Why 95% of Enterprise AI Fails, and What Government Can Teach Business
In conversation with Dr. Laura Gilbert, founder of i.AI, the UK government’s Incubator for Artificial Intelligence at 10 Downing Street.
Roughly 95% of enterprise AI deployments fail to deliver tangible benefits, says Dr. Laura Gilbert, because they start with the technology instead of the problem. She has seen it from the inside: she founded i.AI, an open-source AI engineering team within the UK government, and has advised nearly 50 governments worldwide.
A former Oxford particle physicist, hedge-fund AI modeller and healthtech co-founder, Laura now builds AI that solves real public problems. In this interview she explains why “AI-first” is a red flag, why radical transparency is the only route to public trust, and how competitive kickboxing shaped the way she leads.
Watch the full interview here.

Dr. Laura Gilbert,
AI, Data Science & Public Policy Expert
Key takeaways
- Start with the problem, not the tool. Declaring yourself “AI-first” shows you don’t understand the core issue.
- Trust is earned. Publish progress and open-source the code so others can inspect it.
- Passive “human-in-the-loop” creates false security. Design systems that demand active judgment, not a rubber-stamp “agree” button.
- Augment human empathy; don’t replace it. CADDI supports advisors rather than substituting for them.
How did you get from physics, finance and startups to 10 Downing Street?
Originally, I was a particle physicist at Oxford University. Overnight one Christmas, the government cut £80 million in funding for particle and astrophysics, causing immediate job losses for non-tenured staff. I was offered a position at Fermilab in Chicago, during a minus-40-degree winter, for $800 a month. I chose instead to move into quantitative finance, building market-trading AI models for hedge funds for three years.
It was technically fascinating, but generating returns for wealthy individuals wasn’t personally fulfilling. On the side, I helped a friend whose medical technology app had crashed after spending £10,000 with a developer. I co-founded that healthtech startup, left finance, and spent nine years building the business until we exited in March 2020.
During the COVID-19 lockdown, I spotted an advertisement for Director of Data Science at 10 Downing Street. I created 10DS, the Downing Street Data Science team, and grew it from four to 25 people within 18 months. After ChatGPT launched, I founded i.AI, an open-source engineering team inside government created to solve public sector problems directly with AI.
I think the biggest risk might be that we just don’t do it. We’re in a space where we’re not going to be able to run the National Health Service long term if we can’t afford it… We need the upside; we need the opportunities.
How did you address public skepticism when launching i.AI?
When i.AI was announced, the public response on social media split into two concerns. People either believed the government was too incompetent to build AI properly, or feared it would use AI malevolently to strip away benefits.
You cannot simply tell people to trust you; you have to actively prove it through competence and alignment with their best interests. I adopted a mantra of radical transparency. Breaking away from civil service norms that historically discouraged public discourse, we published our progress, spoke at conferences, and open-sourced our code on GitHub.
Open-sourcing our work lets citizens, security researchers and independent experts inspect it, offer feedback and catch vulnerabilities early. Trust in government is currently lower than trust in advertisers. Consistently demonstrating competence and publishing code transparently over time is the only way to rebuild public confidence.
How does CADDI, built for Citizens Advice, make government more human?
CADDI is an internal AI assistant that sits alongside advisors at Citizens Advice, the UK charity that helps people with legal, housing and financial problems. It relies on volunteers, and expecting them to memorize all UK legislation is impossible, so searching for answers manually during consultations was slow.
Rather than replacing human advisors with a public chatbot, we built CADDI to work next to them. When a citizen asks a complex question, the advisor prompts CADDI, which retrieves accurate policy answers in real time.
The benefit to Citizens Advice is not just that you get good advice, but you want to talk to a human who’s going to give you a cup of tea and a tissue and give you that human feedback and support that you don’t get from a chatbot.
In evaluations, CADDI achieved a 60% increase in successful case resolution speed and a 2.5-fold increase in advisor confidence. The core mission is using AI to augment human empathy, not to replace human connection with cold automation.
Why do so many enterprise and public sector AI projects fail?
Many organizations fall into a bandwagon trap, assuming that deploying a tool like Copilot means they have successfully adopted AI. In reality, roughly 95% of enterprise AI deployments fail to derive tangible benefits, because they start with technology rather than the underlying problem. Declaring you are “AI-first” shows you don’t understand the core issue.
The scale of the problem: A single UK public consultation generates text equivalent to 400 Brexit withdrawal agreements. Processing them manually costs £100 million a year and takes three to six months. We built an AI system that analyzed consultation themes within an hour.
But that example carries its own lesson. Public consultations are often flawed in the first place: they happen after policies are decided and tend to engage lobbyists rather than everyday citizens. We automated a process without first asking the core question: how do we bring citizen voices into policy creation earlier?
Organizations must stop focusing on “adopting AI.” Instead, identify what is broken, determine the outcome you want, and evaluate whether AI is the right tool.
How can AI systems keep humans truly accountable?
“Human-in-the-loop” is often misused in technical design. If an AI system generates a decision and presents a button for a human to click “agree,” behavioral incentives lead the person to approve passively without critical review. That increases risk by creating a false sense of security and encouraging people to tune out.
To solve this, we developed Policy Creator at the Tony Blair Institute, the policy think tank. It does not hand down pre-packaged answers. Instead, it requires the user to actively define outcomes, set policy parameters, and evaluate evidence-based trade-offs.
Policy Creator also simulates multi-agent stakeholder personas, representing everyone affected by a policy, from asylum seekers and police to human rights advocates and lifeguards, to show decision-makers how different groups will experience it. Genuine human-in-the-loop design forces active critical thinking rather than turning humans into rubber stamps.
How has competing in Savate kickboxing shaped your leadership?
I took up kickboxing as a university student despite having no athletic background. Initially, every time I got punched in the nose, I cried. But I realized I always stuck to things I was already good at, like math and music, so I forced myself to persevere through a difficult challenge until I mastered it.
Competing internationally taught me a vital psychological lesson. During multi-day tournaments, when I was exhausted, bruised and cutting weight, my brain would tempt me to give up so I could rest. What unlocked my motivation was realizing that my students were watching and relying on me to perform.
I learned that it’s difficult for me to go that final mile just because I want to. But if I feel like there are people relying on me, then I can.
I carried that mindset into Downing Street, whether fighting to bring data into high-stakes decision rooms or canceling wasteful multi-million-pound contracts. I apply it when mentoring female leaders, too. When a female colleague hesitated to claim credit for her work, I reminded her that staying quiet harms her, deprives the organization of strong leadership, and hurts everyone following behind her. Recognizing who relies on your leadership gives you the determination to stand up, take credit and push through difficult challenges.
Thank you Dr. Laura Gilbert for sharing your journey!