Vol. III · Essays on work after AI
Saturday, October 3, 2026
evolving.work
Trust · Org Design · Essay № 025

The New Bottleneck Is Not Intelligence. It Is Accountability.

AI has made intelligence cheap.

H Héctor Reinoso May 18, 2026 14 minute read Share · Save

Part of “After the AI Demo: The Real Constraints of the Agent Economy”

A mid-sized financial services firm ran the pilot. The AI compliance tool found more suspicious transactions than the human analysts did. The vendor documentation was clean. The economics worked. Six months in, the projection was a 40% reduction in headcount and a measurable lift in detection rates.

Then the General Counsel asked a single question, and the deployment stalled for seven months.

“If this system misses something we should have flagged and we face an enforcement action, who exactly is on the line?”

The vendor’s contract capped liability at the subscription fee. The Chief Compliance Officer’s license and the firm’s charter were not on offer. The model, no matter how accurate, could not be sanctioned, fined, or professionally disciplined.

The bottleneck was never the intelligence.

Intelligence Is No Longer the Scarce Resource

I work with these systems every day, and the capability gap I used to see is gone. Three years ago there was a real ceiling on professional-grade output — legal briefs, diagnostic differentials, financial models, strategic memos. That ceiling has moved. A first-draft contract from a modern model is competent. A radiology pre-read is, in many domains, at or above the average human baseline. The “wait until the AI gets good enough” argument has quietly run out of runway.

And yet adoption in regulated and high-stakes domains keeps stalling — at rates the capability story cannot explain.

The Deloitte 2024 enterprise survey is the clearest signal I have seen: 58% of Fortune 500 leaders named “unclear accountability for AI errors” as their top deployment barrier, against 41% for technical limitations. Adoption stalls outranked engineering stalls by seventeen points. A 2026 Accenture/Wharton report on co-intelligence puts the mismatch in one line: “While intelligence can be scaled, accountability does not share the same flexibility” — even as nearly three-quarters of knowledge workers now use AI, much of it unauthorized shadow AI.

Capability is racing ahead. Accountability is not. That mismatch is what is actually killing deployments.

Organizations Are Accountability Systems, Not Information Processors

When I discuss this with executives, I find it helps to ask what an organization provides beyond information and output.

We talk about organizations as if they exist to process information and produce outputs. They do. But they also exist for a less glamorous reason: to make sure that when things go wrong, someone is on the line. Professional licensing, malpractice insurance, fiduciary duty, signature requirements, sign-off chains — these are not bureaucratic friction. They are the operating system of trust.

A lawyer does more than generate a brief. A doctor does more than synthesize a differential. A tax adviser, a PM, an engineer — each does something the AI cannot. They take responsibility. Their license is the collateral. When the work fails, there is a known mechanism — malpractice, bar sanctions, license revocation, civil liability — to respond.

The American Bar Association’s 2024 guidance puts it plainly: “Lawyers who use AI tools remain fully responsible for the accuracy and integrity of their work product. The technology does not diminish the professional obligation to supervise and verify.” The University of Chicago Law Review, writing on AI as “risky agents without intentions,” draws the same line for the broader economy: “People should not be able to obtain a reduced duty of care by substituting an AI agent for a human agent.”

AI systems cannot be sued, sanctioned, disbarred, or stripped of a credential. They have no legal personhood, no fiduciary relationship, no skin in the game. When AI enters a workflow, the chain of accountability must still terminate in a person, an institution, a license, or a contract. The philosopher Andreas Matthias called this the “responsibility gap” more than a decade ago — and framed it as a structural feature of autonomous systems, not an edge case.

You can run an agent across the workflow. But the agent is not on the line.

The Three Percent Problem

An accuracy figure tells us how often a system succeeds. It leaves open who must answer for the failures.

Suppose an AI gets it right 97% of the time. In a marketing draft pipeline, that is more than enough. In a legal filing, a credit decision, a clinical recommendation, a security response — it is a catastrophe waiting for a quorum. Professionals are not responsible for the AI’s average performance; they are responsible for 100% of outcomes. Any error rate above zero requires active judgment about when to trust the output and when to override it. “97% accurate” sounds reassuring until you remember it is also the math of “a malpractice suit every thirty-three filings.”

In June 2023, attorneys Steven Schwartz and Peter LoDuca filed a federal personal-injury brief in Mata v. Avianca citing six court decisions. All six had been fabricated by ChatGPT, and the lawyers had not verified them. Judge P. Kevin Castel sanctioned both attorneys and their firm $5,000, found that reliance on AI output without verification constituted professional negligence, and triggered a wave of court orders requiring lawyers to verify and disclose AI-generated content. The defining sentence in the ruling was not about ChatGPT. It was about the attorneys: their signature on the brief was a legal declaration of personal responsibility for its accuracy.

The same logic runs through the rest of the regulated economy. Federal banking regulators — through SR 11-7 and successor guidance — require a traceable, accountable human decision-maker for every AI-assisted credit and risk output. Physicians remain liable when AI diagnostic tools contribute to errors, even when the AI outperforms the average clinician on labeled test data. As Mike O’Keeffe of Corlytics put it in a 2026 FinTech Global analysis, as firms rely on AI to interpret regulation, monitor transactions, and evaluate risk, “the human chain of responsibility becomes obscured.”

A second wrinkle explains why the “average accuracy” frame misleads. Human errors are idiosyncratic; AI errors are correlated. One physician misreading one chart is a tragedy for one patient. One AI model misreading the same way across ten thousand charts is a class-action lawsuit and a CMS investigation. Iyad Rahwan and Manuel Cebrian describe the “moral crumple zone”: humans absorb the liability when AI fails while the AI absorbs the credit when it succeeds. It is structurally unfair. It is also how courts and insurers are routing the risk.

Objections Worth Taking Seriously

Three objections deserve attention because they test different parts of this argument.

What if AI reliability eventually surpasses human baselines tenfold — and in narrow domains like radiology or document review, it is approaching that? Insisting on human sign-off may add harm by reintroducing human fallibility into a more reliable process. True on average; it misses the variance. AI errors cluster while human errors scatter, and a single correlated model failure can harm thousands at once — radiology has already seen one model’s bias propagate identical false negatives across cohorts at multiple institutions.

The institutional fix that would resolve this — legal personhood for AI, vendor liability regimes, certification boards — moves in decades, while capability moves in months. Organizations have to plan inside the current accountability regime, not the future one they hope is coming.

The second objection has teeth too. Professional bodies have a long, documented history of using accountability language to defend market position. When the bar says “only licensed professionals can be accountable,” part of what is defended is billable hours. That is real — and not the whole story. The accountability bottleneck is genuinely structural where harm is catastrophic and irreversible — medical malpractice, criminal justice, financial fraud — even when guild interest happens to coincide with it. The right response is to distinguish accountability that is structurally necessary (someone must own catastrophic downside) from accountability that is mere regulatory friction (insurance and vendor contracts can substitute). Domain by domain, not a wholesale judgment call.

The third draws on history. Electricity required new liability law. Automobiles required driver licensing and mandatory insurance. Aviation required certification boards and black-box mandates. Every transformative technology forced new accountability architectures into being, and AI will be no exception. Mostly correct on a twenty-year timeline — and misleading as a guide to a 2026 deployment decision. The adaptation depends on factors we cannot yet measure: how legible AI errors will be to courts, how many catastrophic failures the EU AI Act and FDA guidance need before crystallizing into operative law, how fast insurers build products for AI risk. Assuming the future regime exists when it does not is how organizations end up with real exposure.

None of these changes the near-term answer. They change how much humility we should hold about the long term.

What Accountability Architecture Looks Like

The organizations getting AI deployed in regulated industries are not waiting for the institutional regime to catch up. They design accountability architecture into the workflow itself. The Deloitte data is striking: companies with explicit AI accountability frameworks deploy AI 2.3 times faster than those without. Deployment stalls when nobody has worked out how to exercise that accountability in practice.

The version I keep refining in client work has four moving parts.

Step 1: Identify the accountability terminus. For every AI-assisted workflow, name the single human or institutional actor who is legally, professionally, or contractually on the line if the output is wrong. If the answer is “no one” or “the vendor,” the architecture is incomplete. The terminus is a named role with consequences attached — the CMO, the underwriter, the attending physician, the partner whose signature goes on the filing. A person with something to lose. The failure mode is accountability drift: the terminus is identified on paper, then operationally ignored as they delegate, approve in bulk, and assume someone downstream is checking. That is governance theater, and it is common.

Step 2: Classify the decision stakes. Not every AI-assisted decision needs the same architecture. A 2x2 on reversibility and harm magnitude is enough to start. RED (irreversible, high magnitude — clinical diagnosis, credit denial, legal filing, security incident response) requires a full accountability checkpoint. AMBER (medium — draft contracts, marketing content, customer-facing communications) requires human review before external use. GREEN (reversible, low magnitude — internal summaries, meeting notes, scheduling) can proceed with an audit trail. Revalidate with legal and compliance annually, and reclassify whenever workflow scope expands — because it will.

Step 3: Design the handoff protocol. Define exactly how AI outputs reach the terminus holder. Specify what they see, what they certify, and what happens if they certify without reviewing. The handoff is the operational mechanism by which accountability is actually exercised, not just assigned. Implement the certification as a real process step, not a click-through, and test whether it changes behavior. If average review time is under two seconds and override rate is zero across a thirty-day window, the protocol is theater.

Step 4: Build the audit and correction chain. Track AI errors after they occur and trace each back to the terminus. Distinguish AI errors from handoff failures from accountability failures, and tie findings to specific consequences — retraining, process change, recertification. Track the human override rate: if reviewers approve 99% of recommendations for a task class with no incidents, that class becomes a candidate for reduced oversight; if override rate is climbing, the model is drifting and the audit chain is doing its job.

Epic Systems’ clinical decision support shows what happens when this is done well — and badly. Hospitals that deployed Epic’s sepsis prediction algorithm without explicit accountability architecture saw alert fatigue exceed 90%; physicians simply ignored the AI. Hospitals that built explicit governance — physician committees reviewing alert logic, documented override rationales, tiered alerts distinguishing “advisory” from “requires action” — saw 40-60% higher physician compliance and shortened time-to-full-adoption by 18-36 months. Same technology. Different accountability architecture. Different outcomes.

This generalizes. Brookings has documented that automated decision-making at scale diffuses accountability when there is no architecture to concentrate it — no single human reviews each output, and responsibility quietly vanishes into the workflow. The fix is not to slow the AI down. It is to design where the human attention goes.

What Survives Is Accountability, Not Output Generation

The legal definition of an agent has always been someone who acts and bears the weight of acting. Acting carries consequences. Generating outputs does not.

The professionals who survive this decade will not be the ones who resist AI, nor the ones who race it on output volume — that race is over and they lost it. They will be the ones who understand that their real job has always been accountability, and that AI has made the accountability job more valuable, not less.

This applies upward as well as downward. The founder fantasy of running a company with fifty agents collides with the same wall: agents execute, the founder still owns every error. The org chart shrinks. The accountability surface area does not.

For the AI businesses that scale in regulated industries, the play is not just to ship a better model. It is to design accountability-compatible workflows — to give the licensed professional, the credentialed institution, the named officer a defensible liability position they can stand behind in court, in front of a regulator, or in front of a board. The vendors building this are quietly winning the enterprise market while the demo-day vendors are still showing slides.

Even autonomous vehicles, which have the cleanest case for capability surpassing humans, have not resolved the accountability question. Waymo started with human backup drivers and now uses remote operators. Tesla makes it explicit that drivers remain responsible when FSD is engaged. The cars outperform humans on most metrics. The lawyers and the insurers still know whose name to put on the policy.

The scarce resource in the AI economy may not be intelligence. It may be accountable judgment — the kind that comes with a license number, a contract signature, and a name on the door.

If you are leading AI adoption, the question to ask in the next deployment meeting is not “is the AI good enough?” It almost always is. Ask instead: when this AI is wrong — and it will be — who specifically is on the line, what exactly do they have to lose, and have we built the handoff that lets them actually exercise judgment rather than rubber-stamp the model?

A vague answer leaves the deployment exposed to a costly stall. A clear answer gives you an accountability architecture to work with. In this account, providing the model is the easier part.

References

AI strategyAI transformation
About the author
Héctor Reinoso writes about how organizations should be designed when intelligence becomes abundant. He works on enterprise AI adoption inside a large industrial organization, after 25 years as a strategy consultant and transformation executive across four continents. Harvard MBA, industrial engineer, and still writes the code he writes about. More →
Continue reading
Org Design · Oct 2, 2026

When Coding Gets Cheaper, the IT Handoff Gets Harder to Justify

If your supplier uses coding agents while you still specify and verify the work, it is time to reconsider what you are outsourcing.
Org Design · Sep 26, 2026

The Myth of the 50-Agent Founder

The fantasy of a solo founder running a company with 50 agents misunderstands what agents do to cognitive load.
Org Design · Sep 24, 2026

When Time Collapses, What Breaks First?

Time was an invisible governance system in organizations, not just a constraint.
Org Design · Sep 21, 2026

AI Does Not Eliminate Bottlenecks. It Moves Them.

AI does not eliminate bottlenecks.
Org Design · Sep 18, 2026

The More Abstract the System, the Harder It Is to Fix

Abstraction is how organizations scale: metrics replace direct observation, processes replace individual judgment, dashboards replace frontline presence.
Trust · Sep 15, 2026

The Correctness Economy

AI made volume free; correctness became the binding constraint.