Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Agents Are Reshaping Enterprise Workflows
Transcript
- Lucas: So last week I was talking to a friend who runs operations at a mid-size logistics company — about three thousand employees. And she told me they're piloting something called an AI agent. Not a chatbot. Not a copilot. An agent. Luna: What's the difference? I feel like we hear these terms thrown around interchangeably. Lucas: Right. And the difference matters. A copilot — like Microsoft's Copilot or GitHub Copilot — is essentially a productivity layer. It helps you write an email, summarize a meeting, generate code. But it's still reactive. You prompt it, it responds. An agent is different. An agent is given a goal — like 'resolve all tier-one IT support tickets from the past 24 hours' — and then it autonomously figures out how to do that. It calls APIs, it queries databases, it writes and executes code, it escalates only when it hits something it can't handle. Luna: So it's more like a virtual employee than a tool. Lucas: Exactly. And that's why it's a bigger deal than the previous wave of AI tools. The agent paradigm shifts the relationship between humans and software. Instead of you driving the tool, the tool drives itself toward an outcome. You're more of a manager or a supervisor. Luna: Okay, but is this actually happening in enterprises today, or is it still mostly hype? I feel like every vendor is slapping 'agent' on their product now. Lucas: It's a mix. There's real adoption, but also plenty of marketing. Let me give you two concrete examples. Salesforce — they've been rolling out something called Einstein Service Agent. It's designed to handle customer service inquiries autonomously. They claim it can resolve about eighty percent of routine cases without human involvement. And they're pricing it per conversation, not per seat, which is a fundamentally different model. Luna: That's interesting — per conversation pricing. That aligns incentives. If the agent doesn't actually resolve the issue, the company doesn't pay. Lucas: Right. And the other big one is Microsoft. They announced Copilot Studio lets you build custom agents. So a company could build an agent that connects to their internal systems — their CRM, their ERP, their HR database — and have it handle tasks like onboarding new hires or processing expense reports. Microsoft is positioning this as the next phase of their AI strategy. Luna: So what does this mean for the average knowledge worker? Is this a threat to jobs, or more of a shift in what work looks like? Lucas: I think it's a shift. And the data we've seen so far suggests the best outcomes come from augmentation, not replacement. A study from MIT last year looked at agents in customer support — they found that agents plus humans handled complex cases fourteen percent faster than humans alone, and customer satisfaction actually went up. The key is that the agent handles the routine, repetitive stuff, and the human focuses on the nuanced, high-judgment cases. Luna: So it's like the classic automation story — but with a twist that the automation is now intelligent enough to handle variability. Lucas: Exactly. And that variability is what makes this different from earlier RPA — robotic process automation. RPA was brittle. If a form field changed position, the bot broke. An agent using a large language model can adapt. It can read the context, understand that the field moved, and still navigate the interface. That's a huge leap. Luna: But with that power comes new risks, right? I mean, giving an AI agent API access to your CRM and HR systems — that's a massive trust and security question. Lucas: Absolutely. And that's why we're seeing a whole new category of software emerge — agent governance and monitoring tools. Companies like Credal and Mpathic are building platforms that let companies set boundaries: this agent can read customer data but not write to the billing system, this agent can only operate during business hours, this agent must get human approval for any action over five hundred dollars. It's basically identity and access management for AI agents. Luna: So the CIO's job is getting more complex. They're not just managing software licenses and security patches — they're managing autonomous digital workers. Lucas: Exactly. And that's a shift that's happening right now. A survey from Gartner this spring found that thirty-eight percent of organizations are already piloting or deploying AI agents in at least one business function. The most common use cases are customer service, IT operations, and sales follow-up. And the companies that are doing it well are the ones that treat it as a change management challenge, not just a tech implementation. Luna: What about the small and medium business side? Is this only for large enterprises with big IT teams? Lucas: It's trickier for SMBs, but not impossible. The good news is that many of the platforms — like Salesforce and Microsoft — are offering these agent capabilities as add-ons to existing subscriptions. So a small business using Microsoft 365 can start experimenting with agents in Copilot Studio without a huge upfront investment. But the governance piece is harder when you don't have a dedicated IT security person. You have to rely on the platform's guardrails. Luna: And what about the cultural side? If employees feel like they're being managed by an AI, that could create friction. Lucas: That's a real concern. There were some early experiments where companies deployed agents to monitor sales calls and give feedback — and the sales reps hated it. They felt like Big Brother was watching. So the successful implementations are the ones where the agent is positioned as a helper, not a judge. For example, a customer service agent that surfaces relevant knowledge articles to the human rep in real time — that's seen as helpful. An agent that flags a rep for not using the right script — that's seen as surveillance. Luna: So transparency and design matter a lot. The agent's purpose needs to be clear, and employees need to have some control over it. Lucas: Right. And that's where the best product teams are focusing right now. They're designing agents with explainability in mind — so when an agent takes an action, it can tell you why. And they're giving humans the ability to override or pause the agent at any point. That's critical for trust. Luna: Let me ask about the economics. Right now, AI agents are still relatively expensive to run because each call to a language model costs money. Does that limit adoption? Lucas: It does, but costs are dropping fast. OpenAI's GPT-4o is significantly cheaper per token than GPT-4 was a year ago. And there's a new generation of smaller, task-specific models — like Microsoft's Phi-3 — that are cheap enough to run for routine tasks. So the unit economics are improving. But companies still need to be smart about where they deploy agents. You don't want an agent that costs ten cents per interaction handling a task that previously cost a penny. Luna: Right, you need to find the high-leverage use cases first. Lucas: Exactly. And the companies that are seeing the best ROI are starting with narrow, high-volume, low-complexity tasks. Think password resets, order status inquiries, scheduling. Tasks that eat up a lot of human time but don't require deep judgment. Once the agent proves itself there, they gradually expand to more complex workflows. Luna: I've also been reading about something called multi-agent systems — where multiple agents collaborate on a task. Is that happening in practice? Lucas: It's early, but yes. For example, a company might have one agent that handles customer inquiries, another that checks inventory, and a third that processes refunds. They communicate with each other through shared data or APIs. It's essentially creating an automated workflow that used to require multiple people and handoffs. The challenge is orchestrating them so they don't get into loops or conflicts. That's why we're seeing startups like CrewAI and AutoGen emerge — they're building frameworks specifically for multi-agent collaboration. Luna: So we're moving from single-agent assistants to entire AI teams. That's a pretty radical shift. Lucas: It is. And I think it's going to change how we think about organizational structure. If you can have an AI agent that handles all procurement requests, you don't need a procurement department of ten people — you need one procurement specialist who manages and audits the agent. That's a different skill set. It's less about doing the work and more about supervising and improving the automated system. Luna: That sounds a lot like the shift that happened in manufacturing with industrial robots. Factory workers became robot operators and maintenance techs. Lucas: Exactly. And history suggests that while some jobs are eliminated, new ones emerge. We're already seeing job postings for 'AI agent managers' and 'workflow automation architects' — roles that didn't exist two years ago. The question is whether the pace of change will be faster than the labor market can adapt. Luna: And that's where policy and education come in. If we're going to have a workforce that manages AI agents, we need training programs that prepare people for that. Lucas: Totally. And I think that's something companies and governments need to start thinking about now, not later. Because the technology is moving fast. The agent capabilities we have today are primitive compared to what's coming in the next two to three years. We're going to see agents that can negotiate contracts, agents that can write code across multiple repositories, agents that can conduct entire customer interactions from start to finish. Luna: Speaking of the future — what's one thing you're watching closely in this space over the next twelve months? Lucas: I'm watching for the emergence of industry-specific agents. Right now most agents are general purpose — they handle customer service, IT support, sales. But I think we'll see agents built specifically for healthcare, for legal, for logistics. An agent that understands HIPAA compliance, that knows the legal standards for document review, that can optimize a supply chain route in real time. That's where the real value will be unlocked. Luna: And that requires deep domain expertise in the training data and the architecture. It's not a one-size-fits-all. Lucas: Right. And that's why the companies that will win in this space are the ones that combine AI talent with deep industry knowledge. It's not enough to have great engineers — you need people who understand the workflows and the regulations of a specific sector. Luna: It's a fascinating moment. And honestly, conversations like this make me realize how much I rely on podcasts and newsletters to keep up. It's a lot to follow. Lucas: I hear you. And you know, that's exactly why listener support matters to us. Shows like this — they take time to research and produce. And they're ad-free, which means they rely on people who find them valuable chipping in a little. If you've gotten something out of this episode, buying us a coffee is a great way to keep it going. That's buy me a coffee dot com slash fexingo. Even a couple of dollars genuinely makes a difference. Luna: Yeah, it really does. It's a small way to support independent content you value. Lucas: Anyway — back to agents. I think the other big trend to watch is the integration of agents into existing enterprise software. Microsoft, Salesforce, SAP — they're all embedding agent capabilities directly into their platforms. So within the next year, using an agent could be as normal as using a dropdown menu. Luna: And that's going to lower the barrier to entry dramatically. You won't need a team of data scientists to deploy an agent. It'll be a checkbox in a settings panel. Lucas: Exactly. And that's when the real transformation happens — when it becomes invisible, just part of how work gets done. The question is whether organizations are ready for that shift culturally and operationally. Luna: I guess that's the million-dollar question. Thanks Lucas. This was a great unpacking. Lucas: My pleasure. And for everyone listening, I'd love to hear how your organization is thinking about AI agents. Drop us a note — we read every one. Until next time.