Source: Unite.AI
Andrej Jonovic is the CEO of XBP Global, a multinational technology and services company focused on mission-critical enterprise workflows. XBP Global is applying agentic AI across regulated environments including healthcare, public sector and financial services, where precision, security, data sovereignty and auditability are essential. In this interview, Jonovic discusses adaptive process orchestration, open-weight models, intelligent document processing, human adjudication and what it takes to turn AI from a technical capability into measurable operating results.
You began your career as a transactional lawyer before moving into mergers and acquisitions, corporate transformation, and ultimately leading XBP Global. How has that background shaped the way you evaluate the governance, commercial risk, and organisational impact of deploying artificial intelligence across mission-critical workflows?
I tend to look at AI through the operating model first. Coming from transactions and transformation, you get used to asking whether the structure actually holds once you move past the presentation and start executing.
That becomes very important with AI because a model can perform extremely well in isolation and still create a bad operating result. In healthcare, public sector and banking, processing precision, regulatory compliance and auditability are non-negotiable. You have to understand where the data lives, what the system is allowed to do, how decisions are controlled and what the cost base looks like when you run the workflow at volume.
Governance has to be designed into that architecture. I want to know where the model is running, what information crosses the enterprise boundary, what gets recorded in the audit trail and what happens when the workflow reaches a decision that requires a different level of control.
That is probably where my background has the biggest impact. I am generally less interested in what the technology can do in a controlled demonstration. I want to know whether you can put it inside a mission-critical process and rely on it.
XBP Global describes agentic AI as an evolution beyond conventional rules-based automation and robotic process automation. What changes technically when a workflow moves from following predefined instructions to using AI agents capable of inference, decision-making, and coordinating multiple tasks?
The architecture becomes much more adaptive.
Traditional workflow automation is deterministic. You have rules, you know what happens at the next step and you hardwire more logic as the process becomes more complex. That works very well for predictable work. The difficulty comes when the number of rules becomes enormous or the process changes faster than the system can be rebuilt.
What we are doing now is adaptive process orchestration. You can retain deterministic rules engines where precision and compliance require them, then introduce intelligent models into the parts of the workflow where inference adds value. Agents can coordinate tasks, access information and select the next action without somebody having to hard-code every possible route through the process.
The other change is that the underlying model becomes more replaceable. We are seeing enterprise-grade open-weight models improve very quickly. If the pipeline is designed properly, you can integrate a stronger model as the ecosystem develops without rebuilding the entire operating process around it.
So I think the important technical asset increasingly becomes the workflow itself: the orchestration, domain knowledge, controls and data around the model.
Many companies are experimenting with AI but struggle to move from pilots into production. What are the most common architectural or organisational barriers preventing enterprises from deploying agentic workflows at scale?
Clients are moving past the experimentation stage. They have seen the demos and heard very aggressive ROI promises. At some point, somebody has to put the technology inside a real process and make it work.
That is where the architecture gets difficult. A production workflow has to connect with the environment the company already has. Sensitive information may need to remain in a private cloud or on-premise. Existing systems still have to be accessed. Regulatory requirements have to be met. The economics also have to survive at enterprise volume.
We recently won work around the maintenance life cycle of a global airline fleet. Every maintenance item will be vectorized and stored in a private vector database inside the client’s private cloud. The agentic infrastructure manages access to that information and supports reproduction of the records when aircraft go off-lease or are sold.
That is a very different problem from proving that a model can answer a question about an aircraft record. You are dealing with an entire operating environment, with data controls and a process that has to work repeatedly.
Organizationally, the gap is usually ownership. Somebody has to understand the process deeply enough to define what production actually means. Without that operational discipline, the pilot can stay impressive for a very long time without becoming useful.
XBP Global combines automation with human-in-the-loop oversight, particularly for high-consequence decisions. How do you determine which decisions an AI system can execute autonomously and where human judgement must remain part of the process?
We design the human role around adjudication.
In a high-consequence workflow, you usually have a large amount of work that can move through the pipeline with a high degree of confidence. Then you have edge cases where the information is incomplete, the circumstances are unusual or the consequence of getting the answer wrong is too high.
When the pipeline reaches one of those cases, it routes the transaction to a subject matter expert. That person resolves it and the judgment can then be fed back into the system so the model continues to improve.
I think that feedback loop is important. Human oversight has more value when it improves the workflow over time. The expert is bringing domain knowledge into the cases where the model still needs it.
The autonomy threshold also depends heavily on the environment. Healthcare adjudication, public-sector records and regulated financial processes have different requirements from a low-consequence internal task. The acceptable confidence level has to reflect what the decision actually does once the system acts on it.
Your clients operate in regulated industries and the public sector, where data sovereignty, security, and auditability are essential. How can organisations deploy large language models within protected environments without compromising the capabilities that make these systems useful?
Open-weight models are changing the answer to that question quite significantly.
For a long time, there was an assumption that getting the best model performance meant sending information into somebody else’s environment. That creates an obvious problem when the data has to stay inside the enterprise boundary.
We can now take enterprise-grade open-weight models and integrate them into deployed agentic pipelines while keeping the data inside the client’s environment. That can mean private-cloud or on-premise infrastructure, depending on the requirements. You still get the benefit of advances in the broader model ecosystem while maintaining the controls the organization needs around data sovereignty.
The architecture also has to be able to change. There is no reason to assume that today’s model will remain the best model for a specific task. Our training protocols are designed so we can integrate new models into an existing pipeline quickly.
For regulated clients, that flexibility is extremely valuable because the workflow can continue improving while the data boundary remains intact. That is also how you begin addressing requirements such as the EU AI Act without treating compliance as something that gets added after the technology has already been deployed.
A significant amount of enterprise information remains trapped in documents, images, legacy systems, and physical records. How is intelligent document processing evolving, and what role does it play in turning fragmented data into information that AI agents can reliably use?
I think intelligent document processing is becoming part of the infrastructure for agentic workflows.
Historically, a lot of document processing was template driven. You knew what the document looked like, you knew which field you wanted and the system extracted it. Enterprise information is rarely that clean. Healthcare alone can involve EOBs, complex correspondence, claims information and records that arrive in different formats.
The newer approach is much more context aware. You can take unstructured information, understand what is relevant and convert it into structured outputs that downstream systems can actually use. That is important because an agent cannot reliably orchestrate work if the information underneath it is incomplete or incorrectly interpreted.
Scale also changes the problem. We operate environments that process enormous volumes of images, claims, payment transactions and other documents. At that level, a small difference in extraction quality becomes a very large operating issue.
So IDP increasingly sits at the front of the intelligent workflow. It turns fragmented information into usable data, gives the next part of the process something dependable to act on and creates the foundation for more automation further downstream.
XBP Global has discussed migrating legacy, rules-based processes into agentic AI pipelines and achieving higher levels of first-pass resolution. What must an AI system understand about an organisation’s processes, exceptions, and institutional knowledge before it can deliver dependable outcomes?
A lot of institutional knowledge never made it into a clean process document.
It sits inside old systems, rule sets and the judgment of people who have spent years operating the process. They know which information matters, which rule applies under a particular circumstance and what happens downstream when the wrong decision gets made.
When you migrate that work, you have to preserve that operating knowledge. In our world, that often means combining deterministic logic with AI rather than assuming every rule should disappear immediately. There are parts of a regulated workflow where a rules engine is extremely useful because the requirement is precise. There are other parts where an intelligent model can handle variation much more effectively.
Domain expertise tells you where each belongs.
That is also why first-pass resolution cannot be viewed as a model score by itself. The workflow has to produce an outcome that survives the next step in the process. If it creates another escalation somewhere downstream, you have moved the problem rather than resolved it.
The system has to understand the process well enough to know what a successful transaction looks like from beginning to end.
XBP Global is applying automation not only to client operations but also to its own transition toward an AI-first operating model. What have you learned from using the company as an internal proving ground, and how do you balance productivity improvements with workforce development and change management?
We have been pretty aggressive about applying this internally because I think you learn very quickly when your own economics are attached to the result.
Our AI-first automation efforts are now expected to generate $65 million to $75 million in annualized operational efficiencies, with approximately $35 million of benefit expected during 2026. Revenue per employee reached roughly $89,000 in the second quarter, up from about $82,000 the previous quarter, and we expect that to approach $100,000 by the end of the year.
Those are useful measures because they tell us whether the operating model is actually becoming more productive.
The workforce implications are real. We are estimating roughly a 20% reduction in global headcount by the end of 2026 as we move toward a higher-productivity operating model. Change management has to acknowledge that productivity changes staffing requirements.
At the same time, the people who remain inside these workflows have to become more capable. We need domain experts, architects and operators who can structure work, evaluate outputs and take responsibility for higher-value decisions.
Running that transformation inside our own company has made the conversation much more practical. You see very quickly which automation creates durable efficiency and which activity is just somebody using AI more often.
Enterprise AI projects are often promoted through demonstrations rather than measurable business results. Which metrics should leaders use to evaluate an intelligent workflow, and how should they account for accuracy, exception rates, processing time, operational cost, and risk reduction?
I would measure the workflow at two levels.
At the operating level, I want to know the accuracy, first-pass resolution, processing time, exception volume and cost to process the work. In a regulated environment, I also care about whether the transaction is auditable and whether the controls hold up consistently at volume.
Then I want to see whether those improvements reach the economics of the business.
At XBP, our adjusted gross margin reached 24.9% in the second quarter, our highest level to date, while revenue per employee increased to approximately $89,000. Our annualized operational efficiency target also moved up to $65 million to $75 million as we identified additional automation opportunities.
Those measures tell you something different from a benchmark score. They show whether the technology is changing the cost structure of the operation.
A workflow can have excellent model accuracy and still be a bad deployment if it is too expensive to run or creates too much manual rework. I think leaders need to follow the result all the way through. The technology has to produce an operating benefit that shows up somewhere you can actually measure.
Looking ahead, do you expect agentic AI to remain a layer added to existing enterprise systems, or will it become the primary orchestration layer through which people, software, data, and business processes interact?
I think more of the workflow will become agentic.
The enterprise will still have systems, data repositories and applications that perform specific functions. What changes is how work moves between them. An agentic pipeline can retrieve information, coordinate tools, trigger actions and determine when a transaction needs another type of intervention.
You can already see that in the airline example. The maintenance records sit in a private vector database inside the client’s environment, while the agentic layer manages how that information is accessed and used through the maintenance life cycle.
I expect that type of orchestration to become much more common.
The underlying models will continue to change. Open-weight models are improving rapidly, and there will be another generation after the ones we are using today. That makes the orchestration layer increasingly important because the enterprise needs something durable around those models.
If you build the pipeline around the process, domain knowledge, security requirements and enterprise controls, you can continue taking advantage of better models as they arrive. That is where I think a lot of the long-term value will sit.
Thank you for the great interview, readers who wish to learn more should visit XBP Global.