The EU AI Act’s high-risk system obligations must be applauded. However, the act only sets minimum, floor-level standards. Michael Eichsteadt, VP of Engineering, iManage, advocates the notion of practical AI, making the undeniable business case for this technology. He explains how embedding trust, transparency and governance into engineering processes can enable such an outcome.
AI is everywhere – embedded in tools, workflows and products. And yet for many organisations and individuals, it has yet to deliver the transformative value it promises. There remains a gap between the promise of AI and what it actually does.
Here is what’s easy to overlook: practical AI is only possible when trust is engineered in from the start, not bolted on at the end. Trust must be a foundational design requirement, alongside governance, explainability and transparency.
The EU AI Act’s high-risk system obligations, effective from February 3, 2026, deserves real credit. For the first time, there is a codified, enforceable standard holding AI developers to account for the risks their systems create. It is a meaningful step.
But the act sets a floor and not a ceiling. Trust-first engineering demands more. Trust, governance, explainability, security and transparency must be non-negotiable first-class requirements across the full AI development lifecycle. When companies get this right, they will put themselves on the road to delivering AI capabilities that users and organisations will not be able to live without.
The right mix of talent and mindset
Trust-first engineering starts at the top. Leadership has to make explicit the committment that trust is a priority. If it’s not centered on how decisions get made, it won’t be in design decisions.
From there, it’s a matter of assembling the right people. AI sits at the intersection of computer science, infrastructure engineering, mathematics, statistics and – in the Generative AI era – even linguistics because language and semantics are central to how these systems operate.
There is no single ‘unicorn’ hire who spans all of these fields – and no shortcut that substitutes for genuine depth across them. Development teams without this mix of skills will struggle to anticipate how AI systems can fail in ways that undermine trust.
When it comes to software development, modern Generative AI tools are genuine accelerators, but they do not replace the expertise required to build AI that is reliable, safe and trustworthy. They should complement development teams, not stand in for them.
The other reality is that everything in AI is experimental and rapidly changing. Nobody knows how a use case will look six months from now. Keep teams small and nimble, iterate quickly and have a prioritisation plan to manage every unexpected output. Get the people right, and everything else follows – including specifics around governance, transparency and feedback loops.
Engineering trust at multiple levels
Trust-first engineering operates at two distinct levels. Governance is preemptive in nature. Following sound governance principles and standards helps prevent a wide range of downstream failures. Think of it as the architectural blueprint for trustworthy AI.
Transparency is more reactive. It enables stakeholders to ‘look under the hood’ when something goes wrong – giving internal and external users meaningful visibility into what are otherwise opaque, black-box systems.
The two work in concert: governance defines what trustworthy looks like; transparency proves, on an ongoing basis, that you are delivering it.
Fine-tuning the feedback loop
There is one more key engineering discipline to keep in mind, and that’s the evaluation feedback loop for the models. Unlike traditional deterministic systems, AI models are probabilistic and non-deterministic – a fundamental difference from traditional systems that requires an entirely different approach to evaluation.
Legacy organisations accustomed to more deterministic logic can apply straightforward tests around if/then logic, quality testing and load testing. AI is a different animal. A user could enter the same prompt twice and receive different responses. And the usage pattern isn’t a slow, steady drip in the background – it’s highly variable and prone to sudden spikes. How do you test against these scenarios?
This is the ‘muscle’ that most legacy engineering teams haven’t built yet: a standardised way to evaluate these outputs to ensure that the model is performing as intended. It comes back to talent. The right mix of people is what ensures organisations can execute this final, crucial step: engineering trustworthiness directly into the AI systems themselves.
Trust-first isn’t just the right thing to do – it’s the smart thing to do
The argument for engineering trust-first AI systems is not just moral or regulatory. It is commercial. Where AI handles sensitive data, informs consequential decisions or operates inside regulated industries, the cost of a trust failure is not just a bad press cycle – it can be a damaged customer relationship, a regulatory investigation or a brand crisis.
The organisations that will build AI that users and businesses genuinely cannot live without are those that treat trust as a core feature: one that must be architected, tested and maintained with the same intentionality as any other part of the system.
Taking this approach is the surest path to making an undeniable business case for AI. Without trust, even the most advanced systems struggle to gain traction – a limitation no amount of technical sophistication can overcome.
