the same question: where, in this institution, does judgment actually live, and how do we keep it from quietly leaving?

What Decision Design is not

A concept defined only by what it includes drifts. A concept defined by what it is not stays sharp. This one needs the discipline of negative space.

Decision Design is not AI adoption strategy. Adoption strategy ends when the tool is in the workflow. Decision Design begins there.

Decision Design is not governance in the conventional compliance sense. Governance asks whether actions conform to rules. Decision Design asks who has the authority to commit the institution to the action in the first place, in a domain where the relevant rules were written before the technology that now executes within them existed.

Decision Design is not automation. Automation removes decision space. Decision Design preserves decision space and architects who occupies it.

Decision Design is not digital transformation. Digital transformation is concerned with operating model and value creation. Decision Design is concerned with what happens to authority and accountability inside the transformed model, which most DX frameworks treat as out of scope.

Decision Design is not AI ethics. AI ethics articulates the principles a system should embody. Decision Design specifies the institutional architecture through which a system's outputs become commitments the organization can defend.

Decision Design is not workflow optimization. Optimization treats the workflow as the object. Decision Design treats the workflow as a substrate on which the judgment architecture is mounted.

Each of the five fields above is doing real work. None of them, on their own or in combination, designs the thing that is currently breaking.

The problems Decision Design exists to address

Worth naming them in plain language, because the failure modes they describe are already in motion inside organizations that would describe themselves as well-governed.

Authority allocation ambiguity. Nobody at the firm can produce, in writing, the answer to which categories of judgment we have delegated to the model. The answer exists in practice. It does not exist in policy.

Escalation gaps. The threshold for surfacing AI-handled matters to a human is implicit, inherited from the previous workflow, and unaligned with the actual risk distribution of model output.

Ritualized Human-in-the-Loop. The human is in the loop. The human is not, in any meaningful sense, judging. The audit trail is intact. The judgment is not.

Accountability diffusion. The signature is on the file. The internal sense of ownership for what the signature commits is thinner than it used to be, and getting thinner. No one party absorbs the difference. It distributes.

Hollowed human oversight. Oversight roles persist on the org chart, but the work those roles actually do has migrated from substantive review to procedural confirmation.

Disappearance of institutional judgment. Several years downstream, the organization no longer contains people who can reconstruct, in their own terms, why a given class of matter was decided the way it was. The decisions exist. The judgment-bearing humans who would once have stood behind them do not.

These are not six separate dysfunctions. They are six visible surfaces of the same underlying void: the absence of a deliberately designed architecture for institutional judgment in an environment where judgment is now distributed across human and machine actors.

What implementation actually looks like

A concept that does not survive contact with implementation is a slogan. So, in the spirit of being concrete:

For a law firm beginning to take Decision Design seriously, the first move is a matter-type matrix for AI review boundaries. NDAs, standard commercial agreements, and high-volume routine work move to AI first-pass review with human attention scoped to deviations and exceptions. M&A definitive agreements, regulatory submissions, novel transaction structures, and litigation strategy stay human-led with AI confined to research support and pattern surfacing. The matrix is published internally, owned by a named partner, and revisited quarterly.

Built on top of that is a rule-based escalation regime. Specific triggers — model confidence below threshold, output divergence from prior matter precedent, client risk classification, exposure above defined monetary tier, regulatory novelty flag — automatically route matters to senior review. The rules are written, not folkloric. They are auditable. When they are wrong, the failure mode shows up in the log rather than in a malpractice claim.

Layered on that is override governance coupled with structured decision logging. When a reviewer changes the model's recommendation, the change is captured with reasoning in a structured format. Periodic aggregation of those overrides becomes a primary input to model selection, prompt refinement, and partner training. The override log is, in effect, the institution's evolving theory of where its automation systematically gets things wrong.

At the workflow level, the firm formalizes a three-tier review architecture: form review, substantive review, and committed judgment. The model conducts form review and part of substantive review. A human takes substantive review the rest of the way. A named principal performs the committed-judgment step — the act in which the firm's authority is actually staked. Each tier has defined inputs, outputs, and handoff conditions. Nothing is left to the assumption that of course the partner saw it.

To allocate human review capacity intelligently, the firm implements confidence-tiered routing. High-confidence model output receives summary-level review; medium-confidence output receives partial review; low-confidence output receives full review. Reviewer attention is treated as the scarce institutional resource it actually is, and routed accordingly.

The final element, and the one most firms will resist, is a deliberately designed apprenticeship architecture. Junior lawyers are required to work a defined volume of matters without AI assistance during their first years, even when this is operationally inefficient, specifically to build the catalogue of friction that produces the senior partner's eventual sense of wrongness. Separately, a structured training program teaches juniors to identify what the model failed to flag — treating finding the absent flag as a discrete skill that has to be built rather than absorbed. The economic case for this looks bad in any given quarter. The economic case for not doing it looks worse over a decade.

None of these elements is novel in isolation. Pieces of each exist in firms that have been thoughtful about their AI integration. What Decision Design proposes is that these pieces stop being treated as discrete operational improvements and start being treated as components of a single institutional architecture — the architecture through which the firm preserves its own capacity to judge.

What is actually at stake

The decision to adopt AI in professional services has been made. Not by any individual firm, but by the market the firms operate inside. Clients expect it, competitors deploy it, and the firms that opt out will not have the option for long.

The decision that has not been made — and that remains genuinely available, for now, to institutions willing to take it — is the one underneath. Who, inside this institution, will be designed to carry judgment, on what terms, with what authority, traceable through what record.

Institutions that decline to draw their Decision Boundaries deliberately will discover, ten years from now, that they no longer contain people who can carry judgment. Institutions that decline to architect accountability continuity will discover that they can no longer explain, internally or externally, why they decided what they decided. Institutions that decline to preserve decision logs will discover that their own history is, to them, opaque.

These are not AI problems. They are problems the AI layer has made visible. The absence of judgment architecture has always been the quiet condition of most large organizations. Until recently, the human reviewers in the loop carried enough residual judgment that the absence did not show. The reviewers are still there. The judgment they once carried has begun to thin.

The work of drawing Decision Boundaries is unglamorous. It produces no immediately measurable outcome. It is the kind of work that justifies itself only on the timescale of institutions, not quarters.

It is also the only work that keeps the institution's judgment inside the institution.

The choice in front of professional service firms — and not only them — is whether they will become organizations that press the confirm button, or organizations that continue to carry what the confirm button is supposed to represent.

That choice is arriving faster than the technology curve suggests. It is already in the building.