AI for Payments: Why "Prompt & Pray" fails and what scales safely [Chapter 4 - AI Deep Dives]
George
Chief of AI at Moveo
September 18, 2025
in
✨ AI Deep Dives
Welcome to another installment of our "AI Deep Dives" series! We're demystifying how to implement AI systems in enterprise environments.
In previous chapters, we showed why the Multi-Agent System architecture is a more robust alternative to the Wrapper method. In Chapter 3, we discussed the dangers of the "Prompt & Pray" approach, the belief that a more advanced Large Language Model (LLM) can solve everything on its own.
In payments and accounts receivable, close isn’t good enough. A 1% error rate isn’t a rounding error. It’s a chargeback spike, a complaint, or an audit exception. ‘Prompt & Pray’, a single mega-prompt or opaque wrapper, can look impressive in demos, but it doesn’t survive policy gates, sequencing, or audit trails at enterprise scale.
What works instead is a blend of deterministic and probabilistic AI: deterministic flows encode business logic, governance, and control, while probabilistic models handle interpretation, empathy, and language.
Now, let's explore a practical example from an industry where sensitivity, structure, and compliance are fundamental pillars: payments & accounts receivable (AR).
Managing payments is more than just transactions, it's a human interaction that often deals with delicate situations for the customer.
For the business, it's a critical function for financial health. For leaders, the scoreboard is Days Sales Outstanding (DSO), right-party contact rate, promise-to-pay kept, chargeback losses, complaint rate, and audit exceptions. For the customer, it can be a moment of vulnerability that requires empathy and clarity. On top of that, it's a highly regulated industry where any misstep can lead to complaints, fines, and reputational damage.
This is where the "Prompt & Pray" approach is especially dangerous. Trying to cram all the rules for tone, eligibility checks, and authentication steps into a single "mega-prompt" is a risky gamble.
When the system fails (and it will fail), the consequences are severe. The model might use coercive language, skip mandatory steps, or even create payment plans with incorrect parameters. These errors, even if rare, are unacceptable and can be extremely costly.
Think of a powerful LLM as an autopilot: superb at keeping things smooth, but every airline still uses checklists, gates, and air-traffic control. In payments and accounts receivable, those external controls are consent, authentication, disclosures, and audit logs.
The risk of non-compliance with "Prompt & Pray"
The fragility of the "Prompt & Pray" method in a regulated context, like in payments & AR, is alarming.
Simply instructing the prompt to "be gentle and empathetic" doesn't guarantee the model won't use a phrase like, "Pay now or we will escalate the case". This seemingly minor slip-up can easily lead to a formal complaint.
Instead of a fluid and safe experience, the lack of a rigid, controlled structure can result in:
Tone Violations: the absence of a post-response compliance check can allow the AI to use prohibited, threatening, or inappropriate language.
Process Failures: the AI might skip mandatory steps, such as consent verification, or simply fail to follow the correct sequence of a negotiation flow.
Fabricated Data: in more extreme cases, the model can "hallucinate" and create non-existent payment parameters or fake transaction numbers, leading to confusion and auditing nightmares.
Example
Customer: “I can’t pay in full. Can I set up €150 monthly starting next month?”
Prompt & Pray: Creates the plan on the spot. It skips authentication, never verifies consent, and even hallucinates a start date that was not confirmed. Later, the customer disputes it. There is no audit trail, so the plan is reversed, finance spends hours reconciling, and a complaint is filed.
Guardrailed flow: The system first confirms consent, then verifies identity with OTP, then creates the plan with the exact parameters. A compliance agent checks tone before sending, and every step is logged. The result is an enforceable plan, fewer disputes, and a clean audit trail.
→ Read Chapter 3 of "AI Deep Dives": The Problem with Prompt & Pray
The power of structure: How a Multi-Agent System works
As covered in Chapters 2 & 3, the fix isn’t “a smarter prompt”, it’s a smarter Multi-Agent System: small, specialized models where language helps, and deterministic flows where policy lives.
1. Tone compliance
Sensitivity in payments & AR begins with language.
In a Multi-Agent System, the task of ensuring the appropriate tone isn't left to the main Response Agent. Instead, a dedicated Post-Response Compliance Agent is responsible for analyzing the generated response before it's ever sent to the customer.
This agent operates with clear, deterministic rules, filtering out any prohibited language.
2. Structured negotiation flow
Setting up a payment plan, for example, requires a sequence of steps that cannot be flexible. A Response Agent (the LLM) can lead the conversation in an empathetic and natural way, collecting information about the customer's financial situation. However, the process flow is rigidly controlled by a deterministic system:
Consent: The customer must explicitly agree to share financial information.
Eligibility & Options: The system checks policy rules and presents only valid payment plan options, never inferred by the model.
Disclosures: the system ensures all legal and contractual disclosures are made.
Confirmation: The customer provides final, explicit confirmation.
Receipt: the system issues a formal receipt or transaction record.
The logic here is clear: the LLM is the "interlocutor," but an external rules engine ensure the business process is executed with precision and compliance.
3. Insights and Resegmentation
A Multi-Agent System does more than resolve the immediate interaction.
An Insights Agent analyzes the conversation to detect outcomes such as promise to pay, hardship, legal threat, or right-party confirmation.
These insights then feed into an agentic flow that determines the next best action: sending a personalized follow-up on a different channel, scheduling a reminder for the promised payment date, escalating a legal threat to a human agent, or re-engaging non-responders with a different strategy.
By continuously resegmenting customers based on real signals, the system becomes adaptive rather than static, ensuring that every follow-up is timely, relevant, and aligned with the goal of increasing accounts receivable.
Read Chapter 2 of "AI Deep Dives": Wrappers vs. Multi-Agent Systems in enterprise AI
From Conversation Signal to Next Action examples
Here is how the Insights Agent translates conversation signals into concrete next actions that keep accounts receivable moving forward:
From "Prompt & Pray" to robust production
The final lesson is that in sensitive and regulated domains, AI cannot be a generic tool. It must be an intelligent component within a broader, more robust system.
Payments & AR is a perfect example of how combining artificial intelligence with a well-thought-out system architecture is crucial for governance, compliance, and reputation. By adopting a structured approach, companies can be confident that their AI systems will operate with the precision, security, and auditability that corporate operations demand.
Next in Chapter 5, we’ll show why simply exposing APIs to a model (“Tools & Pray”) still fails on process control, argument validity, and auditability, and how to combine tool access with planners, validators, and flows to keep payments/AR safe.