AI at Work

di Neil C. Hughes

Lavoro Tecnologia

Why Model Choice Is the Wrong AI Obsession

Why are enterprises spending so much time comparing AI models when their strongest competitive advantage may already be sitting inside their own data? My guest today is Emma McGrattan, Chief Technology Officer at Actian. Emma argues that model choice is becoming less distinctive while data quality, shared business meaning, governance, and culture increasingly determine whether AI works in production. We discuss why executives often rate data maturity more highly than the teams closest to the data, what changes when an autonomous agent becomes the consumer, and how data products and enforceable contracts can help organizations move beyond impressive prototypes. Emma also shares a remarkable example of an expense agent approving a $300,000 spend through one thousand individually permitted transactions. If your organization is trying to move AI from pilot to production, this conversation offers a practical place to begin. Episode description Why are so many enterprise AI programs still stuck in pilot mode when powerful models are widely available? In this episode of AI at Work, I speak with Emma McGrattan, Chief Technology Officer at Actian, about why model selection may be receiving more attention than it deserves. Emma argues that models are becoming increasingly interchangeable, while an organization’s proprietary data, business definitions, governance controls, and data culture remain much harder to reproduce. Emma explains why familiar terms such as customer, revenue, and churn can mean different things across sales, finance, tax, and operational teams. AI cannot reliably fill those gaps without context, ownership, lineage, and measurable quality. We discuss why enterprise data designed for dashboards and human interpretation must be treated differently when an AI system or autonomous agent becomes the consumer. The conversation examines Actian’s data governance research, including the finding that 83% of organizations face governance and compliance challenges and that executives rate data maturity 12 percentage points higher than operational managers. Emma describes what this gap looks like in practice, from duplicate customer records and undocumented pipelines to a business field called “Revenue Final Version 7 Verified and Revised.” We also discuss data products, data contracts, executable governance, and BARC research suggesting that organizations using data products and contracts were 3.4 times more likely to report success with AI at scale. Emma recommends beginning with one valuable use case, defining the data product it requires, agreeing the contract around it, and proving that the organization can deliver reliable results before expanding. Agentic AI raises the stakes further because the human who previously handled ambiguity may no longer be present at every step. Emma shares the example of an expense agent that approved one thousand $300 purchases. Every transaction sat below its individual limit, yet the combined spend reached $300,000. Finally, Emma explains what a strong data culture looks like. Employees need permission, access, and skills to question an AI answer, trace its lineage, and recognize when the result does not make sense. Learn

Who Controls AI Agents When Software Starts Acting Alone With Oasis Security

What happens when software can reason like a person, move at machine speed, and use every permission it receives? In this episode of AI at Work, I speak with Adam Ochayon, VP of Product Strategy at Oasis Security, about why autonomous AI agents create a different identity and access problem from the software businesses have secured for decades. Adam explains that traditional software is fast but generally deterministic, while people can reason but usually act slowly enough for organizations to intervene. AI agents combine reasoning with machine speed. They are goal-seeking, capable of choosing different routes to complete a task, and likely to use any access available to them. Permissions that might remain unused by an employee can become active risk almost immediately when assigned to an agent. We discuss why static roles are poorly suited to this behavior. An agent may need different permissions from one session or task to the next, depending on the person directing it, the data involved, its recent behavior and the action it is attempting. Adam argues for contextual authorization that can grant narrowly defined access at the right moment, monitor behavior continuously and provide a kill switch when an agent begins operating outside its approved purpose. Ownership presents another problem. Employees can create, modify and reuse agents across several business systems. Some agents act on behalf of a person, while others receive their own autonomous access. In both cases, companies need to know who remains accountable, which credentials the agent holds, what happens when its owner leaves and how its permissions are retired. Adam connects these questions with the wider problem of nonhuman identities, including service accounts, API keys and secrets. Companies that have not brought those identities under control may find AI adoption magnifying weaknesses that already exist. Security teams also face a delicate balance. Excessive friction encourages employees to bypass approved systems, while unrestricted access creates unacceptable exposure. Adam believes identity teams can become advisers to the business by creating controls that permit faster adoption with clearer boundaries. We also discuss Cyera’s announced $1 billion deal to acquire Oasis Security and what the combination of data security and access governance may signal about the direction of enterprise AI security. Adam closes with a practical sequence for leaders: discover which agents exist, understand their access, assign ownership, define policies, monitor activity and enforce controls across cloud and on-premises systems. Who owns your AI agents, and would your organization know when one moves beyond its approved purpose? Share your thoughts with me.

Measuring Whether AI Is Actually Working With Nexthink

How can a business tell whether workplace AI is producing a meaningful result rather than another encouraging adoption chart? In this episode of AI at Work, I speak with Scott Pope, Director of Value Advisory at Nexthink, about a problem facing many technology leaders. AI tools are reaching employees quickly, but deployment, usage, and business value are often treated as though they describe the same thing. They do not. A company can distribute thousands of licenses and report active users without knowing whether work became faster, easier, less expensive, or less frustrating. Scott argues that AI value has to be defined before a rollout begins. Productivity may matter most to a chief executive or HR leader, while a CFO may focus on cost and an IT support manager may watch ticket volumes. Each stakeholder is working with a different currency of value. Without a baseline, the business cannot measure the gap between its starting point and the result it hopes to achieve. That distinction matters because familiar IT measurements can create a misleading picture. Scott says a decline in support tickets does not automatically prove that the employee experience improved. People may have stopped reporting problems, created workarounds, or accepted friction as part of the job. Infrastructure can appear healthy while employees continue to lose time at the device, application, or workflow level. We discuss why digital employee experience, often shortened to DEX, has moved from a specialist IT concern into a wider business conversation. Work happens where employees interact with laptops, virtual desktops, mobile devices, applications, and services. Monitoring servers and cloud platforms remains useful, but it does not reveal every delay, failed interaction, or workaround experienced by the person trying to complete a task. Scott explains how observability can help organizations understand which AI tools employees are using, where adoption is deep or shallow, and which teams may need support. He is also careful to distinguish visibility from proof of value. Knowing that an employee opened an AI application is a starting point. It does not show whether the tool saved time, improved a decision, reduced cost, or produced a better customer result. The conversation also considers why one AI tool will not suit every role. Different teams work with different information, processes, risks, and desired outcomes. A persona based approach can help a business decide which technology fits the work rather than asking every employee to adopt the same product. It can also reveal where people need timely guidance instead of a training session delivered once and quickly forgotten. For Scott, the people question is where many programs become difficult. Providing access to software has become relatively straightforward, but changing established behavior takes communication, evidence, and a reason employees can recognize in their own work. Leaders often explain what AI could do for the business while giving less attention to the personal value for the person expected to use it. The opportunity is a workplace where technology problems are identified earlier, employees receive help at the moment they need it, and AI investments can be connected with measurable results. The risk is that businesses mistake purchasing and activity for progress while adoption becomes uneven and employees quietly carry the cost of poor implementation. Does your organization know what its AI tools are changing for employees, and which measure would give you the clearest answer? Listen to the episode and share your thoughts.

Redesigning Enterprise Work Around AI Employees With Ema

What happens when companies stop adding isolated AI tools and begin redesigning entire business processes around AI employees? In this episode of AI at Work, I speak with Surojit Chatterjee, founder and CEO of Ema, which stands for Enterprise Machine Assistant. We discuss why the debate about AI replacing jobs often misses the larger business question: how should organizations redesign work when intelligent systems can coordinate tasks, access enterprise knowledge and complete workflows across multiple applications? Surojit describes this as “agentic business transformation.” Instead of giving every employee another chatbot or assistant, organizations can use coordinated AI agents to manage processes that cross departments, systems and approval chains. People remain responsible for setting boundaries, reviewing sensitive decisions and deciding when an agent has earned greater autonomy. He shares the example of Wipro, where an Ema-powered system called WiproNow supports around 240,000 employees across 65 countries. According to Surojit, it covers approximately 70 use cases spanning the employee journey from recruitment to retirement, connecting with over 100 enterprise applications. The reported results show why workflow-level automation matters. Average response times for employee requests reportedly fell from five days to less than five seconds, while employee satisfaction increased by almost 20 percentage points. Surojit also says the number of people needed for this work fell from roughly 1,000 to 550, with employees reassigned to other areas. We also discuss why companies do not need perfect data before beginning. Surojit argues that capable AI systems can identify contradictions, missing information and undocumented processes as they work. This can expose the informal knowledge that organizations often discover only when an experienced employee leaves or goes on vacation. Trust remains the deciding factor. Surojit compares deploying an AI employee with hiring a talented new colleague. Leaders provide context, test performance, review early decisions and gradually increase autonomy. Clear boundaries remain necessary for sensitive issues involving areas such as employee relations, healthcare or financial decisions. The practical lesson is that meaningful AI returns come from redesigning work across teams rather than measuring prompts, tokens or individual productivity gains. Is your organization preparing AI to own complete workflows, or giving employees another tool to manage? Listen to the conversation and share your thoughts with me.

How SS&C Blue Prism Helps Businesses Escape AI Pilot Purgatory

Why are some businesses generating measurable value from AI while others remain surrounded by pilots, rising costs and impressive demonstrations that never reach daily operations? In this episode of AI at Work, I speak with Brad Hairston, Director of Strategy at SS&C Blue Prism, about the operational and cultural foundations that separate productive AI programs from expensive experimentation. Brad spent 30 years in consulting before joining SS&C Blue Prism around seven and a half years ago. He now works within the company’s Customer Zero program, which deploys SS&C’s automation technology internally before it reaches customers. Brad says the program has helped SS&C grow revenue by approximately one billion dollars without adding headcount. We discuss why AI programs should begin with the business outcome rather than the latest model. Brad explains why companies making progress connect their automation investments with corporate strategy, build on existing robotic process automation and create reusable governance, security, orchestration and measurement practices. Brad also challenges the idea that AI agents will replace every deterministic automation. Rules-based digital workers remain useful for predictable processes, while AI agents can support work that requires reasoning and adaptation. Combining both approaches can also provide greater control over cost. Our conversation examines what should happen before an AI agent receives permission to make payments, update customer records or initiate business processes. Brad recommends defined roles, limited permissions, human approval for higher-risk decisions, complete audit trails and an orchestration layer connecting agents with people, APIs and digital workers. We also discuss how companies can give employees access to no-code automation while maintaining common standards and oversight. Brad describes the federated model used inside SS&C, where individual business units build automations through shared platforms, templates and governance. For leaders feeling overwhelmed by daily announcements from OpenAI, Anthropic, Google and other providers, Brad offers simple advice: take a breath, return to the business problem and begin with a process where the outcome can be measured. Is your AI program building reusable capabilities with every deployment, or simply adding another experiment to the pilot queue? Please share your thoughts with me.

Keeping Human Intent at the Center of AI Creativity With Freepik

If anyone can produce a professional-looking image or video with AI, what will make audiences care about one piece of content over another? In this episode of AI at Work, I speak with Joaquín Cuenca, co-founder and CEO of Freepik, about how generative AI is changing creative work, business workflows, and access to professional production. Freepik serves over one million paid subscribers, while Joaquín says the platform attracts over 70 million monthly visitors. At that scale, Freepik has seen the difference between an impressive AI demonstration and a tool people can rely on for real creative work. Joaquín argues that generating something attractive is easy. Producing something that reflects a precise idea, maintains consistency, and creates an emotional response requires direction, judgment, and human intent. We also discuss what Joaquín calls the no-collar economy. His view is that lower production costs will allow individuals, smaller companies, and modestly funded creative teams to pursue projects that previously looked too expensive or risky. That could create opportunities for storytellers, photographers, audio specialists, performers, and other creative professionals. Joaquín also acknowledges that some existing roles will be affected as machines take over repeatable production work. For companies adopting creative AI, Joaquín recommends looking past licenses, activity, and content volume. Experimentation has value while teams are learning, but businesses eventually need to connect AI adoption with revenue, costs, brand performance, or another measurable return. We also consider the threat of AI slop. Better tools cannot provide taste, purpose, or a compelling story. As technical production becomes easier, those human qualities may become the greatest source of differentiation. Will easier production produce a new generation of creators, or will businesses fill every channel with forgettable content? Listen to the conversation and share your thoughts with me.

Keeping Humans Accountable in an AI First Workplace With Nansen

What does an AI first workplace look like when every employee has an agent but every person remains responsible for the outcome? In this episode of AI at Work, I speak with Alex Svanevik, co-founder and CEO of Nansen, about how his company is integrating AI agents into daily operations while retaining human judgment, security boundaries, and quality control. Nansen has around 80 employees, and Alex says each person has been given an AI agent. His own agent, Winnie, prepares draft agendas using previous meetings, company objectives, strategy, and cultural context. Alex then works with the agent to improve the agenda before the meeting begins. His use of AI extends beyond routine administration. Alex describes building the first version of a Nansen product through Telegram while walking with his daughter. By the time he returned home, the agent had created a working product that later became a command line interface used by thousands of people. There is also a lighter side to this deeply connected life. Alex and his wife occasionally use their respective agents to broker disagreements. As someone who has been married long enough to appreciate the commercial possibilities of automated diplomacy, I suspect this could become an unexpectedly popular category. The workplace message is serious. Nansen expects employees to use AI across much of their work, but Alex says the human must own the quality, output, and result. Employees cannot blame the tool for inaccurate, generic, or poorly reviewed work. Alex compares the review process with sending a disappointing meal back to the kitchen. The first output may be acceptable, but reaching a high standard often requires several rounds of feedback. He believes judgment and taste will become strong sources of differentiation as average quality becomes easier to produce. We also discuss the security tension surrounding workplace AI. Alex argues that companies must consider the risk of avoiding AI because attackers and competitors are using it. His preference is to provide employees with approved tools and safe environments rather than leave them to assemble uncontrolled alternatives. One of his most practical recommendations concerns machine readable information. Documents, code, designs, spreadsheets, and diagrams must be accessible to both employees and agents. Nansen has moved internal work toward GitHub repositories, Markdown documents, CSV files, and other formats agents can process. Making everything readable only by machines would create a different problem. People must retain the ability to inspect, understand, and approve the work. The aim is shared accessibility rather than transferring complete control to an agent. Evaluation becomes especially important when agents influence financial decisions. Nansen tests trading agents through backtesting, measuring whether they can interpret data, judge the significance of news, and produce profitable decisions. A separate optimizer or coach then recommends improvements to each agent’s strategy. Alex closes with four human traits he believes will matter in an AI first workplace: high agency, good problem selection, judgment and taste, and clear communication. Experimentation amplifies those qualities, provided people avoid unnecessary risk and retain ownership of the result. Could giving every employee an AI agent increase productivity while making personal accountability even more important? Listen to the episode and share your thoughts with me.

Rethinking Legal Work Through Agentic Law With Norm AI

In this episode of Tech Talks Daily, I speak with John Nay, founder and CEO of Norm Ai, about Agentic Law, AI native legal services, outcome based pricing, and the proposed legal framework for companies managed by AI agents. John has worked on the application of AI to law and public policy for around 14 years. His research predates the current generative AI era and includes GovDeVec, an early attempt to train neural networks on legal and government text so they could identify concepts embedded across large bodies of policy information. The arrival of frontier language models opened a different category of legal automation. Deterministic systems can complete forms and apply fixed rules, but language models can also examine precedent and guidance before applying it to a new situation. John separates this work into three layers. The first covers deterministic rules and repeatable automation. The second uses model based analysis to interpret documents and apply legal guidance. The third preserves human supervision for legal advice, consequential decisions, client communication, and final approval. We discuss how this structure works inside an enterprise. An AI agent could conduct an initial compliance review of marketing communications against SEC or FINRA rules. A human professional would then review the findings and complete the determination. Norm Law applies a similar model to legal services. Documents received during a transaction can be processed immediately by AI agents, with the results presented to an experienced attorney. The attorney decides whether to contact the client, negotiate with the counterparty, request additional information, or move the matter forward. For John, the value includes time savings and broader coverage. A legal team conducting due diligence may lack the time or economic incentive to inspect and cross reference every document in a data room. AI agents can examine a wider set of material and identify inconsistencies that could otherwise remain unnoticed. Outcome based pricing changes the incentive structure. A law firm charging a fixed price can use AI to review additional evidence without adding hourly fees to the client. John acknowledges the limitations. Predictable transactions can be priced around outcomes more easily than litigation where scope, duration, and strategy may change dramatically. The operating model also creates new roles. Norm brings together practicing attorneys, legal engineers, and AI engineers. Legal engineers translate professional knowledge and client preferences into agent behavior, while AI engineers build production systems and connect agents with live workflows. Another part of the conversation concerns supervisory AI. As companies deploy agents that advise customers or take commercial actions, human reviewers may be unable to inspect every decision at machine speed. Norm Ai is developing agents that monitor other agents for compliance with laws, regulations, and company policies. We also discuss Delaware’s proposed Artificial Intelligence Company initiative. The regulatory sandbox would test a legal entity managed by an AI agent while retaining human involvement, capitalization requirements, disclosure obligations, and government oversight. John argues that autonomous agents will increasingly take consequential economic actions. The policy question is whether this activity develops within established legal systems or moves toward jurisdictions and technical environments offering fewer controls. The supplied episode brief also provides significant company context. Norm Ai recently announced a $120 million Series C at a reported $1.2 billion valuation, bringing total funding above $260 million. Norm says organizations representing over $30 trillion in assets under management use its technology for legal and compliance work.

How Bridge Uses AI to Remove Workplace Communication Friction

How much productive time disappears because somebody misheard an instruction, missed part of a meeting, or could not fully express an idea? In this episode of AI at Work, I speak with Paul Lee, CEO of Bridge and InnoCaption, about communication friction and why it deserves greater attention in the workplace AI conversation. Bridge provides AI-powered real-time captioning, transcription, translation, meeting summaries, and meeting intelligence. InnoCaption provides AI and human-powered telephone captioning for eligible Americans who are deaf, hard of hearing, or have a speech disability. Paul explains how the experience gained from captioning over 30 million calls is informing Bridge’s approach to workplace communication. Research shared by Bridge says one in six working-age adults experiences hearing loss. It also reports that 37% of employees with hearing loss lose over five hours each week because of communication gaps, while nearly 20% lose over ten hours. Those losses can appear through repeated conversations, missed context, reworked tasks, and weaker decisions. Paul introduces the curb cut effect, named after the sidewalk ramps created for wheelchair users that also help parents with strollers, cyclists, and travelers carrying luggage. He believes workplace captions can produce a similar result. Technology designed for people facing the greatest communication barriers can improve comprehension, attention, and recall across a much wider workforce. We also discuss how accurate transcription can turn meetings into searchable company knowledge. Paul shares how his own team uses AI to consolidate brainstorming notes and reduce 100 ideas to a manageable set of choices. The system organizes the information, while people remain responsible for deciding what happens next. Paul also considers multilingual collaboration, AI translation that preserves meaning and nuance, and why AI ROI should include decision quality, participation, knowledge retention, and product development speed alongside immediate time savings. For business leaders, his advice is to understand work at the department, team, and individual levels before choosing a tool. Setting an arbitrary AI adoption target can create poor incentives, while studying repetitive tasks and employee frustrations can reveal where AI will offer genuine value. Where is communication friction quietly consuming time inside your company, and could accessibility technology help everyone participate more fully? Listen to the conversation and share your thoughts with me.

How Taxd Is Building AI Tax Automation With Humans in the Loop

Would you trust an AI system to prepare your taxes if it could not reliably tell HMRC guidance from information published by the IRS? In this episode of AI at Work, I speak with Arjun Kumar, cofounder of Taxd, about AI tax automation, digital tax filing, and the continuing role of human judgment in regulated financial services. Arjun began his career at PwC after joining through a school-leaver program. While working in expat tax, he and his cofounder saw how professional services firms often relied on offshoring and annual cost reductions rather than sustained investment in technology. Their attempt to promote a different approach internally eventually led them to create Taxd during the pandemic. We discuss Arjun’s prediction that routine tax compliance will become increasingly autonomous. When the required data already exists across tax portals, bank accounts, payroll systems, investment platforms, and brokerages, software can connect those sources and complete much of the repetitive work. AI can also help review hundreds of transactions for landlords, sole traders, and small business owners. However, tax advice often depends on jurisdiction, personal circumstances, and overlapping rules. Arjun recalls seeing customers use AI as a tax advisor, only to receive guidance drawn from the wrong country. A confident answer from a chatbot can become expensive when HMRC and the IRS are discussing entirely different tax systems. Arjun explains why Taxd combines software and AI with access to human accountants. We also discuss real-time tax reporting, Making Tax Digital, privacy, anonymized data, and how patterns across tax filings can help customers identify relevant deductions and questions. For founders, Arjun shares why specialist edge cases can provide a strong opening. Taxd began with expat tax, using its founders’ existing knowledge to serve customers whose needs were often poorly covered by general accounting services. Could your business automate routine compliance while preserving human responsibility for the decisions that carry real consequences? Listen to the episode and share your thoughts with me. **The Team at TAXD have kindly offered a discount code for listeners of the podcast. Use TECHTALKS to get 10% off any tax filing services (please note, this applies to filing only and excludes our advisory services).

1 di 6