> Markdown version of [/videos/100226-hard-problems-hide-in-boring-places-turning-accounting-workflows-into-ai-products](https://www.wearedevelopers.com/videos/100226-hard-problems-hide-in-boring-places-turning-accounting-workflows-into-ai-products). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Hard Problems Hide in Boring Places: Turning Accounting Workflows into AI Products

How do you deploy LLMs in fintech without failing strict compliance audits? Discover how Pliant’s dual-LLM architecture balances real-world reasoning with fully traceable, fintech-grade rule enforcement.

- **Speakers:** [Oleksandr Korotkykh](https://www.wearedevelopers.com/@oleksandr-korotkykh), [Tolga Sümer](https://www.wearedevelopers.com/@tolga-sumer)
- **Event:** World Congress 2026 Europe
- **Published:** July 10, 2026
- **Duration:** 21:52
- **URL:** https://www.wearedevelopers.com/videos/100226-hard-problems-hide-in-boring-places-turning-accounting-workflows-into-ai-products

## Summary

Alex and Tolga from Pliant explore the complexities of integrating AI into B2B payments and accounting—domains where errors and compliance breaches carry real financial consequences. To automate expense policy enforcement, they initially tested traditional deterministic rule engines, which proved unscalable and rigid in the face of unstructured data. Conversely, a naive LLM approach offered the necessary real-world reasoning to detect hidden policy violations (like identifying specific alcohol brands embedded within grocery receipts) but failed fintech-grade auditability requirements.

The team experimented with a standard hybrid model where an LLM compiled policy PDFs into logic for a deterministic rule engine. However, unstructured edge cases exposed the limitations of static rules when processing nuanced OCR line items. To reconcile real-world judgment with verifiable operations, Pliant decoupled their pipeline into a dual-LLM architecture. During onboarding, an LLM extracts policy rules with direct citations to the source document, submitting them for human-in-the-loop approval. At runtime, the LLM evaluates transaction data against this tightly constrained schema, returning specific broken rule IDs and explicit reasoning. 

This architecture highlights a crucial engineering distinction: auditability and strict technical reproducibility are not the same. By trading baseline latency for grounded, traceable AI judgment, developers can successfully deploy models in heavily constrained, high-trust environments. Ultimately, applying AI to concrete workflows demands testing against the messiest, uncleaned real-world data as early as possible to expose flawed structural assumptions long before production.

**Keywords:** accounting workflows, B2B corporate payments, expense policy compliance, deterministic rule engines, OCR receipt processing, human-in-the-loop AI, fintech auditability, dual-LLM architecture, LLM policy extraction, structured data schemas, AI verification, unstructured data parsing, financial workflow automation, reproducibility vs auditability

## Chapters

1. **Managing corporate payments with programmable virtual cards** (00:02) — Treating corporate credit cards like scoped API tokens enables granular control over merchant spending and expiration limits.
1. **Extracting compliance rules from unstructured expense policies** (03:11) — Company expense policies are typically long, obfuscated legal documents that hide complex compliance requirements deep in text.
1. **Analyzing structured transaction data and invoice line items** (05:33) — Transactions arrive as complex JSON payloads containing detailed accounting metadata and OCR-extracted receipt items for evaluation.
1. **Evaluating deterministic rule engines for transaction compliance** (06:38) — Traditional rule engines offer processing speed and transparency but struggle to scale with rapidly changing bespoke customer configurations.
1. **Replacing manual code creation with naive LLM prompts** (08:49) — Directing an artificial intelligence model to blindly assess compliance handles natural ambiguity well but completely destroys system auditability.
1. **Compiling natural language policies into strict deterministic rules** (10:16) — Using underlying language models to generate executable code initially seems like the perfect structural balance of flexibility and transparency.
1. **Failing to catch real-world nuance with fixed lists** (11:45) — Code-based pipelines fail when receipt items require generalized contextual knowledge, forcing engineering requirements to mandate flexible judgment alongside auditability.
1. **Designing an auditable language model evaluation architecture** (15:07) — Extracting simplified rules linked directly to original source documents allows dynamic evaluation while enforcing strict boundaries and human-in-the-loop approvals.
1. **Surfacing system reasoning and rule violations visually** (17:11) — Returning specific violated policy IDs and grounded reasoning ensures user interfaces can explicitly map decisions back to trusted documentation.
1. **Trading operational latency and reproducibility for actual accuracy** (18:15) — Utilizing large language models in functional paths requires processing queues in periodic batches and ultimately accepting non-deterministic execution states.
1. **Exposing hidden failures using messy production data immediately** (20:51) — Constructing robust AI features necessitates testing against raw, unstructured real-world data instead of sanitized inputs to discover actual limitations early.

## Related Moments

- [Testing AI limits in enterprise software design](https://www.wearedevelopers.com/videos/100340-ai-driven-development) (from "AI Driven Development")
- [Securing heterogeneous legacy payment infrastructure against AI](https://www.wearedevelopers.com/videos/100331-fighting-the-next-wave-of-cybercrime) (from "Fighting the Next Wave of Cybercrime")
- [Transitioning from demos to real business processes](https://www.wearedevelopers.com/videos/100328-the-limits-of-llms-in-real-world-applications) (from "The Limits of LLMs in Real-World Applications")
- [Identifying and fixing over-engineered AI calls through observability](https://www.wearedevelopers.com/videos/1465-event-driven-architecture-breaking-conversational-barriers-with-distributed-ai-agents) (from "Event-Driven Architecture: Breaking Conversational Barriers with Distributed AI Agents")
- [Addressing corporate compliance challenges with secure enterprise AI](https://www.wearedevelopers.com/videos/1665-secure-and-private-ai-deepmask) (from "Secure and Private AI - DeepMask")
- [Shifting from AI hype to enterprise operations](https://www.wearedevelopers.com/videos/100328-the-limits-of-llms-in-real-world-applications) (from "The Limits of LLMs in Real-World Applications")

## Related Articles

- [MLOps – What’s the deal behind it?](https://www.wearedevelopers.com/magazine/125-mlops-what-s-the-deal-behind-it)
- [MLOps And AI Driven Development](https://www.wearedevelopers.com/magazine/82-mlops-and-ai-driven-development)
- [Stephan Gillich - Bringing AI Everywhere](https://www.wearedevelopers.com/magazine/489-stephan-gillich-bringing-ai-everywhere)
- [Panel Discussion: Responsible AI in Practice - Real-World Examples and Challenges](https://www.wearedevelopers.com/magazine/488-panel-discussion-responsible-ai-in-practice-real-world-examples-and-challenges)

## Related Jobs

- [Principal Software Engineer, Enterprise AI Platform](https://www.wearedevelopers.com/jobs/ext/1467292-principal-software-engineer-enterprise-ai-platform) at **GitHub**
- [AI Software Engineer (Germany)](https://www.wearedevelopers.com/jobs/48317-ai-software-engineer-germany) at **Sunhat**
- [Senior AI Agent Software Engineer (Go, Python) (m/f/x)](https://www.wearedevelopers.com/jobs/48277-senior-ai-agent-software-engineer-go-python-m-f-x) at **Dynatrace**
- [Staff Software Engineer, Copilot Experiences](https://www.wearedevelopers.com/jobs/ext/164361-staff-software-engineer-copilot-experiences) at **GitHub**
- [Machine Learning Engineer](https://www.wearedevelopers.com/jobs/ext/588393-machine-learning-engineer) at **Twilio**
- [AI Operations Manager (all genders)](https://www.wearedevelopers.com/jobs/48263-ai-operations-manager-all-genders) at **envelio**