AppZen’s Anant Kale Follow-up on AI in the Back Office: A Serious Insights Conversation
Fraud detection is a revealing area for measuring how quickly financial AI is maturing. In our December 2025 conversation, AppZen co-founder and CEO Anant Kale described an AI-versus-AI contest in which synthetic receipts and altered documents are pitted against systems trained to identify anomalies in both documents and behavior. This follow-up begins with AppZen’s reported jump from 0% to 70% of detected fakes over 14 months and asks whether the change reflects more fraud, better detection, or both.
Kale’s answer points to product maturity rather than a single technical breakthrough. The larger shift, however, is from finding suspicious transactions to acting on them. AppZen now positions agents between its detection models and human auditors, where they apply context, follow standard operating procedures, request missing information, approve or reject work, and escalate uncertainty.
That makes governance part of the workflow. Kale describes simulating changes to standard operating procedures before deployment, monitoring human-handled exceptions for recurring patterns, publishing precision and recall targets, sampling production decisions, and directing uncertain cases to people. Those practices make the boundaries of delegated authority visible.
The workforce consequences are equally explicit. Kale says finance leaders are evaluating agents as substitutes for portions of outsourced and shared-services work, not merely as assistants to existing teams. In that setting, change management requires executive direction, measurable outcomes, new roles for people who manage agents, and domain experts willing to learn to redesign the work.
Top takeaways from the interview:
- Finance AI is moving from detection to resolution. AppZen’s earlier systems narrowed the auditor’s queue; Kale says agents can now apply context, follow a customer’s standard operating procedure, and take or recommend the next action. AppZen claims this can remove as much as 95 percent of the work an auditor would otherwise perform on a flagged transaction.
- Governance has to operate at the speed of the agents. Kale describes testing policy changes against historical work, comparing agent decisions with human results, analyzing recurring exceptions, publishing precision and recall targets, sampling production decisions, and escalating uncertainty. Trust comes from visible limits and repeatable assurance, not a claim of perfect accuracy.
- Autonomous finance is becoming a workforce strategy. Kale says customers are using agents to reduce business-process-outsourcing and shared-services headcount. Successful adoption depends on top-down goals, serious change management, and experienced, curious employees who can translate domain knowledge into governed digital work.
The Anant Kale interview
In our last interview, you said repeating fraudulent patterns were a major signal. More recently, AppZen said it went from zero to detecting 70 percent of fakes in 14 months. Was that increase more about fraud or more about detection?
It was about finding new ways to identify things that are not real. It was not one unlock that enabled all of it. You can create fake receipts with different models. You can create or modify receipts with these tools and PDF editors; you can work with images; and you can use templates from fake-receipt websites. There are multiple ways of doing it, and there is no single solution where you can say, “I found this.”
We are learning from the data and from patterns we previously missed. We have found other techniques that make this more visible. Sometimes it also requires behavioral analysis: something could be wrong, so I go to the next level of detail to understand why the pattern looks this way. I would call it product maturity, with technology allowing us to detect more and more of this activity.
Are those techniques also producing false positives or false negatives?
False positives absolutely happen. The false-positive rate when searching for fraudulent receipts is much higher than what we allow for transactional AI models that process an invoice, audit something, or code something. Those models require very high precision.
For fraud detection, what matters more is recall: can we find something that is worth looking at, even if it is not always wrong? If you optimize only for precision, you are going to miss many things—the false negatives. There is no point in having a leaky funnel.
AppZen cites 87 percent autonomous processing, a 90 percent reduction in manual audits, and operating-cost reductions of up to 50 percent. Can you walk through what a good deployment behind those numbers looks like?
What has evolved considerably is the use of models and agents to take on the next level of human work. Think about our product in accounts payable or expenses. It was initially created to find a problem, such as a policy violation in an expense report, highlight it, and let a human auditor review it and decide what to do next.
That alone creates a lot of value because you reduce the review population from 100 percent to closer to 10 percent. But after an item reaches a person, the auditor still has to evaluate what the AI found. Suppose it identifies that someone bought a number of gift cards. The auditor looks at what they were for, sees that the explanation says it was for a customer trip, and notices that the employee works in marketing. The auditor combines context about the person and the reason for the purchase, then decides whether it is acceptable, a policy violation, or something that requires more information.
That level of analysis was not possible with our earlier technology. Now, once our model finds something wrong, we do not immediately hand it to a human. We hand it to an agent. The agent analyzes the result and applies the context I just described. First, it makes sure the model got it right and that the alert is not a false positive. Then it determines whether to reject the expense, ask for more information — for example, whether it was pre-approved by a manager — or let it go. Those are decisions and actions a human would have taken that we can now hand to an AI agent.
That has been the most revolutionary change for our customers during the last six to eight months: agents can evaluate how a decision should be made, consider the options, and remove a large layer of work. We say it can take away 95 percent of the work the human auditor would have done.
That is the unlock in AI Agent Studio, where customers can build their own agents. The issue we all face, and which is especially important in finance, is knowing how a decision was made. There has to be governance. Our approach to building an agent is similar to building a standard operating procedure: this is what I look at, this is the function I perform, and the AI should not look outside those defined boundaries.
The governance around how the decision is made is visible in the SOP. A human can confirm that this is how the work is supposed to be done; the AI agent follows it, and the decision is auditable because the reasoning is visible. That has been a big unlock for us. We now deploy these agents for customers, and they also come standard in our product.
That sounds much more deterministic than generative. Companies change policies and practices, however, and the operating procedure has to change with them. How do you manage change on the AI side without creating drift or allowing a generative system to misinterpret the new policy?
At the end of the day, the process is logical and deterministic. Arriving at that logic is both the fun part and the difficult part. Nothing is static. If I change a policy, I do not automatically know what drift it will create or how decisions will change.
This technology allows us to express what the policy change means for the SOP, then simulate the revised SOP. We can show which decisions and actions would have changed – the good, the bad, and the ugly – before the company implements it. That is super powerful, and it is what people are now using when they deploy agents.
You do not simply create an agent, create an SOP, and deploy it. You simulate it, test it, compare it with human results, and examine the consequences. Things will go faster; that is a given. But what else is happening? What is the impact? You want to see that before implementing the agent or the new SOP.
What happens when an agent keeps handing the same kind of exception to people, and they manage each case without updating the process? Does the system monitor that pattern and recommend a better automation or a change to the SOP?
We have thought about that and built a feature we call continuous transformation. It uses another agent to analyze what is happening and look for patterns in execution. It reviews transactions handed to humans, observes the decisions people make, and starts giving recommendations.
It might say, “When this is handed to people, this is what they do, so I should add that decision logic to the existing SOP.” It might recommend looking at other factors that people use implicitly when they make a decision. It can show how to change an operating procedure or identify an opportunity to add an SOP or an agent that does not exist today.
We produce a report that shows customers the work people perform today. It identifies SOPs already defined as good practices that the company is not using and estimates the work those SOPs could remove. It also identifies SOPs that exist but need to be updated because they do not reflect how people actually make decisions. Maybe the actual practice is wrong, but at least the organization can see how decisions are being made and has an opportunity to change it.
There will still be areas where we do not have enough data, evidence, or patterns to understand the work. We call the report the Hybrid Workforce Blueprint. If a company wants to move toward a hybrid workforce, more decisions have to move from a human worker to a digital worker — an agent. The blueprint shows where agents already perform work but the SOP needs updating, where work is not handled by agents but could be, and where we simply do not know enough yet. It creates a continuous-transformation process for the customer.
I encourage people to say “good practice” rather than “best practice” because “best” suggests it cannot improve. As organizations optimize with AI, do they risk deciding they have found the optimal process, then stop looking and stop innovating? What happens when a new technology or method disrupts a process the models have never seen?
Expecting AI to have answers for everything is unrealistic; we are far from that. It works for processes that are well-defined, have gates and governance, and use deterministic rules. Many processes fall into that category, from legal review of a contract to finance and accounting. Much of that work can be distilled into rule-based or action-based decisions, which makes it well suited to optimization by AI and execution by agents today.
But when you are thinking about a genuinely new way of doing something – even when we are deciding what to do in our own product – that is not the realm of AI, at least not the AI I know. You can throw everything at AI and ask it for a solution, and it will produce something, but that does not mean the answer goes anywhere.
To think creatively about what to apply, people need the intellect to understand what is causing the problem and deep knowledge of the domain. AI does not replace that.
Frontier-model vendors often sell more magic than they can deliver. AppZen now has a Trust and Assurance Program, with some customer-sensitive material covered by an NDA. Do you see external certification, industry benchmarks, or other independent assurance becoming necessary beyond AppZen’s proprietary program?
The program is for our finance customers, and some of its metrics are customer-sensitive, which is why they are covered by an NDA. The idea is that finance teams make decisions that are critical to operating a company and, for public companies, meeting their obligations. Those decisions need controls. Finance has SOX controls and other general controls designed for rule-based decision-making. If one person creates something, another approves it, and another executes it; the duties are separated.
When those decisions are automated by AI agents, customers need a reason to trust the system. We cannot just say, “Trust us; we know we are right.” For each finance workflow we automate – auditing expenses, coding or entering accounts-payable invoices, or paying vendors – we identify key decisions that customers rely on us to make.
For those decisions, we publish precision and recall. We might say that we are 95 percent accurate and cover 94 percent of cases. That means we know we will be wrong in 5 percent of cases, and we know we will miss some cases. We publish those figures so customers understand the commitment; we do not claim 100 percent accuracy. We also believe those results can be better than human performance when the same test is run on people.
Training on our data and reporting a result is not enough. The trust program periodically revisits the criteria. Each month it samples data across customers and checks whether the system meets the benchmark. Human reviewers inspect decisions and reasoning manually. If something is wrong but remains within the boundaries described in our documentation, we know the measured error remains within the commitment. We also run the assurance tests whenever we change or tune a model or introduce a new model.
That becomes quality assurance for AI. Very few organizations are setting benchmarks or service levels for AI, but we have to because finance teams are delegating critical financial decisions from people to agents. That is what the program is about.
In the AP Inbox Service Center, an agent may communicate directly with vendors. If an agent gets something wrong, where are the guardrails, stopping points, and remediation? Does every communication go through a person?
Every agent has guardrails. The explicit guardrails define when it wakes up and acts, the situations in which it works, the decisions it makes, the information it examines, and the actions it can take.
One guardrail is always present: if the agent is not confident about its decision, it hands the work to a person and includes the reasoning that led to the escalation. By default, we do not force an action when the system is uncertain.
If an agent reads an email from a vendor but is not clear about what the vendor is asking – perhaps it detects a dispute but the SOP does not tell it how to answer – it analyzes the request, presents its interpretation, and tells a person to take over. That guardrail is important because this is finance. The same applies when processing an invoice. If the tax rate is 18 percent but the calculated amount does not reconcile, the agent cannot make up an answer. It says that the numbers do not add up, shows what it extracted, and asks a person to take over.
Our number-one rule is to err on the side of human action rather than fabricate an answer and claim it is right. We might tell a customer that the system handled 90 of every 100 emails autonomously. For the remaining 10, it may think it knows the answer, but it will not send it. It creates a draft, shows its work, and requires a person to act.
Companies often say automation will free people to do something else. What are you actually seeing as customers apply agents to finance outsourcing and shared-services work?
“We will always find something else for people to do” was the old story. I do not think that is the case anymore.
Our customers generally have two kinds of outsourcing. One is work sent directly to a business-process-outsourcing company, such as Accenture or Tata Consultancy Services. These arrangements can involve hundreds of people. Customers see a lot of churn, they have to wake up in the middle of the night to make things work, they travel to the centers, and there is a large total cost.
Customers are telling us that if an agent can do the work better at a similar cost, they are ready to use it. Massive companies come to us and say, “We have 200 people in Eastern Europe or 300 people in Bangalore doing this work. If I can remove 60 percent of those roles, I do not need them. I might even hire some people locally and make the operation easier.”
How much work can be taken over by a digital worker is central to what our agents do. In earlier years, leaders would not answer that question directly. Today, every board and CEO is asking what the company is doing with AI and whether it is producing that kind of result. BPO providers that promise transformation have little incentive to remove people because their business model depends on those people.
The second type is a company’s own captive or global capability center. Those people are employees, not a third party that can be wound down on a contract timeline. Companies see that as more of a challenge: what do we do with these people? But I am seeing it happen there as well, although at a slower pace than in BPO arrangements.
Not every company adopts this technology at the same speed. What obstacles separate the companies that are moving successfully from those that are not?
The key obstacle is change management. Every software project has required change management for decades, but this is different because it requires a change in people’s attitudes. The company is replacing work that people perform, and that is sensitive. Operators have the knowledge and are invested in the current process, so there is a natural pushback.
If the company does not explain why it is doing this and why it matters, the project will not work. The direction has to come from the top: this is how the company intends to operate, this is the change it will implement, and this is the result it expects. If reducing the headcount in a team by 60 percent is the outcome, leadership has to say so. Without a top-level initiative and a clear definition of success, these projects do not work.
Sometimes the goal exists but is not backed by enough change management. There have to be champions and winners. A manager may move from leading a team of 20 people to managing 100 agents. That person needs a new role, training, and a reason to lead the change. If the company does not create and train those champions, everyone will push back, and the project will fail.
We advise customers to invest in change management alongside the technology. We can help them understand what resistance they will encounter, but if they do not invest in that work, the implementation will not succeed.
In earlier research, I described the “horizontal employee”: someone empowered by AI who can move across functions rather than remain tied to one job. An accounts-payable expert, for example, might learn agentic tools and join a team addressing accounts receivable. Are you seeing that behavior among your customers?
I am seeing forms of that transformation. People who are innovating recognize that, to succeed in their future jobs and careers, they need to lean into AI, understand the tools, and become familiar with how to use them. They have already figured out how to work with ChatGPT, create presentations, and make themselves productive. They often become effective transformation agents on their teams and can work horizontally because other teams can use them as well.
Many are younger, but that is not the essential attribute. The biggest attribute is curiosity: they want to learn something new, are open-minded, and are willing to roll up their sleeves and do the work.
Domain knowledge and experience still count for a lot in the areas every company is trying to automate. If you want to go beyond superficial automation, such as creating a slide deck or doing simple analytics, you need to understand the subject deeply. The best transformation agents are people rooted in the domain who also lean into AI and adopt it.
Workday, SAP, Oracle, and other platform companies are building finance agents. Where does that leave a specialist such as AppZen, and what creates defensible differentiation when the large vendors have access to the same models?
If it were that simple, the frontier models would already have done it. To exceed human decision-making on cost, speed, and depth of workflow, you have to go very deep into the problem. A frontier model might do 50 or 60 percent of the work, but the advantage does not come simply from access to data or a model. Everybody has access to similar models and increasingly to the same agent frameworks.
Differentiation comes from choosing a task and going deep enough to automate every aspect of it. That is a focus problem. The large companies have brilliant engineers; the question is whether they will focus on a specific workflow or on the biggest market pulling at them.
Most companies that say they have agents really have fancy copilots or chatbots. In the early days, you asked a chatbot a question and got an answer. Now the chatbot is smarter and can suggest what you could do and how you could do it. For them, that is an agent.
For us, an agent works in the background and does the work without requiring the user to watch it. That level of autonomy happens only when you go very deep into solving a specific problem. I am less worried about the large platforms on that dimension.
About Anant Kale
CEO, AppZen
Anant Kale is the co-founder and CEO of AppZen, the leader in AI-driven finance automation. Under his leadership, AppZen has pioneered the use of artificial intelligence to transform global finance operations for the world’s largest enterprises. With a vision to bring autonomous, intelligent systems to corporate finance, Anant has scaled AppZen into a trusted partner for Fortune 500 companies, helping them improve compliance, reduce spend, and unlock efficiency. Before founding AppZen, he held leadership roles in Fujitsu America. Anant earned a Bachelor of Science in Finance and Engineering and an MBA from Mumbai University.
LinkedIn: https://www.linkedin.com/in/anantkale/
For more serious insights on AI, click here.
For more serious insights on learning, click here.
Did you enjoy the Sean Iannuzzi interview? If so, like, share or comment. Thank you!
Leave a Reply