A shift is underway in how serious companies talk about artificial intelligence, and it has very little to do with what the technology can do. The capability question has largely been settled. AI can read a document, approve a loan, screen a candidate, price a policy, and flag a fraudulent transaction faster than any team of people ever could.

The harder question, the one that keeps surfacing in boardrooms and post-mortems, is what happens when the system gets one of those decisions badly wrong, and whose name sits next to the consequence. For a long stretch, that question was left unanswered. Models were treated as experiments, errors were treated as teething trouble, and responsibility was something an organization figured out after the fact, if at all.

That era is closing. As AI moves into credit, hiring, supply chains, and clinical care, the cost of an unexplained mistake has stopped being a line in a dashboard and started becoming a financial, regulatory, and human reckoning. NervNow spoke with leaders who are living inside this problem every day. None of them sounds anxious about the technology. What they share is a firm conviction that a machine cannot be the place where accountability finally stops.

A Machine Can Execute a Decision, But It Cannot Be Held Responsible for One

Venkat Ramana, who leads NthEye and ValuePitch and has spent more than two decades building verification, lending, and surveillance products, puts the matter in language that leaves no room for evasion. Accountability, in his view, does not evaporate because a machine carried out the task. He describes AI as an execution engine, a way of saying it can carry out a decision while the responsibility for that decision stays with a person, and the whole argument follows from there.

When an algorithmic pricing model produces an outcome that wipes out a margin, the person who answers for it is the executive who owns revenue. When a screening tool introduces bias into hiring, the head of human resources answers for it. The technical team is responsible for the fidelity of the system, while the fiduciary and reputational weight rolls up to whoever deployed the technology in pursuit of their own targets.

Aksheshkumar Shah, the founder and chief executive of Cogniify.ai, arrives at the same destination from a slightly different road. At his company, the teams and leadership responsible for deploying a system are the ones held responsible for what it produces, and the reasoning is almost philosophical in its simplicity. A company uses AI as a tool, and the tool does not get to make its own choices.

When an outcome goes wrong, the inquiry turns to whether the failure was rooted in design, implementation, oversight, or the quality of the data feeding the model, and responsibility is assigned to the team or manager closest to that cause. The final accountability, though, always rests with the business using the AI, regardless of how mature the underlying technology happens to be.

“

AI is an execution engine, not a scapegoat. The accountability sits squarely with the executive who owns the business function.

Venkat Ramana — CEO, NthEye & ValuePitch

Anupam Shrey, co-founder of the climate fintech Plutas.ai and founder of DigiSafe Insurance Broking, raises the stakes considerably, because in his world a wrong decision arrives as money already gone rather than as a recommendation someone can revise. In the kind of parametric model his work involves, a payout can fire automatically the moment a rainfall threshold is crossed, reaching a recipient within hours, with no claims process and no adjuster standing in between. There is no dashboard to correct after the fact, so a mistake becomes financial and human the moment it happens.

For that reason, Shrey says, his organizations built accountability into the architecture before a single model was written, because that is the only sequence that works. Every system has a named owner, one individual who carries documented authority over how it is deployed, where its limits sit, and when it is shut down.

Gurudatt Bhobe, who leads engineering and AI at the identity verification company IDfy and has spent a decade in the field, frames the same conviction through the texture of the systems themselves. AI, he points out, is probabilistic by nature, which is a precise way of saying it deals in likelihoods, and a system that deals in likelihoods has to be wrapped in guardrails before anyone leans on it.

At IDfy, high-stakes decisions are designed from the outset to keep a human in the loop, and AI outputs are validated and post-processed to push accuracy as high as it will go before a decision is allowed to stand. The point, across all four companies, lands in the same place. The machine can be fast, capable, and even autonomous, and it still never becomes the entity that carries the consequence.

You Can Outsource the Algorithm, But Not the Liability That Comes With It

If the first principle is that a human must own the outcome, the second is its uncomfortable corollary, and Shrey names it more directly than most are willing to. He calls it the governance paradox that nobody discusses honestly: the enterprise carries the reputational and regulatory risk, yet very often it does not own the algorithm generating the outcome.

When a vendor’s model misfires inside your deployment, the customer does not telephone the vendor. They telephone you. The regulator does not open a file on the vendor. It opens one on you. The liability, he argues, sits with the business that switched the system on, no matter who built it.

Ramana is equally unsentimental about the temptation to believe otherwise, calling it a dangerous corporate illusion to imagine that accountability can be handed off to a software provider. His image for it is exact: vendors supply the engine, but the enterprise drives the car. Contractually, a vendor’s liability is narrow and specific, confined to matters like service-level breaches, intellectual property infringement, data privacy violations, or gross negligence.

The commercial and operational fallout of a bad AI decision, however, is entirely an internal matter. No indemnification clause repairs broken customer trust, and none of them satisfy a regulator who has come knocking. Internal teams, therefore, are held wholly responsible for how a vendor’s model is fine-tuned, tested, and deployed within the specific contours of the business.

“

The liability sits with the business that deployed the system. The customer calls you. The regulator investigates you. Regardless of who built the algorithm.

Anupam Shrey — Co-Founder, Plutas.ai & Founder, DigiSafe

What separates these leaders from the wishful thinking they describe is that they have built machinery to hold the line. Shrey operates a model he calls structured shared liability, in which vendor contracts carry real AI performance warranties: explicit commitments on accuracy thresholds, fixed timelines for bias audits, and notification windows for drift.

These provisions are not yet standard across the industry, and he is candid that they should be. Alongside them, his teams run independent validation layers internally, testing a vendor’s outputs against their own actuarial benchmarks before anything reaches a live customer. A vendor that cannot produce explainability documentation for its model does not get to integrate.

Prerak Shah, the co-founder and chief technology officer of Cogniify.ai, describes a discipline that begins on the first day of any partnership, when everyone involved must understand exactly how responsibility will be divided through firm contracts, compliance standards, and operational transparency. Vendors are accountable for the integrity, security, and technical performance of the tools they provide, while internal teams carry the implementation, the ongoing use, and the final business decision.

The idea of shared accountability matters, he stresses, and it has a hard edge to it: the organization deploying the AI can never fully transfer its own accountability to the vendor. Bhobe anchors the same relationship in instrumentation. Because IDfy relies heavily on in-house models built to be explainable, its discipline with outside vendors is to know precisely what is sent to them and precisely what comes back, with the boundaries of responsibility ultimately settled in the contract.

The Move From Case-by-Case Improvisation to a Named Owner on Every System

One clear sign that AI accountability has matured is that these organizations have stopped deciding it incident by incident. Prerak Shah describes a governance posture that now recognizes several distinct layers of responsibility working together, with technical leaders, operational teams, and executive management each holding a defined part.

Ownership, in his telling, has settled into a formal structure instead of arriving incident by incident, because using AI more widely across a business is impossible without comprehensive frameworks for accountable use. Accountable leadership, he argues, can no longer be treated as optional. It is a structural condition for AI to grow inside an organization without growing dangerous.

Ramana describes the same evolution as leaving behind what he calls the Wild West phase of case-by-case experimentation. AI ownership at his companies is no longer filed away as an isolated IT project. It is treated as a core operational risk vector, governed by a matrixed model in which the technology leadership owns technical and data integrity, the legal and risk function owns regulatory compliance, and the line-of-business leader owns the commercial result. There is, he insists, no ambiguity in this arrangement, because before any AI system is deployed a specific human executive has signed off on the trade-off between its risk and its reward.

“

If everyone owns it, no one owns it. One executive, one system, clear authority. That is the only structure that survives an incident.

Anupam Shrey — Co-Founder, Plutas.ai & Founder, DigiSafe

Shrey is willing to show how his own organizations learned this lesson, and the honesty is telling. When DigiSafe first folded AI into its risk assessment workflows, ownership was loosely distributed: the technology team owned the model, the business team owned the customer relationship, and when something went wrong, responsibility moved between them like a hot potato.

The arrangement held until a climate data anomaly caused the parametric triggers to behave unexpectedly, forcing a significant recalibration. The post-mortem was useful, he says, but the ambiguity it exposed was not something he was prepared to live with.

Today the company runs a named-owner model, one executive per production system with documented authority to pause, modify, or decommission it. Each system is reviewed quarterly against business performance, fairness metrics, and regulatory exposure. The board receives an AI risk dashboard every quarter that reads as a leadership briefing, written in the language of which systems are operating within their parameters, what incidents occurred, and what regulatory deadlines are approaching, with the engineering detail left out.

That last detail points to a gap Shrey believes sits at the very top of most organizations. Reflecting on industry data showing that only a fraction of boards currently receive any AI-related metrics at all, he describes a governance hole precisely where accountability needs to land, and warns that technology teams cannot be expected to absorb strategic risk on leadership’s behalf indefinitely.

At IDfy, Bhobe reports that this clarity is already in place, with defined outcomes and named owners for AI systems built on close to a decade of deployment experience, and with responsibility assigned to the relevant domain teams as new use cases emerge. Aksheshkumar Shah characterizes his own company as moving steadily in the same direction, drawing the lines of responsibility among technical teams, management, and executives in advance so that no one is left improvising once a problem surfaces, on the conviction that AI can no longer be allowed to remain ambiguous in a serious business.

Governance That Lives Inside the Work, Not in a Slide Deck Reviewed Once a Month

When the conversation turns to how all of this is enforced, the leaders converge on a single idea: the governance that works is the kind you stop noticing, because it has been built into the daily rhythm of the work and not bolted on afterward as a ceremony. Shrey is direct about how he came to believe this, having learned the hard way that policy documents living outside the workflow are not really governance at all. They are, in his words, liability protection that happens not to protect you.

His framework begins with an AI system registry that catalogs every model in production, including those embedded inside third-party platforms, recording each one’s purpose, data sources, decision scope, risk classification, and named owner. You cannot govern what you cannot see, he points out, and in most organizations nobody can name every AI system touching a customer or a financial decision, which makes that invisibility a risk in its own right.

From that registry, Shrey’s systems are sorted into three tiers by consequence. Informational systems offer insight while a human decides; advisory systems strongly shape an outcome; and autonomous systems, where his payout triggers live, act directly and carry the most stringent controls. Every tier has defined escalation thresholds, pre-deployment bias reviews, and continuous drift monitoring with a twenty-four-hour resolution commitment when an alert fires. Every incident, including the near-misses, goes into a structured log that leadership reviews, because institutional memory matters more than people tend to think.

Ramana builds toward a similar robustness through what he describes as a defense-in-depth model, on the recognition that conventional software quality assurance was never designed for probabilistic systems. His framework rests on three pillars. Before any high-stakes model goes live, an independent internal team subjects it to adversarial red-teaming meant to force edge-case failures and surface hidden biases.

Once a model is in production, automated tripwires watch for drift and flag the moment its outputs wander beyond an acceptable variance, triggering a pause or a rollback. And any use case classified as high risk, meaning anything that touches customer finances, data privacy, or safety, must clear a standing cross-functional review board of legal, technical, and business leaders before a single line of code reaches production.

“

Governance that lives outside the workflow is just liability protection. The kind that works is built into the operating rhythm so deeply that nobody notices it is there.

Anupam Shrey — Co-Founder, Plutas.ai & Founder, DigiSafe

Bhobe describes a more understated but equally deliberate version of the same instinct at IDfy, where a low-confidence output or inference automatically triggers a human-in-the-loop workflow tailored to that particular use case, so that the system itself decides when to summon a person and no one has to remember to check. He places heavy emphasis on instrumenting inferences at every stage, leaving a clear decision trail that can later be audited or fed back to improve the models.

The two leaders from Cogniify.ai round out the picture with a governance posture built on recurring system audits, performance monitoring, bias detection, compliance review, and human validation before any output is allowed into a business-critical function. Prerak Shah underscores transparency and regular auditing as the spine of their model, with continual human oversight concentrated in the areas that affect people most directly, such as predictive analytics, customer service automation, and data intelligence, supported by legal protections, operational controls, and ethical guidelines that travel alongside the technology and keep pace with it as it changes.

Placing Human Judgment Where the Risk Truly Lives, Not Everywhere at Once

The final tension is the one every leader in this conversation is negotiating, because the whole promise of AI is speed and autonomy, and a clumsy insistence on human oversight can erode the very value the technology was meant to create. Shrey reframes the question with precision. The real question is no longer whether humans belong in the loop, but where exactly they belong within it.

In a system where money is meant to move the instant a weather threshold is crossed, requiring human sign-off on every payout would defeat the purpose of building it. So his team designed what he calls a human-at-the-threshold model. The AI operates autonomously inside pre-approved decision corridors defined by data quality, historical accuracy ranges, and payout ceilings, and human judgment is summoned only by the decisions that fall outside those corridors.

A routine payout within historical norms executes on its own. An anomalous climate pattern, an unusually large payout, or a degraded data source flags for human review before anything fires. Sampled audits then run continuously on the autonomous decisions to confirm the corridor is still calibrated correctly, an arrangement he calls imperfect but honest, because it is built around the real risk profile and makes no attempt to manufacture the appearance of control.

Ramana reaches the same conclusion through a concept he calls risk-stratified autonomy, on the logic that inserting a human bottleneck into every process destroys the very return the technology was supposed to deliver. His organizations segment by risk. For low-stakes, high-volume, reversible work such as internal data sorting or first-line query routing, the AI runs with full autonomy. For high-stakes, irreversible decisions, including credit approvals, vendor contract generation, or anything touching human welfare, a human-in-the-loop architecture is mandatory, with the AI acting as the fastest and most comprehensive analyst available while a human remains the final, accountable adjudicator.

“

Speed is a competitive advantage, but speed without guardrails is just a faster way to crash.

Venkat Ramana — CEO, NthEye & ValuePitch

Bhobe adds a layer of nuance that the others largely assume, which is that the word AI now covers an enormous range, from straightforward statistical models to large language models, and that the right choice depends on the criticality of the use case and the organization’s appetite for risk. Where explainability matters most, he says, the design may lean toward statistical or rule-based systems, while leaning on a language model brings its own requirement for human oversight depending on what is at stake.

The Cogniify leaders express the balance in plainer terms but no less firmly. Aksheshkumar Shah describes automating the repetitive work while ensuring that every critical decision keeps a human involved, on the view that AI should augment professionals and never replace responsible leaders. Prerak Shah frames it as a collaborative model of intelligence in which automation adds to productivity while a human expert keeps the decision-making meaningful, supplying the context, ethics, and strategic priorities that an autonomous system, left to itself, has no way to weigh.