The AI Adoption Podcast

One agent per person is manageable. A million per person may not be. As AI agents multiply, the question for leaders shifts from actions and decisions agents can take to who remains in control of them.Ramesh Raskar, Associate Professor at MIT Media Lab and leader of Project NANDA, argues that the value in AI will accrue to networks of agents rather than to a single centralised model or to artificial general intelligence (AGI).Ramesh contends that agents differ from Internet of Things devices because they can settle protocols between themselves, carry general knowledge, and negotiate and form coalitions. He maintains that regulating AI models alone is like making engines safe while ignoring the rules of the road. He expects the bulk of AI to run on home and office machines rather than in the cloud, which, if networked AI prevails, would leave today's data centre build out ahead of demand. He also warns that agent work becomes addictive leading to unhealthy working practices.• A $100 marketplace in which millions of agents compete and collaborate to complete a job with payment going only to the agents that did the work successfully.• Meta's Muse as "the AOL of the mid 90s": it removes friction, but inside a walled garden.• The new HAL, the human agent loop, which Ramesh calls "maybe the most important problem" for researchers.Listen before your organisation signs its next AI infrastructure commitment. Listen to get insights into the future of AI.Chapters00:00 Networked AI, not AGI, is where the value will sit04:38 Three traits that set agents apart from the Internet of Things11:10 Bicycle or crutch: the risk of cognitive surrender17:42 Edge AI moves agents onto home machines20:17 Data centres: needed, but perhaps not at this pace23:17 Renting AI versus owning it25:12 The ratio at which humans lose control, and the new HAL27:16 Managing digital labour without drowning in agent requests30:16 Agent work becomes addictive

Asked whether it would advise a white man differently, Claude said it would. It named the differences: charge more, take a different approach. The bias surfaced only because my guest thought to ask.Leah-Sunshine Garrett, Founder of Dawn Intelligence and co-host and co-author of AI for Equity, argues that AI literacy is not AI fluency, and that casual users of AI will not even consider the bias in the answers they receive.Leah makes the case that users should be sensitive to the information shared with a model about, for instance, their gender and race. They should question where every answer comes from rather than accepting the first one. She argues that organisations need oversight and community discussion before adopting any tool: whether it is needed, whose data feeds it, who it might harm, and who owns the company behind the model. She is critical of Grok, which she says is being used by the far right to spread misinformation and to create deepfake images of women and children, with very little internal or external regulation. On recruitment, she describes a loop in which fewer entry level roles drive mass applications, which drive AI screening, which many companies say is filtering out strong candidates.• Asked what it would tell her if she were a white man, Claude told Leah it would advise her to charge more and take a different approach.• "I could automate this, but should I?": Leah's test for organisations, and for a generation that she says cannot boycott AI completely because "it is everywhere".• Over 500 job applications after redundancy, then over 4,000 applications in three weeks when Leah hired for a marketing role at Dawn Intelligence.• Screening tools that reject CVs without the exact month and year of every role, and interview bots that reject candidates who look away for 10 seconds.• Dawn Intelligence, which Leah says forecasts incidents of harm to 98% accuracy, down to street level and the hour.Press play before your next AI procurement decision: Leah's questions belong in the meeting that approves it.Chapters• 00:00 Gen Z, pervasive AI, and the need for literacy• 03:16 Why equity matters in AI development• 05:21 How language models can amplify bias• 07:42 Questions companies should ask before adopting AI• 10:10 Who are large language models built to serve?• 12:15 AI as an assistant, not a source of truth• 13:55 Using AI to forecast and prevent harm• 15:26 Deepfakes, regulation, and protecting young people• 20:22 AI recruitment and the entry-level jobs squeeze• 27:24 Neurodiversity, communication, and AI• 29:32 Why equitable AI matters to society

Tokens are a form of money that few employees or AI users understand. Leaders can budget for a salary with confidence. They cannot forecast an AI agent's token usage with any accuracy, and once that agent runs a core service, switching it off stops being a realistic option.Dimitris Perdikou, Digital Advisor at Version 1 and former Government Chief Engineer at the Department for Science, Innovation and Technology, argues that the future team mixes people and AI agents, and that its running cost needs managing upfront and throughout.Digital roles are changing, for instance, delivery managers generate board packs automatically, user researchers spend more time with users, and Dimitris expects broader, hybrid roles to emerge over the next six to twelve months. Cutting entry level roles because AI automates their mundane tasks is, in his view, shortsighted, since new entrants bring fresh thinking. Tokenomics, the emerging discipline of managing the tokens an AI agent consumes, depends on guardrails, model choice, user limits and regular review. Yet, token costs can only be estimated. Outcome based contracts can shift some of that risk from organisations to suppliers, which matters for small and medium sized businesses with little financial cushion.• Tokens described as "a bit of a monopoly money type thing", priced in units few employees understand.• A product team of a business area specialist, a user designer and three or four engineers, in place of eight engineers and a product manager.• Half an hour every Friday afternoon reviewing what agents did and what they cost.• Token limits for each user at Version 1, with a process to extend them where worthwhile.• Some chief financial officers putting the brakes on AI implementation, and the approach that brought finance teams on board.Listen before your next budget round commits to an agent you cannot afford to pause.Chapters00:00 Understanding Business Value in AI Adoption02:11 The Evolution of Digital Roles05:04 AI's Impact on Work Dynamics09:13 Data Accessibility Challenges11:08 The Future of Entry-Level Roles14:18 Balancing Soft and Technical Skills15:53 The Role of AI Agents in Organisations18:35 Explaining Tokenomics23:22 Managing AI Costs and Resources28:18 Building Confidence in AI Adoption31:19 Comparing Public and Private Sector Experiences

Leadership teams can explain how they make money, but not their data. Ask about data architecture, and the reply my guest voices from the boardroom is a question: “Why do I need to know that?” This conversation makes suggests that this is a cultural problem rather than a technical one.Claire Thompson, Chief Data Analytics Officer for the Corporate Bank at Barclays, argues that AI is primarily a people and culture challenge. An organisation’s culture decides the extent to which AI opportunities becomes part of everyday work.Claire makes the case that generative AI is probabilistic rather than deterministic. It can give a different answer to the same question each time, and that answer may not be right, so it cannot be rolled out like a conventional digital programme. She argues that disconnected AI pilots deliver some value but not the value that matters, and that investments should be anchored to business or client outcomes. Governance, in her view, is much easier to put in at the outset rather than once agents are sprawling across the organisation. She contends that the more people around the leadership table who understand AI, data and technology, the more likely an organisation is to succeed.• Three routes bring AI into an organisation: AI tools in everyone’s hands, capability vendors build into their applications, and bespoke internal models. Each needs different governance.• The scaling checkpoint: personal use within clear data boundaries is fine; share a tool with 100 people and it goes through triage.• The hidden dependency: a locally built tool becomes critical to a process, then fails, with no backup plan.• Role modelling: leaders who say “it’s not a theory, this is how I’ve done it” move colleagues from curiosity to confidence.• Job anxiety is not new in data analytics; the value sits in “the valuable knowledge that people have in their heads”.Listen before your next AI pilot is approved.Chapters00:00 AI as a People and Culture Challenge05:06 Transforming Processes with AI09:47 Successful AI Adoption Strategies11:34 Addressing Job Security Concerns with AI12:41 Human-Agentic Teams in the Workplace13:35 Senior Management’s Role in AI Implementation15:24 The Data Management Challenge16:54 Moving from Local AI Use to Strategic Implementation18:21 Governance in AI Adoption23:45 Improving Data Maturity in Organisations25:37 Cultural Implications of AI Adoption

Opening a hydrant map, as an example of vital infrastructure, to AI speeds up repairs and widens access. The data that tells a crew to fix the hydrant near the hospital first could also show an adversary where a city is weakest.Dino Ravnić, Co-founder and CEO of GIS Cloud, argues that AI is the final step in opening geographic information systems (GIS), the software organisations use to record and analyse their physical assets by location, to people who are not specialists, and that this works only with a layer of control between the data and the AI.Dino makes the case for AI as a virtual colleague in the workflow, handling triage such as comparing a field photo of a hydrant with what was reported, while a person signs off. Feeding an entire dataset into a large language model, in Dino's view, burns budgets on tokens and invites hallucination, so a GIS should sit in between as the system the AI queries. On sovereignty, Dino observes that customers often want data kept within the EU and that a vendor's origin matters more than before. The conversation closes on citizens, for whom public data on schools, kindergartens and outages already exists but sits in portals that are hard to use.• A solar plant built with each panel recorded as a polygon on a map, so decision makers see progress from frames to wiring in real time.• An example of GIS would be hydrant repairs ranked by proximity to schools, hospitals and homes for elderly people who are hard to evacuate in a fire.• A separate permission in GIS Cloud governing whether a user may read or edit shared data through AI.• "Once you start to share the data, then you lose control." Dino Ravnić• Shadow AI: staff dragging data into ChatGPT or Claude without the organisation being aware.Press play to understand the interplay between GIS that holds a map of assets, AI and geopolitics. Chapters00:00 The Importance of Data Control in AI04:41 Use Cases of GIS in Organisations07:52 Integrating AI with GIS09:50 AI vs. Internet of Things: Understanding the Difference11:15 Democratisation of GIS: Making Data Accessible13:54 Enhancing Decision-Making with GIS17:09 Geopolitical Implications of GIS Data18:16 Governance and Sovereignty in GIS Data20:56 Balancing Democratisation and Control24:40 The Impact of GIS on Citizens

This is Part Two of a two-part conversation, and it moves from the global picture to adoption inside the organisation.Ericson Chan, Group Chief Information and Digital Officer at Zurich Insurance Group, makes the case that boards should judge AI by the value it delivers across a portfolio, not by a pipeline of individual projects, proofs of concept and use cases. Ericson explains that agents mark a shift from generating content to executing work, so an unreliable answer can become an unreliable action inside a business process. For now, Zurich orchestrates its agents through a deterministic workflow, with a predefined path from one agent to the next, and gives each agent an operating envelope. He argues that the transformation is gradual but not sequential, because technical capability is moving faster than organisations can absorb it. Boards, in his view, must govern AI as a portfolio that is an integral part of the enterprise strategy and capital allocation.• ROAI, Ericson's term for return on AI, judged across a portfolio rather than demanded from every experiment• An operating envelope for each agent: identity, permissions, data access, permitted actions, escalation thresholds and monitoring• Zurich's Hyper Challenge agentic AI hackathons, with hundreds of participants from 20 to 30 countries• "Half the team will never sleep": managers leading teams of people and agents• Junior tasks such as research, drafting and reconciliation, and an apprenticeship model Ericson says must be rethoughtListen to the closing exchange, where Ericson turns the question on the host and students and new employees turn out to face the same test of judgement.Chapters00:00 Holistic Value in AI Adoption02:40 The Rise of Agentic AI06:36 Understanding Deterministic vs Non-Deterministic Approaches08:14 Gradual or Radical Transformation in AI Implementation09:48 Governance and Strategic Planning for AI11:43 Managing AI as a Portfolio13:14 Collaboration Between Humans and AI Agents16:07 Cultural Shifts in Organisations with AI18:53 AI Literacy for the Next Generation22:22 Balancing AI Automation and Learning Opportunities

Three regions are building AI for three different reasons. The US pushes frontier models towards AGI, China embeds applied AI into manufacturing and logistics, and Europe builds trust through risk based regulation. A global insurer has to operate inside the three at once.Ericson Chan, Group Chief Information and Digital Officer at Zurich Insurance Group, makes the case that a multinational needs one AI strategy and variations on implementation.Ericson argues that regulators in Europe, the US and China start from different places but converge on safety, transparency and accountability, which Zurich codifies as STAR. The harder divergence is cultural, and an insightful metaphor is customers raised on Astro Boy and Doraemon see technology as a helper, while Western science fiction offers Terminator and Skynet. He contends that AI owned by the technology function fails, because the operating business must own process redesign, adoption and realise value. Productivity gains, in his view, become a bar competitors will cross, so the larger opportunity lies in redesigning the value chain end to end and preventing risk before it happens.• STAR: Security, Transparency, Accountability and Reliability as Zurich's non-negotiable principles across jurisdictions• 75 to 80 percent of Zurich's workforce actively using AI tools, up from a quarter two years earlier• A global AI strategy as a constitution backed by a playbook, with local teams owning implementation and value• Zurich's AI360 strategy: AI for Everyone, AI at the Core and AI for the Future• Small language models based on decades of Zurich's own knowledge as a source of competitive advantageThis is Part One of a two-part conversation. Listen for the regional picture before Part Two turns to adoption and implementation in detail next week.Chapters00:00 The Importance of Transparency in Leadership03:16 Global Perspectives on AI Development07:53 AI as a Threat and Opportunity08:39 Challenges and Opportunities in Scaling AI12:39 Cultural Influences on AI Adoption14:44 The Need for a Unified AI Strategy17:56 Coordination in Decentralised Implementation19:23 Leadership Priorities in AI Adoption23:57 Transforming Business with AI26:52 Zurich's AI Strategy: AI 360

The Home Secretary has charged PoliceAI with making policing in England and Wales "the most AI enabled police force in the world", yet each force keeps its own technology, risk appetite and culture. Proving that an AI tool works is not the hard part; scaling it within one force and then from one force to the next is.Dylan Alldridge, Deputy Director for Strategy, Oversight and Coordination at PoliceAI, argues that national adoption will be won through evidence, education and unglamorous paperwork and bureaucracy reductions.Dylan makes the case that AI should first target the "unsexy", process heavy work of case files, disclosure schedules and transcription, freeing time officers can spend with victims. He sets the Home Office target of around six million hours, or 3,000 full time equivalent roles, against workforce fears about jobs, and is candid that the use of released time rests with each force's chief officers. He treats public trust as something earned by showing as well as telling, through an AI registry every force will be required to maintain. He also names the risk of moving early: vendor lock in and rising token costs, the metered charges for using AI models, which could leave the public sector dependent on AI it cannot afford.• Pathfinder forces trial and evidence a tool before the case is made to scale it across policing.• "Two, three hours a day" back in an officer's pocket is the prize from fewer bureaucratic tasks.• The target is not a brief "to find 3000 people's jobs", and there is a recognition that jobs will change.• A public AI registry lets any citizen scrutinise their force's use of AI and the testing behind it.• AI, in a phrase Dylan credits to PoliceAI's director Alex Murray, can "put the humanity back into policing".Press play for a frank account of adopting AI across 43 organisations that the public perceives as one.Chapters00:00 Introducing PoliceAI and the case for a national centre05:13 Scaling AI across 43 independent forces07:03 Pathfinder forces and the unglamorous paperwork10:00 Culture, ease of use and the six million hours target14:05 Roles, jobs and time back for officers16:39 Data, governance and consistent standards21:51 Public trust and the AI registry25:02 The impact on victims and society27:12 Acting early and the risk of unaffordable AI31:18 The victim comes first

NatWest holds every AI system to Europe's standard, wherever it runs. One rulebook across the UK, the EU and North America, chosen because governing to several standards at once multiplies the risk work rather than reducing it.Paul Dongha, Head of Responsible AI and AI Strategy at NatWest Group, makes the case that AI governance earns its keep only when it reaches the branch counter and the customer, not just the risk committee or the AI Ethics Panel.The conversation moves from the enterprise risk frameworks banks have run for decades to the ways those frameworks bend, and sometimes strain, under autonomous agents. Paul argues that trust, not compliance, is the real asset AI governance protects. He sets out the need for a diverse ethics panel that argues each use case beats any scoring formula, and the provision of role-based training reaches staff who never touch an AI model but still field customer questions about it.Highlights• NatWest applies the EU AI Act standard to its AI internationally, so a system built in the UK can move into the EU with no further compliance work.• The EU AI Act sorts systems into four tiers, from prohibited to low risk; a credit lending decision counts as high risk because it changes a person's life.• Guardrails block toxic output and stop confidential data leaving the building; hallucinations, factually wrong output, are treated as a managed risk.• Persona-based training runs across the bank, from data scientists testing for bias to branch colleagues who need to answer a customer's question about AI.• Ethics trade-offs are settled by reasoned debate against the bank's values, in Paul's words "not computational. You can't apply a formula."Press play, for a working blueprint of AI governance that reaches the customer, not just the committee.Chapters00:00 The dual nature of AI: risks and benefits02:00 Responsible AI in banking03:09 Governance of agentic AI in financial services08:29 Aligning with the EU AI Act12:06 Risk assessment and categorisation in AI15:16 Cross-jurisdictional governance challenges17:30 The importance of trust in AI governance20:19 AI literacy: building a knowledgeable workforce22:33 Persona-based training for AI governance24:46 The role of AI ethics councils29:11 Balancing business needs and ethical governance

Enterprise AI is not working at the speed the market promised. The reason is not the models. It is the complexity that decades of growth leave inside any company. Dhiraj Rajaram, founder and CEO of Mu Sigma, argues that databases record actions and not the decisions and perceptions that produced the actions.He makes the case that many business cases fail because leaders solve problems in the vertical silos, as separate nodes rather than a connected constellation or network. He sets out a field of context built from three traces, actions, decisions and perceptions. He argues that the unit that matters is the decision, not the data point. He offers a physics analogy that treats data as mass and context as energy, and holds that a great deal of an organisation's context is still dark, as influences of human judgement and organisation culture remain hidden. Some highlights from the conversation.● A store's records show 32 door frames sold in a month, not the discounts approved or the weather that drove the sales.● The big D is decisions, not data; leaders optimise truth against time on a continuous basis.● Spaghettification: decades of markets, geographies, products and acquisitions leave complexity that blocks the business case.● A home improvement retailer reframes contractors from a segment to own into a distribution channel to collaborate with.● Context behaves like a field: a small amount of content creates a large amount of context, and much of it remains dark.For any board funding agentic enterprise AI, this conversation reframes the work that has to happen first.Chapters● 00:00 Challenges in building enterprise-level AI business cases● 04:45 Complexity in large organisations and its impact on AI● 07:03 Problems as constellations, not silos● 10:42 From databases to a field of context● 12:05 Making decisions more visible with AI● 18:24 Unpacking action, perception, and decision traces● 20:57 The three I's: intention, intelligence, initiative● 24:31 Real-world example: problem redefinition in organisations● 32:28 Key takeaways and future steps for enterprise AI