Anant Kale, CEO and Co-Founder of AppZen on AI and Fraud in the Back Office: A Serious Insights Interview
Generative AI has shifted from novelty to infrastructure, and as it evolves, it is rewriting the risk profile of basic finance workflows. Expense reports and invoices once felt like back-office hygiene; now they sit on the front line of an AI-versus-AI contest where synthetic receipts, invented vendors, and subtle document edits can be generated with a prompt.
The skill barrier to fraud has collapsed, but the accountability expectations for finance leaders have not. AppZen occupies that fault line. Anant Kale and his team have spent the last several years turning policies and standard operating procedures into AI-native controls that inspect every transaction before money moves, treating spend as a continuous stream of signals rather than a pile of forms to spot-check.
This conversation with Kale explores what that new reality looks like from inside hundreds of finance operations: how generative AI has changed the volume and character of expense fraud, what defensive AI can see in documents and behavior that humans generally cannot, and where the current limits of “AI vs. AI” still leave room for risk.
The discussion moves beyond tools into operating models, such as how organizations can rewrite policies for synthetic receipts, shift from sample-based audits to pre-payment reviews, align better with HR, Legal, Compliance, and Finance around shared playbooks, and prepare for regulators and boards that now expect explainable, auditable AI decisions. It also looks ahead to 2026–2028, where leading teams will treat AI Agents as digital coworkers embedded in T&E, AP, and vendor management, working continuously to keep pace with an evolving fraud landscape rather than chasing it from behind.
How has generative AI changed the nature and volume of expense fraud you see across your customer base, and what patterns stand out most right now?
Generative AI has meaningfully increased both the volume and sophistication of receipt fraud across our customer base. The skill barrier has collapsed: what once required Photoshop expertise or access to a template now takes a single prompt, which is why we’ve seen AI-generated receipts quickly rise to a sizable share of fraudulent submissions. What stands out most is not isolated, egregious cases but repeatable, template-driven behavior—employees reusing the same synthetic restaurant layout, the same server name, or patterns that don’t match their normal travel. Fraud typically starts small, often as a “lost receipt replacement,” and escalates when the employee isn’t caught. CFOs increasingly view this not as an expense hygiene issue but as a systemic control risk.
When an employee submits an AI-generated fake receipt, what exactly does AppZen’s AI look for that a human approver is likely to miss?
We try to mimic what a great auditor would do, but at machine scale:
Template and layout anomalies. We compare the visual structure of the receipt against known patterns for that merchant and category. Tiny differences in logo placement, fonts, spacing, or line-item ordering can be statistically “off” even though they look perfectly fine to a manager eyeballing the claim.
Metadata and image fingerprint checks. Our models look for inconsistencies in image metadata, repeated hashes, and signs of synthetic generation. For example, we can flag when the same underlying image has been resubmitted with slightly modified amounts, or when metadata does not match the claimed capture context.
Merchant and price validation. AppZen validates that the merchant exists at that location, that the date and time align to the employee’s travel pattern, and that prices and taxes are within expected ranges for that vendor and region. A human approver usually does not cross-check a Hilton breakfast against a market benchmark.
Behavior over time, not just one receipt. Our “Fraud Pack” models and Employee Spend Trend reports look at sequences: repeated use of the same template, the same unusual server name, or identical receipts from different countries submitted months apart. This pattern analysis is where AI has a huge edge over a one-off human review.
Contextual policy interpretation. Because our models are trained on organizations’ standard operating procedures (SOPs) and policies, they can interpret nuanced rules (for example, when an alcohol limit should or should not apply) at the line-item level, rather than rubber-stamping “looks like a dinner.”
Where are the current limits of “AI vs. AI” in this space—what kinds of forged images or documents are still hardest to detect reliably?
The toughest fakes are high-quality AI receipts that are printed or displayed and then photographed, which strips away metadata and makes the image look authentically “captured.” Subtle manipulations of authentic receipts—changing only the date, tip, or attendee count—also sit in a gray zone where realism is high and risk signals are low. New or very small merchants with limited historical templates are another challenge because there is less ground truth to compare against. Defensive AI is strongest when patterns repeat; one-off, low-dollar synthetic receipts designed to blend into normal behavior will always require a careful balance between vigilance and over-policing.
How do you balance aggressive fraud detection with the risk of false positives that frustrate employees or slow down legitimate reimbursements?
AppZen’s approach is based on risk scoring rather than binary decisions. The platform evaluates every expense report, card transaction, and supporting document before any payment is made. Low-risk items are auto-approved to keep the process fast for honest employees. High-risk items are escalated for review.
Each alert includes a reason code that explains the underlying concern. For example, the system may note an inconsistent template, a mismatch between location and travel records, or arithmetic anomalies. This transparency helps approvers make informed decisions and avoids the frustration of unclear rejections.
We also maintain a strong human review loop. High-risk items always receive human evaluation before action is taken. The models are continuously retrained using feedback from millions of transactions, which allows us to reduce false positives over time while maintaining strong fraud detection.
Generative AI lowers the skill barrier for fraud. What changes are organizations actually making to policies, training, and controls in response, beyond buying new tools?
Leading organizations are modernizing their policies to explicitly address synthetic receipts. Many have added clear language stating that AI-generated or reconstructed receipts are prohibited, even when the underlying spend was legitimate. This removes ambiguity and sets firm expectations.
Companies are also shifting from sample-based audits to continuous and pre-payment auditing. This is possible because AI handles the volume and frees humans to focus on exceptions. Corporations are increasing their use of corporate cards and virtual cards, which reduces reliance on employee-provided documentation altogether.
Cross-functional fraud playbooks are becoming standard. Finance, HR, Legal, and Compliance align on how to investigate anomalies, how to classify intent, and when an issue should lead to coaching or disciplinary action. Education is also part of the strategy. When employees understand that the organization uses sophisticated AI to review all spend, the perceived likelihood of detection increases, which reduces misuse.
AppZen is embedded in finance workflows like AP and T&E. How are customers using the same AI stack that spots fake receipts to manage broader finance risks, such as fake invoices or vendor fraud?
The same underlying AI models that detect fake receipts also strengthen risk controls across Accounts Payable, vendor onboarding, invoice processing, and finance-related email workflows. In AP, AppZen identifies fake or altered invoices by analyzing layout, language patterns, and semantic meaning. The system also flags suspicious vendor bank changes, duplicate invoices, and activity that deviates from historical norms.
AI Agent Studio allows companies to convert their finance standard operating procedures into AI Agents. These Agents execute vendor onboarding steps, invoice exception checks, and approval flows in a consistent and autonomous way. Because T&E, card transactions, and AP all run on the same platform, customers gain visibility into cross-channel risk. A person who manipulates a travel receipt may also submit questionable purchasing requests, and the system can connect those dots automatically.
What are regulators, auditors, and boards asking you about AI-driven fraud, and how is that shaping your product roadmap and data practices?
Regulators and auditors have become much more focused on documented preventive controls. Governments worldwide are implementing e-invoicing and real-time digital reporting to combat VAT/GST fraud. Frameworks like the United Kingdom’s ECCT, which penalizes companies that fail to prevent fraud, are increasing pressure on organizations to demonstrate active monitoring for misconduct. Boards want clear evidence that AI-driven decisions are explainable and auditable.
This has influenced our roadmap in several ways. We invest heavily in detailed audit trails, transparent reason codes, and a simulation environment where customers can test how an AI Agent will behave before deployment. Boards also ask about data residency, model training boundaries, and privacy protections. This is why we built capabilities like Private ZenLM, which allows customers to personalize model behavior without mixing their data with that of other tenants. Compliance and security expectations are now major drivers of AI adoption in finance.
How do you think about privacy and employee trust when your system is effectively profiling behavior over time to flag potential bad actors?
We maintain strict boundaries around the data we analyze. AppZen only processes financial transactions such as expense reports, card activity, invoice documents, and vendor requests. We do not monitor employees outside these workflows. We operate under GDPR, CCPA, and ISO 27001 controls and use strong encryption and access restrictions.
We promote transparency with customers about how behavioral models work. The purpose is to protect company funds and enforce policy, not to profile employees personally. When an exception appears, the system provides a clear explanation of why it was flagged. This allows managers to have factual conversations rather than relying on subjective judgments. In practice, this increases trust because employees see that decisions are based on evidence.
Many organizations already have legacy expense and ERP systems. What distinguishes an “AI-first” finance operation from one that simply bolted an AI auditor onto its existing stack? What happens when other systems adopt AI and disagree on a finding?
An AI-first finance operation places AI at the core of the workflow, rather than at its edges. Every transaction is reviewed before payment by AI Agents that interpret company policies and SOPs. In this model, policies become the source of truth and are expressed as code through AI Agents. The outcome is consistent decision-making across expenses, cards, and AP.
When different systems disagree, AI-first organizations rely on the policy itself. They use the SOP as the governing rule and allow a human reviewer to resolve edge cases. These decisions are then fed back into the models to increase accuracy. AI-first teams also measure automation rates, accuracy, fraud detection, and false positives, and use that data to tune the environment continuously. By contrast, legacy bolt-on AI modules often sample transactions or operate after the fact, which limits their value.
Looking out to 2026–2028, do you expect AI-generated fraud to keep escalating, or will defensive AI catch up—and what will “good” look like for a finance team that wants to stay ahead of this arms race? (Or is staying ahead a pipe dream at this point?)
Attempts at AI-generated fraud will continue to increase because generative tools will keep improving and remain easily accessible. However, the success rate of fraud does not need to rise. Defensive AI is evolving rapidly and already detects behavioral and document-level patterns that humans or generic tools cannot see.
We expect that, by 2028, leading finance teams will use AI to audit all spend before payment, standardize their SOPs using AI Agents that act as digital coworkers, and maintain strict model governance aligned to regulatory expectations. We see them unifying controls across T&E, AP, and vendor management, ensuring that no pattern falls through the cracks. Staying ahead is absolutely realistic, and it requires ongoing investment and continuous refinement. Fraud prevention has always been an evolving environment rather than a one-time project, and the organizations that adapt to changes quickly will maintain trust, accuracy, and financial integrity.
About Anant Kale
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 management, click here.
Did you enjoy the Nashay Naeve interview? If so, like, share or comment. Thank you!
Leave a Reply