A Conversation on AI, Trust, and What Startups Get Wrong
An Expert Perspective with Maya Murad, AI Specialist & MIT Researcher, Principal PM at Microsoft AI Frontiers
Views expressed are Maya's own and do not represent her employer.
There are people who talk about AI, and then there are people who have spent years inside it - building it, stress-testing it, watching it fail, and figuring out why. Maya Murad is firmly in the second camp.
A researcher who wrote her thesis on responsible AI at MIT, Maya has worked at the intersection of AI policy, product, and deployment for years. She has collaborated with NASA on foundation models for weather prediction, advised on responsible AI governance at scale, and today works at one of the world's most influential technology companies, Microsoft.
We sat down with Maya as part of our expert series on AI and robotics in industry. What followed was one of the most honest, wide-ranging conversations we have had on the subject.
"Responsible AI Is Not a Checklist. It's a Mindset."
The conversation opened with the question of responsible AI. It is a term that appears on almost every company's website, in almost every product announcement. But what does it actually mean in practice?
Maya's answer was simple yet rigorous;
"It's three things," she said. "Can you explain what your AI is doing and how it is being used? Can you prove it is not causing harm? And if something goes wrong, what is your strategy?"
She was quick to expand on what "harm" actually encompasses, a definition that is far broader than most organisations assume. It includes legal risk around data usage, model bias producing differential outcomes for certain groups, security vulnerabilities, and what she called adversarial attacks - where external parties deliberately attempt to hijack AI systems.
"Our understanding of the surface area of what can go wrong is evolving every day," she said. "What we are learning is changing day to day. So it is not a definitive science."
This is precisely why, she argued, responsible AI cannot be treated as a one-time exercise. It is not a framework you implement in year one and leave unchanged in year two. It requires dedicated expertise - red teamers, responsible AI specialists, governance checkpoints - and a culture of continuous monitoring post-deployment.
When pressed on whether the term is becoming more of a marketing badge than a genuine practice, Maya commented "Because it is such a generic term, it can be overloaded. With AI, you need specific expertise. It means you might not ship as fast as you want to ship. It just comes down to whether you’re willing to do the work, especially since there’s little regulation requiring it"
The uncomfortable implication: for many organisations, responsible AI is optional. And a lot of them are choosing to skip it.
The 90% Problem: Why AI Pilots Are Failing
Perhaps the most striking statistic Maya shared was this: according to research from MIT, approximately 90% of generative AI pilots fail.
"They call it pilot purgatory," she said. "A lot of companies jumped on the bandwagon, wanted to experiment, lots of proof of concepts were put in place - but either the proof of concepts didn't deliver on their goals, or there was no pathway to operationalise."
This is the gap that separates organisations genuinely transforming through AI from those endlessly experimenting with it. And according to Maya, this gap comes down to a few critical factors.
The companies that do make the leap tend to be startups - nimble enough to build AI-first from the ground up, without the burden of legacy systems and legacy thinking. They do not treat AI as a hammer looking for nails. They redesign their entire value proposition around what AI makes possible.
Incumbents face a harder road. They have existing systems, existing processes, and existing workforces to navigate. AI gets bolted onto the side of something rather than built into the core. And without a clear pathway from pilot to scale - which often requires significant organisational change - the pilot simply dies.
"The economics don't make sense yet for some of these," Maya added, "because the value to the end user and the overall value proposition is still unclear."
Her advice to organisations trying to escape pilot purgatory: make sure you are solving the most important problems. Make sure there is measurable value even at the proof-of-concept stage. And make sure there is a realistic pathway to scaling - including the data pipelines, the training, and the change management - before you start.
Data: The Unglamorous Foundation of Everything
Ask most people what the biggest barrier to AI adoption is, and they will say regulation, or talent, or cost. Maya's answer is more fundamental: data.
"Your insights are going to be only as strong as your data," she said. "And data is not as glamorous as training machine learning models. So it gets overlooked, or people don't invest the right amount."
In the era of large language models, the challenge has shifted from data quantity to data quality and context. It is not just about having data - it is about having clean, well-structured, accessible data that can be provided to AI systems in the right way, at the right time.
"AI alone is interesting," she said. "But AI plus context, that is where you are truly tapping into something useful. When the two come together and you engineer it well, this is where the real value is."
Her message to any startup building an AI-enabled product: getting your data in order is step zero. Not step three. Not something you figure out later. Step zero.
"Even if you are going down a minimal investment route - buying an off-the-shelf chatbot, plugging in an analytics solution - you still need to connect these systems to your data sources. And your data sources need to be clean. You cannot have real value without one without the other."
Trust: Earned, Not Declared
Trust in AI is not achieved by writing it into your brand values. It is earned consistently, over time, through diligence and transparency.
Maya's framework for building it is practical. Be transparent about what your system can and cannot do. Design your product experience to account for human oversight. Be honest about limitations. And when things go wrong - and they will - have an incident response plan ready before you launch, not after.
She recalled a telling example: an airline chatbot that gave a discount it was not authorised to give, after a user successfully pressured it into doing so. The story is funny until you realise it is also a liability, a brand crisis, and a case study in what happens when organisations deploy AI without thinking through failure scenarios.
"You really need to have some thoughtfulness in place even prior to launching on what you do in case of certain harms," she said. "What is your incident response plan?"
She also made a point that is easy to overlook: trust is not built in isolation. It is built in relationships - with end users, with customers, with stakeholders. And that requires more than technical rigour. It requires a genuine commitment to doing the right thing, even when no regulation is forcing you to.
"It is just about the duty of care," she said. "If you want to build trust, it is something you have to invest in."
AI and Jobs: "We Cannot Do Away With Experts"
The anxiety around AI and jobs is real. The headlines are alarming. Maya's take is more nuanced, and more grounded in the actual mechanics of how AI works in practice.
"I think these headlines are reflecting some macroeconomic factors at play and are not actually speaking to the impact that AI is having, at least from a long-term point of view," Maya said.
Her argument hinges on a point that is counterintuitive but compelling. Even in the tasks where AI performs best - coding, for example - the productivity gains are not as straightforward as they appear. The people who get the most out of AI are experts. Experts who know what good looks like. Who can spot when the output is wrong. Who can nudge the model in a different direction.
"If you give free reign to the AI, the quality of what an expert can produce at the helm of the AI is going to be better than zero-shotted that way."
And the flipside: junior talent that is not developed because organisations believe AI can replace them will leave a critical skills gap. You cannot grow experts without growing juniors first.
"All these companies that are no longer investing in juniors - they will get hit later on. We will need experts. And you do not become an expert overnight."
The future of work with AI is one of collaboration and reallocation - humans focusing on the creative, the contextual, the genuinely difficult - with AI handling the repetitive and the mechanical. But Maya is clear that this outcome is not automatic. It requires leaders who actively choose to make it a reality.
AI in Sustainability: Where the Real Opportunities Are
Given KAP's work in cleantech and the energy transition, we asked Maya where AI can genuinely accelerate progress in sustainability, and where the hype is getting ahead of reality.
During her previous role, her team collaborated with NASA to build what she described as the first foundation model for weather patterns, using high-resolution satellite imagery across multiple spectrums to predict and track major weather events, from cyclones to forest fires.
"I think we are applying knowledge that we actually did on the text domain, using a similar architecture, but on a new modality. And we are able to tap into this as a new source of data for prediction."
More broadly, she sees the strongest AI applications in sustainability coming from optimisation and predictive analytics - grid optimisation, supply chain efficiency, resource allocation. These are areas where traditional machine learning has already demonstrated value, and where generative AI is opening up new possibilities around research, synthesis, and decision support.
Her caution: avoid the trap of believing AI must be central to every solution. Sometimes it is a core differentiator. Sometimes it is complementary. Sometimes it is a distraction. The discipline is in knowing which one you are dealing with.
What Every Cleantech Startup Needs to Know
We closed by asking Maya what she would tell a cleantech or energy startup building AI into their product from day one.
First: establish honestly whether AI is genuinely central to your value proposition, complementary to it, or just noise. If it is core - if your problem is genuinely better solved with AI - then build AI-first. Hire AI talent. Build AI-first processes. Do not treat it as an add-on.
Second: if you are in the core differentiator category, you will likely be building models that have not been built before. That means partnering closely with domain experts - people who know the grid, the supply chain, the ecosystem you are operating in - not just AI engineers who understand the technology in the abstract.
Third:bake in responsible AI from the start. Not as a compliance exercise, but as a design principle.
Fourth: build a culture of experimentation. In AI, things are rarely linear. The winning solution often comes after multiple failed iterations. That is not a bug, it is the process.
And finally: keep one eye on regulation. In sustainability and energy, it is coming.
"The Bubble Nobody Wants to Talk About"
When we asked Maya about the flow of investment into AI - the billions pouring into startups, into infrastructure, into foundation models , her answer was direct
"I think we are in a massive AI bubble. I think investors' expectations of AI are miscalibrated. I think there are going to be some winners, but lots of losers."
She compared the current moment unfavourably to the dot-com bubble, arguing that at least the dot-com era built durable infrastructure. Fibre optic cables installed in 2000 are still in use today. By contrast, the data centre hardware being built now has a useful life of months before it needs upgrading.
Make of that what you will.
Maya Murad is an AI specialist and researcher with a background in responsible AI governance, product development, and machine learning. She completed her thesis on responsible AI at MIT. The views expressed in this interview are her own and do not represent those of her employer.
This interview is part of KAP's expert series on AI and robotics in industry.
Want to feature in our newsletter or expert series? Get in touch at info@kantaradvisory.com
