Not All AI is Created Equal - What Finance Leaders Need to Understand About AI Architecture

The hype surrounding AI is louder than its current functionality and impact. Our Co-founder and CTO, Saurabh Agarwal, shares a no-nonsense download that finance (and even tech) leaders absolutely need to know to separate the crux from the clutter.

TLDR: AI is this new tool in the toolbox. It is changing the pace of building products, the capabilities (unlocked by AI) and how humans interface with them. O2C is a long-term commitment, you can't switch it as often as your iPhone. In a world where we have the power to ship features in days and not months, the true differentiator becomes the foundation - how it is built, down to its architecture.

Retrofitting AI features on outdated automation systems is as good as fitting an F1engine on a horse-drawn carriage. The issue isn't AI; legacy systems it runs on were designed for human usage only.

Think of an AI contract parsing feature that ingests your documents and outputs key details. That in itself can save some time, but AI can do a lot more, but not in a system that has fundamental limitations to automation.

AI-first architecture from day 1

Four years ago, as we started Zenskar, AI was in its early stages. So our vision always included leveraging it. We built our product’s architecture with AI at the core from day one, long before the hype. Over the last two years, we made infrastructure choices that unlock the full potential of AI, including the ability for agents to operate autonomously.

These decisions are why we deliver the most powerful AI in O2C, and this is only the beginning. The advantages will continue to compound because of the foundation we laid then and the choices we continue to make today.

Foundational flexibility - A powerful deterministic model to augment AI

It’s a common misconception to think that AI completely replaces existing systems. The reality is more nuanced: AI can be extraordinarily powerful, only when paired with an equally powerful foundation.

In other words, AI inherits the limitations of the system it's on (the F1 engine on the horse carriage).

Let’s dive into this with an example: Your existing billing system can only handle simple product subscriptions, either prepaid or postpaid, not both in the same subscription. No matter how sophisticated the AI becomes, the system will still be fundamentally constrained by the underlying limitation.

For this system to fully overcome its limitations, it has two poor choices: replace the engine entirely with AI, which requires extensive human oversight, or build itself from scratch to overcome the said limitations in the first place. Either option requires it to rebuild its architecture from the ground up.

We took a different approach. Our deterministic layer supports infinite complexity - every pricing model, billing terms, revenue model, and contract structure. This architecture gives AI a flexible foundation on which to function. As a result, our AI can handle the most complex scenarios without human intervention.

Secure data architecture

Most systems store all the customer data in a single database. This is efficient for storage and operations but creates a major limitation for AI to access, understand, and analyze all of the data present in the system with full freedom, without compromising on privacy, security, and accuracy.

Our architecture allows AI secure and isolated access to the data of individual clients. This allows it to be more than a chatbot overlay. It can access and query customer data (contracts, invoices, payments, usage, etc.) securely.

Powerful actions layer to take assigned actions

LLMs excel at comprehension and generation, but cannot act on it. Function calling is a feature that provides LLMs with the capability to do so. AI tools today are built with this as a foundation, not an afterthought.

Consider a real scenario: When a customer downgrades mid-month, finance teams typically spend hours manually recalculating prorated charges, updating revenue schedules, and adjusting future billings. Traditional LLMs can understand this statement and even generate a response about it, but they can't process it further. 

‍Function-calling allows AI to perform actions on your behalf- such as triggering ingestion or calling APIs. In this instance:

• The downgrade triggers a contract ingestion
• AI calls the Contracts API to pull the latest contract data
• AI analyzes and parses contract terms
• Deterministic rules
  
  • Update billing configurations automatically
  • Adjust future invoice amounts with updated numbers
  • Update performance obligations or revenue schedule for accurate revenue recognition
• AI notifies a human to review and approve the changes

Function calling is a building block to complex workflows and multi-step processes.

Communication with your tech stack

Communication between AI agents, and AI agents and external systems glues together different individual agents and modules for a cohesive O2C process. In our previous example, the AI’s ability to trigger deterministic computation, or notifying a human to approve the changes, is through A2A communication. 

AI communicates with other tools and systems (function calling is limited within the platform) through set protocols - MCP being the one that the industry is converging on right now. We recently launched our MCP server, you can learn more about it here.

Let’s understand its implications with an example. A customer notifies your team that they’ll make a payment after a certain date in a meeting.

• This information can be processed through your notetaker, or an internal update via Zenskar’s AI MCP
• The dunning agent can
  
  • Pause the automated emails
  • Tailor this particular customer’s follow-up emails taking their payment request into account
• It notifies a human to confirm the changes and approve the updated communication

What AI can and can't do today

Finance teams are already using tools like ChatGPT for day-to-day tasks, seeing significant productivity gains from data transformation and by automating repetitive work. That is where AI wins. It extends to:

• Data entry: AI can process invoices, receipts, and financial documents with remarkable speed and accuracy.
• Invoice processing: It can extract information, match purchase orders, and route approvals without human intervention.
• Basic bookkeeping: Transaction categorization, bank reconciliations through pattern matching combined with deterministic rules.
• Standard reporting: Those monthly reports that take hours to format? AI can generate those while you focus on analysis.
• Prediction and alerts: AI can spot trends and predict outcomes in ways that would take humans significantly longer to identify manually.

AI is, however, not suited for computations and calcuations as accuracy remains a primary concern. No matter how sophisticated, AI systems today carry a small probability of errors or hallucinations. In critical financial workflows, this translates to potential compliance risks or incorrect financial reporting. So we rely on rule-based deterministic models for all key math processes such as invoice generation, transaction price calculation, revenue distribution and adjustments, journal entry creation, and so on.

Our approach to bridging this gap is through strategic guardrails for AI. For processes involving money movement or customer-facing communications, all AI-generated outputs go through human verification before execution. We also show AI’s reasoning process and steps so users understand and can review its decisions, while it learns and improves.

As AI accuracy improves and methods are developed to guarantee it, we can gradually reduce these human verification steps. In the meantime, this approach ensures you get AI's power without compromising financial accuracy or compliance.

What’s in a name: AI-first, AI-native, AI-powered, and so on

In this case, everything.

AI-powered: Adds AI features to existing workflows. This is like having a smart assistant who still needs to follow manual processes. Technically, any platform with AI features can be called AI-powered.

AI-native: Built with AI capabilities in its core architecture with intelligent workflows that adapt and optimize automatically. As we've covered in the article, these are not simply AI-powered versions of existing categories, but can do a lot more.

Agentic: AI that can act independently, autonomously to make decisions and take actions without human oversight. While this is the dream (including ours in the future), certain key finance processes cannot be fully autonomous today.

Choose an AI-native solution like Zenskar for impact

The finance leaders who recognize that AI architecture, not just features, determines long-term success will build sustainable competitive advantages. The question isn't whether to adopt AI, but whether to build on a foundation that can evolve with it.

Zenskar ingests contracts and usage data to automate everything downstream – billing, revenue recognition, accounting, collections & analytics with zero workarounds. AI-first has been our way of thinking, building, and working in a time when AI can augment humans. We've built an AI-native platform that maximizes your gains while minimizing the inherent risks of AI, including data security, accuracy, and human control.