How AI Agents Work in Field Service

‍I sat down with Andrew Wolf, Co-founder and CPO of Zinier, who's deployed AI agents in production field service environments. While everyone talks about AI transformation, Andrew reveals the gap between marketing promises and operational reality. We discussed what actually works and the technical considerations that matter.

How AI Agents Work in Real Field Service Operations

Q: How do Zinier's AI agents actually work day-to-day? What would an operations or field team see?

Andrew: When I talk about agents, I think of them across two dimensions. One is capability and the other is domain. Another way of thinking about agents is you can think of them as specialized employees who each have a capability - a type of job or set of tools that agent is trained to perform.

There are lots of different types of tools or capabilities. One is data queries. This is pretty prevalent on the market today, but an agent can be trained to answer questions, summarize your data, perform analytics on the data in your solution. This covers things like "what technician had the highest utilization last month in the Northeast region?" The way it works is you type this into the chat interface, and LLMs that have access to the data can surface this information quite quickly. This may not even be in the field service product itself - it might be in your PowerBI export or something like that.

You can also have an agent that's been trained to answer questions on your documentation that's been uploaded. This typically is like a RAG architecture to be able to answer specific questions around documents and do summaries and other things like that.

The third capability, which is still under development for us but is exciting, is giving agents the power to trigger actions and workflows. I mention workflows a lot in the Zinier ecosystem, but workflows are really what powers everything under the hood. Giving agents the ability to trigger actions is the next step towards agentic automation.

AI agents work as specialized employees with specific capabilities: data queries, documentation access, and workflow triggers - not just chatbots.

The Reality of AI Accuracy in Production

Q: Let's say my operations team was worried about AI making mistakes. What kind of accuracy can we expect?

Andrew: Look, there's no sugarcoating this - with a data type agent, we've been getting them up to well past 95% accuracy. And that's without having gone deep into some of the new stuff that's now available. We're continuing to make these agents better.

With documents, it's a little different because it's stored in a RAG database and then you just got to make sure all the chunking happens accurately and then you can have a link to verify it. But it's getting better.

And I would also just say you kind of have to look to the future as well. Things are improving - you might reach a certain saturation point in terms of how current models are operating on data, but it really gets you to where you need to go. Then improvements continue to happen in the space. We continue to incorporate those improvements, and if you look a year or two into the future, I actually don't expect there to be that much hallucination risk. I think that ends up getting solved in general. So if you think about the future, this feels to me more like a nowish problem that is probably going to go away in a year and a half or two years.

We know we're working with enterprise customers. We know this information is critical. You guys are doing the hard work. We can't afford to give folks hallucinations. We actually go through a refinement process - it's not like just turn it on and go. Every situation is a little unique, and we can keep doing testing and refinement prior to it being deployed to get the accuracy levels up to what we need.

95% accuracy is achievable, but it requires refinement processes that most vendors don't mention. The technology improves rapidly, but implementation matters more than the models.

Balancing Automation with Human Control

Q: If one of our customers needed automation but wanted to keep control over scheduling decisions, is that possible?

Andrew: The recommendation center is actually quite unique to us. We don't know of any other product on the market that has this kind of recommendation center, and it's got a lot of powerful ways to help drive automation and help you manage by exception.

What you can do with recommendations is automate any workflow, any decisions, by creating business logic within a workflow that can automate a decision and trigger a workflow. These are triggered by events, and essentially what happens is an event happens and then it triggers this recommendation. You can accept or reject it, and when you accept it, it triggers a workflow in our system.

For example, if you have a certain type of critical priority task that comes in, you can have a recommendation that says "hey, this critical task has come in, do you want to automatically assign this task to the nearest technician who hasn't been dispatched yet?" But you can get very granular and specific about this. It could be for specific customers where you want to take a specific type of action or do a specific type of scheduling.

Think about all of the different business decisions that an experienced dispatcher might make over the course of a day. You can really codify that with different recommendations to continue to automate. You might have a technician that falls sick or goes on leave, and perhaps you need to automatically reassign all of their assigned tasks to other folks. You can do that with recommendations.

Rule-based recommendations let you codify dispatcher logic while maintaining human oversight. Events trigger suggestions that you can accept or reject.

AI in Remote Field Operations

Q: Some of our customers' field teams work in remote areas with spotty connectivity. Will AI features still work?

Andrew: This is definitely a challenge we see, especially with utilities and telecommunications companies operating in remote areas. Our mobile app is built as a native iOS/Android app, not HTML5, which gives us much better offline capabilities than browser-based solutions.

When technicians are working offline, they can complete jobs, fill in forms, collect signatures, take photos - all of that gets synced when they reconnect. Now, for AI features specifically, some capabilities like the chatbot interface do require connectivity. But we've designed the system so that the most critical operational functions work offline.

We've actually done implementations where regional managers who are mostly sitting in their trucks behind their iPads need to do scheduling. The key insight here is that if it takes longer than a text message to dispatch somebody, they're not going to do it. So we need something that's faster than that. We can give them access on an iPad to do scheduling where it's just one click, two clicks, and their entire ten technicians or sixty technicians - those recommendations are sorted out.

There's also functionality through notifications where you can take action on mobile. You don't have a Gantt chart view on a small screen, but you can handle many scheduling decisions through the mobile interface or use the web interface on an iPad if needed.

Native mobile apps handle offline operations better than browser-based solutions. Critical functions work without connectivity.

Integration Reality with Enterprise Systems

Q: What do our customers need to think about when connecting their existing systems to Zinier?

Andrew: This is where our platform architecture really shines. We have an event-driven, workflow-based architecture that was built from the ground up to be able to connect to other systems and other technologies. This is all enabled through workflows.

We often get questions like "have you integrated with SAP?" or "can you do it?" Because of the approach we've taken with workflows being used to do integrations, the answer is always yes - we can integrate to any other system and also leverage new technologies such as different AI models, AR/VR products, to help keep you at the forefront of technology.

Studio Z makes integrations with our clients' external systems fast and flexible. Integrations are built using our low-code Workflow Builder, which allows anyone with basic knowledge of JSON and JavaScript to define and maintain integrations. Payload transformations and API changes are all handled through workflows, allowing us to tackle even complex and custom integrations. We also maintain a number of out-of-the-box integrations with ERPs, EAMs, CRMs, and other common external systems.

The idea here is really to help future-proof this decision and know that you can leverage new technology and products as they emerge in the market. When new AI models come out, when new capabilities become available, you can incorporate them into your solution without having to rip and replace your entire system.

Workflow-based architecture handles integrations through low-code tools, not custom development. Future AI models can be incorporated without system replacement.

Want to see Field Service AI Agents in action?

As Andrew notes, the accuracy and capabilities will improve rapidly, but implementation methodology matters more than model sophistication. The question isn't whether AI will transform field service operations, but whether your platform can adapt to leverage these capabilities as they mature.

Want to see how Zinier's AI agents work? Schedule a demo to explore Zinier's approach to production-ready AI implementation.

‍