By: Natalie Johnson
When AI is inconsistent in a product review, it is annoying. When it is inconsistent in a hospital or a bank, it is a ruined life. Shekhar Natarajan has built the architecture designed to end it.
Two patients. Same age. Same symptoms. Same test results. Same hospital system. Their cases entered the clinical AI decision-support platform on the same day, minutes apart. The system returned different severity scores. One was flagged for urgent follow-up. One was not.
Neither patient knew this. Neither doctor knew why. The system did not explain itself. It produced an output (a number, a classification, a recommendation), and the number differed between the two, and the platform moved on.
This is not a hypothetical. Researchers testing clinical AI tools have documented exactly this pattern: the same underlying medical facts, described in different words, produce different outputs with equal confidence and no explanation. The system does not reason. It pattern-matches. And the pattern it matches depends on the surface form of the query, the words used, not the facts described. Change the words, change the answer.
“Ask the same question twice, get different answers,” Shekhar Natarajan, founder of Orchestro.AI and the architect of the Angelic Intelligence framework, has said. “For high-stakes decisions like loans or medical treatments, this inconsistency is catastrophic.” He uses the word deliberately. He means it precisely.
One Was for a Loan
Now think about the loan.
Two families. Identical financial profiles, with the same income, the same debt-to-income ratio, and the same credit history. One family submits their application on a Tuesday. The other submits on a Thursday. A field on one form is labeled “monthly income.” On the other hand, it says “income per month.” The system processes both. One application advances. One does not.
No human made this decision. No human can explain it. If either family asks why, the answer they will receive, if they receive one at all, is a version of the same non-answer: the algorithm said so. There is no appeals process that reaches the reasoning, because there is no reasoning to reach. There is only an output, and the output was different, and the difference will determine whether this family builds equity or not, whether their children’s school is in this neighborhood or that one, whether the next generation inherits something or starts from zero.
The two cases, the diagnosis and the loan, feel like different problems. They are not. They are the same problem, wearing different clothes. An architecture that produces inconsistent outputs when the surface form of a question varies will produce inconsistent outcomes wherever it is deployed. In the hospital, it is a treatment missed. In the bank, it is a life redirected. In the courtroom, where AI is informing sentencing recommendations in multiple US jurisdictions, it is a person who goes home to their family or does not. In the hiring platform, it is a career that advances or stalls permanently at the threshold.
In each case, the system performed exactly as it was built to. It was built for performance, at scale, on benchmarks that do not capture what happens in these rooms.
“The failure modes are not just different in degree. They are different in kind. We are not asking for better AI. We are asking for a different kind of AI entirely.”
Why Better Models Do Not Solve This
The industry’s response to inconsistency has followed a consistent pattern: more training data, larger models, better benchmarks. Each generation performs better than the last. The inconsistency persists.
It persists because the industry has been treating a structural problem as a capability problem. More data does not fix an architecture that was never designed to reason. A larger model does not fix a system that has no mechanism for deliberation. A better benchmark captures the problem more precisely; the next generation is optimized for that benchmark, the underlying architecture remains unchanged, and real-world inconsistency persists under conditions the benchmark did not measure.
The existing workarounds have been tried and found wanting. Explainable AI was meant to address transparency, but in most implementations, the explanation is generated by the same model that produced the decision, making it a retrospective rationalization rather than actual reasoning. Human oversight was meant to provide accountability, but in high-volume environments, reviewers process hundreds of automated outputs per shift, and the cognitive conditions for genuine independent judgment are structurally absent. Audit trails tell you who clicked what and when. They solve a documentation problem. They do not solve an accountability one.
None of these approaches reaches the architecture. And the architecture is the problem.
The Room Where Every Answer Is Deliberated
The framework Natarajan has built, Angelic Intelligence (protected by 80 patents), does not add a consistency layer on top of an existing system. It builds consistency into the foundation.
At the center are the 27 Digital Angels: specialized AI agents, each representing a virtue drawn from the moral traditions of every major human civilization, including compassion, justice, truth, protection, wisdom, fairness, and prudence. When a consequential decision is required, they deliberate. Every one of them. On the actual facts of the situation, not the surface form of the query that described it. The deliberation requires consensus. No single agent can override the others. The process is the same regardless of how the question was worded, because the agents are assessing the underlying human situation and its consequences, not parsing syntax.
The result is not a system instructed to be consistent. It is a system whose architecture makes inconsistency structurally difficult. Twenty-seven agents representing justice and compassion, and truth do not reach different conclusions about the same underlying situation because someone used “monthly income” instead of “income per month.”
And when the deliberation concludes, the reasoning is visible, not as a technical log or a probability score, but as a plain-language account of what each agent considered and why. The loan applicant whose application was declined can read it. The patient with the unexpected severity score can read it. The job applicant whose resume was screened out can read it. Not as a document requiring specialist interpretation. As an account, they can understand, contest, and, if the reasoning was wrong, challenge.
This distinction between a technical trace and a human-legible explanation is not cosmetic. A technical trace tells you what the system computed. A human-legible explanation tells you why in terms that the affected person can evaluate. Only the second kind of transparency produces accountability. Only the second kind allows a person to know whether they were treated fairly, and to say so if they were not.
The Same Answer, Every Time
Two patients. Same symptoms. Same test results. They deserve either the same severity score or an honest, explainable account of why their situations genuinely differ.
Two families. Same financial profile. They deserve the same loan outcome, or a clear explanation of a real difference that justifies treating them differently.
This is not a high bar. It is the minimum that consequential decision-making requires. It is what human professionals, including doctors, loan officers, and judges, are expected to provide when their decisions are challenged. It is what we do not currently ask of the automated systems that, in practice, have taken over those decisions in many of the most consequential contexts of human life.
The content Natarajan has published around Angelic Intelligence has attracted billions of views across social media platforms. Rooms full of the world’s most powerful decision-makers have stood when he speaks at Davos, Forbes Middle East, the Future Investment Initiative in Riyadh, and the AI Summit in New Delhi. Not because he is describing a technical architecture. Because he is describing something people already know, in their own lives, without having had the language for it: that the machine gave two different answers, and no one could tell them why, and the difference changed everything.
Trustworthiness requires consistency. Consistency requires architecture. And architecture, unlike a policy, guideline, or compliance checklist, cannot be circumvented by the people who built it. That is the point. That is what Natarajan is building. And that is what the two patients and the two families have always deserved.
Shekhar Natarajan is the Founder & CEO of Orchestro.AI and inventor of the Angelic Intelligence framework. He holds 207 patents and degrees from Georgia Tech, MIT, Harvard Business School, and IESE. His framework has generated billions of social media views and earned standing ovations at the World Economic Forum in Davos, Forbes Middle East, and the AI Summit India.
