Danny Goh, Co-founder & CEO, Nexus FrontierTech. He also co-founded the AI Native Foundation. Danny is a co-author of Becoming AI Native: Charting the Next AI Frontier (Routledge, 2026)
Terence Tse, Co-founder & Executive Director, Nexus FrontierTech. He also co-founded the AI Native Foundation. Ter is a co-author of Becoming AI Native: Charting the Next AI Frontier (Routledge, 2026)
Rob Casper, Principal, Ridgeview Digital
Many financial services companies would have the following experience: A team demonstrates an AI system that drafts contracts, flags fraud, or writes reports. The demo dazzles. The pilot performs even better. Then deployment stalls — not at the engineering desk, but at the risk committee.
The instinct is to up the technology and pour more resources into better models and guardrails. Each round produces a more impressive demo and even better performance. Yet, the same stall at the same committee. It turns out that the obstacle isn’t AI capability. It’s the existence of an accountability gap.
Accountability, not capability
Imagine a hospital that has the best-informed medical AI in the world — one that has read every journal and diagnoses faster than any human doctor. Would the hospital let it write prescriptions unsupervised? No, no matter how brilliant the system is. It’s not due to doubts as to its competence. Instead, medicine has a rule embedded in law: a prescription is only valid after a licensed physician signs it. The AI can prepare, but only the doctor can make it valid. Competence and accountability are two distinct things, and only one can be considered for delegation to a machine.
Regulated industries generally face a choice with AI. When a bank evaluates an AI system, the compliance team doesn’t ask how smart it is — it asks who answers for it if it’s wrong. Existing laws governing financial reporting, banking supervision, and outsourcing, from the US to Britain, the EU, and Singapore, all require naming a specific accountable person. None makes room for AI to be that person.
Two kinds of trust
Mistaking capability for accountability may stem from people using the word “trust” to mean two different things. The first is what software can prove about itself — its accuracy scores, its logs, and its track record. Call this asserted trust: it answers whether a system did what it was built to do, and it’s what the AI industry has focused on strengthening.
Yet what institutions actually need is business trust: a person with credentials and standing taking ownership for an outcome. An auditor signing a company’s accounts or a doctor signing a treatment order isn’t merely confirming that a process ran correctly — they are attaching their professional experience and standing to the result.
The problem is that the AI industry has largely assumed that if asserted trust gets strong enough, business trust will follow. It won’t. Piling up evidence doesn’t create an owner for an outcome. It just collects more evidence to show what the AI-powered process is capable of doing.
The missing step
To build business trust is to use a tool that predates computers by five thousand years: the signature with meaning behind it — the same device a doctor uses on a chart, an auditor uses on an opinion, and a chief executive uses, by law, on quarterly results. Only a signature turns an output into something on which a business can rely.
Most companies today build AI involving two steps: the system does the processing work, and then that outcome gets checked against a standard. What is missing is the third: a signature by a named person, with real authority, taking ownership. Skipping this means nobody takes responsibility for the AI outcomes.
Not all sign-offs are equal
In enterprise AI, sign-offs can be grouped into at least three levels. At the weakest level, a human simply clicks “approve,” with no indication to show they understood what they approved. It is oversight and supervision in appearance only. Much of today’s celebrated “human-in-the-loop” AI is, sadly, operating at this level.
One level deeper is similar to a contractual warranty. Here, a provider merely confirms that the AI output has met a specific standard. But still no individual’s name accountable for the result. What actually matters is the deepest level where a licensed professional reviews the work against a defined standard and signs off on it personally, with reputation on the line.
Building in accountability by design
The goal for companies in regulated industries when creating AI systems is to build a signature into such systems. In this case, when an executive signs off on an AI output, they can put their name behind it on solid ground, enabling a defensible, accountable deliverable.
This can be done with an approach we call “machine-first, human-final.” AI processes the volume. Standardized and normal cases are cleared automatically with a full record attached. Any cases beyond that are routed to a senior professional to review and sign, spending time on judgment rather than digging. A wealth manager, for instance, may today spot-check only a sliver of outgoing client statements. With this model, AI checks all of them, clearing most automatically and flagging only genuine exceptions for sign-off by professionals who can reliably determine whether a surfaced anomaly matters, whether a real risk exposure exists, and whether an edge case has emerged.
Indeed, such a design approach confers an advantage with which no AI model can compete: a growing, signed record of expert judgment, which, arguably, is an institution’s most valuable asset. This, in turn, creates a so-called compounding flywheel, in which every human correction becomes structured training data to further improve the AI in place.
Design, not technology
Business leaders need to stop designing AI around tasks and start designing it around that which needs human accountability. Managers would do well to remember that the winners in cloud computing weren’t the quickest to roll out the infrastructure but the ones who made compliance their selling point. The firms that win in the enterprise AI world tomorrow won’t be the ones with the most clever models. They’ll be the ones whose output a professional is willing to sign.
