AI at Work

If your organization switched off its AI systems tomorrow, would customers, employees or the finance team notice any meaningful difference?In this episode of AI at Work, I speak with Frank Böhmer, founder and CEO of Ninox, about the growing gap between AI activity and measurable business value. Companies can count tokens, prompts, users, copilots and prototypes, but those numbers mainly reveal participation. They do not show whether a customer issue was resolved faster, an audit passed, an error rate fell or a team handled additional work.Frank argues that token consumption is a particularly weak proxy for success. An organization would never judge an employee by the number of meetings attended or the number of words typed into email. Paying attention to how many tokens people consume creates a similar incentive. The CFO wants an operational result, not proof that a model stayed busy.That distinction becomes important as natural language tools make software creation available to people who could not previously build applications. Frank sees genuine value in vibe coding because it reduces the cost of experimentation. Someone can describe an idea and create a convincing prototype in hours. The difficulty begins when that prototype meets the operational world.A production application needs reliable data, authentication, role based permissions, testing, integrations, audit logs and clear ownership. It must continue working as the organization grows, regulations change, employees leave and processes evolve. Frank suggests a more demanding question than asking whether an application can be built in an afternoon: can the company trust it three years later?We discuss where developers fit when AI can generate interfaces and basic code. Frank believes experienced developers move toward architecture, security, integration, data modeling and the consequences of design decisions. Plausible output can conceal incorrect assumptions and forgotten edge cases. As the volume of generated software grows, businesses need people who can distinguish a durable design from a shortcut that only works today.The same reasoning applies to process automation. Automating each step of an inefficient workflow can create a faster inefficient workflow. Frank encourages teams to reconsider the process using the capabilities now available. If AI can bring the right information to the employee immediately, some handoffs, searches and approval steps may no longer be required.For leaders assessing an AI project, Frank offers both a rigorous method and a simple test. The rigorous version begins with a baseline covering process time, cost, errors, delays, people involved and the final outcome. The company then compares those measures with the AI supported version. The simpler test asks what would happen if the agent were switched off tomorrow. If the business barely notices, the project may be an experiment rather than an operating capability.We also revisit shadow IT. Business teams create their own solutions because central IT can never satisfy every request within the available time and budget. Banning creation will not remove that demand. Frank recommends an approved environment where employees can build quickly while identity management, security, backups, auditability and deployment remain under organizational control. This changes IT from the team expected to build everything into the team that provides the safe foundation.Should leaders celebrate wider AI adoption, or reserve the word success for the moment a process becomes measurably better? Listen to the episode and share your thoughts with me.

In this episode of AI at Work, I speak with Rohan Kodialam, co-founder and CEO of Sphinx AI, about the difference between an AI system that sounds capable and one that understands how a business defines truth. Sphinx AI describes itself as an AI research company building human-interpretable context graphs that can support agents across different models and enterprise systems.Rohan argues that consumer AI can feel impressive because many everyday questions are supported by information in the public domain. The situation changes inside a company. Questions about customer churn, product performance, acquisition costs, or operational efficiency depend on definitions, assumptions, and working practices that belong to that organization. A model may reach the data through an enterprise connector and still lack the knowledge needed to interpret it correctly.The most revealing example comes from a retailer using a metric called revenue per store. One team divided revenue by every store in the estate. Another used only stores that were open on the relevant day. Both definitions were reasonable for their intended users, but an AI relying on retrieved documents could select one almost at random and return an answer that looked convincing. The error might remain hidden until somebody compared it with a manually prepared report.Rohan says the problem can be resolved by making the missing business context explicit. In this case, the AI needed a short explanation describing both definitions and when each should be used. It could then consider whether the request came from a real estate or operations user and select the relevant metric. This is a small example of a much larger enterprise problem: the information that separates a useful answer from a costly mistake often lives in people’s experience rather than a formal system.We discuss why retrieval augmented generation can struggle when company material is raw, contradictory, or unsynthesized. Rohan reports that Sphinx AI customers have seen accuracy increases of roughly 35 to 40 percent on existing enterprise benchmarks after curating context, with larger improvements in some cases. These are company observations rather than independent benchmark results, but they show the size of the gap Rohan believes a dedicated knowledge layer can address.Context also changes. Rohan references an example in which an agent reportedly moved from 95 percent accuracy to 65 percent within a month because business processes and priorities had changed. His argument is that companies need context stewards who can review updates, approve definitions, reject unauthorized changes, and maintain an audit trail. He compares this with code review, where machine-generated work still requires accountable human oversight.Security remains part of the design. Rohan says a context layer should explain what information exists while respecting role-based access controls. It should sit virtually above existing systems of record rather than copying sensitive data into a new repository. An agent might know that customer phone numbers exist while remaining unable to retrieve them. It can then tell the user that permission is required instead of inventing an answer or bypassing an established control.Before allowing an agent to act on company data, Rohan recommends two forms of testing. Business experts should create a small gold-standard set of tasks they know well, while the company can build a broader benchmark from repeated work already visible in its operational history. Together, these tests can show whether the agent understands the decisions people actually make rather than performing well on a generic demonstration.Does your organization have a shared, reviewable account of how it defines its most important metrics, rules, and working assumptions? Listen to the episode and share your thoughts.

What should a finance leader be able to prove before giving an AI agent permission to act?In this episode of AI at Work, I welcome Danny Fields, Chief Technology and Customer Operations Officer at Avalara, for a conversation about the distance between a convincing AI pilot and a finance workflow people can trust. We discuss the appeal of faster execution alongside the evidence, controls and business ownership needed when software begins making decisions or recommending actions.Avalara's research provides the starting point. Its global survey of 1,505 CFOs and senior finance leaders found that 92% felt career pressure to demonstrate AI agent ROI, while half reported limited measurable returns. Respondents came from businesses already deploying, piloting or evaluating agents, so these findings describe a particular group rather than every finance department. What interested me was the tension between pressure to demonstrate progress and the work required to make that progress defensible.Danny argues that copying an existing process into an AI workflow can preserve the problems already inside it. He describes why controlled pilots can struggle when production introduces incomplete records and unpredictable inputs. We consider what it means to redesign the workflow, measure accuracy over time and judge value through risk, audit preparation and business outcomes as well as hours saved.We also discuss accountability. Danny's view is that responsibility stays with the team that owns the business outcome, supported by documented roles across finance, compliance, IT and legal. Permission boundaries, recorded approvals and tested incident response become part of the operating model. A CFO may not need to understand the mathematics behind a language model, but should understand what information an agent uses, what it can do and when a person takes over.The conversation closes with a practical test for expanding autonomy: can the business demonstrate value, sustained accuracy, clear ownership, appropriate access and controls that reflect how agents actually work? These are Danny's recommendations, rather than a guarantee of compliance or financial performance. Please share your thoughts with me. What would your team need to see before trusting an AI agent with a consequential finance decision?

How much autonomy should an AI agent receive before it has proved that its work can be trusted?In this episode of AI at Work, I speak with Alexey Korotich, Chief Product Officer at Wrike and a former Dell software leader. Wrike supports around two million users, along with a growing number of agents, giving Alexey a close view of how enterprise teams are trying to move from AI experiments into daily work.Alexey challenges two popular measures of progress. A dashboard built in an hour may look impressive, but enterprise software creates lasting value through trust, context, and reliable delivery. The same problem appears in AI adoption metrics. Prompts fired, tokens burned, and tools deployed can show activity without revealing whether cycle time improved, margins changed, customers became happier, or a workflow produced a better result.The conversation then turns to the reason many AI projects stall. Alexey does not believe every process must be rebuilt from the beginning. He sees a more immediate problem in work that remains undocumented, unstructured, or divided across teams and systems. When employees repeatedly explain how work should be done, much of the time saved by AI disappears. Historical work, process artifacts, permissions, and accountability all give an agent the context it needs to produce a useful result.We also discuss how leaders can set boundaries for autonomous AI agents. Alexey compares agent onboarding with employee onboarding. Start with high-frequency, low-risk tasks such as meeting summaries, request triage, status updates, and risk flags. Review the output, build trust, and raise autonomy gradually. Higher-stakes decisions should remain with people, supported by identity controls, limited access, review channels, and a way to recover when an agent makes a mistake.For Wrike, shared context comes through what Alexey calls a work intelligence graph. It records how work moves, who participates, who owns each stage, and what each person or agent can access. That organizational memory can help agents learn from completed work and human feedback instead of treating every request as an isolated prompt.Should companies measure AI adoption by how often employees use it, or by what their workflows deliver? Listen to the episode and share your thoughts with me.

How do you give AI agents greater responsibility without losing control of the decisions they make or the costs they generate?On AI at Work, I’m joined by Jason Kurtz, Chief Executive Officer of Basware, to discuss governed autonomy, financial accountability, and what businesses need to consider as AI begins taking a more active role in their operations. For finance leaders, the challenge extends beyond choosing a tool. They need to understand what an agent is authorized to do, where human judgment remains necessary, and how to establish whether the investment is delivering useful results.Our conversation looks at the relationship between autonomy and oversight. Giving an AI system permission to complete a task creates questions about responsibility, approval boundaries, and intervention. When should an agent proceed independently, and when should a person review its work? These are business decisions that require input from the people responsible for the process, alongside the teams implementing the technology.We also discuss the CFO’s role in assessing AI value. Understanding the bill is only part of the job. Leaders need to connect spending with outcomes, whether that means improving a finance process, reducing repetitive work, or helping employees make better decisions. The conversation considers why cost visibility and evidence of value need to develop together.Token-based pricing adds another complication. As AI usage grows, businesses can face costs that vary with the amount and complexity of the work being performed. An experiment that appears affordable at a small scale may require a different commercial assessment when it becomes part of everyday operations. Jason and I discuss the questions leaders should ask about pricing, usage, and the relationship between greater adoption and greater expense.The people involved deserve equal attention. Working alongside AI agents may create new responsibilities for training, supervision, and governance. Employees need to understand how these systems fit into their work, what they can reasonably expect from them, and when to challenge an output or escalate a decision. Introducing digital workers brings organizational questions as well as technical ones.For anyone considering AI in finance or across the wider enterprise, this episode offers a way to think about adoption through accountability, commercial discipline, and business purpose. The opportunity is to give teams useful support while keeping responsibility for outcomes clear.Please share your thoughts. How is your organization deciding which responsibilities to give AI agents, and what evidence would convince you that their autonomy is delivering value?Learn more about Basware