AI-powered systems are misclassifying corporate expenses, not because they lack data, but because they misread grammar. What looks like a technical glitch reveals a deeper structural bias at the heart of automation. What does grammar have to do with accounting? More than you think. In corporate and government ERP systems, every expense description is read by an AI model that assigns it to a general ledger category. This happens automatically, often without human oversight. Now consider two phrases:**"Reimbursement for coordination of lodging and transportation services"**vs."The office coordinated lodging and arranged transport." "The office coordinated lodging and arranged transport." They mean the same. But one is far more likely to confuse the AI. Why? Because of syntax. Complex grammatical structures (like nominalizations or nested clauses) lead models to make the wrong call. The result? syntax Misclassification, accounting errors, and potentially audit failure. Misclassification, accounting errors, and potentially audit failure. The finding: AI fails not because it misunderstands—but because it obeys too literally \  A recent study shows that these systems don’t fail due to lack of training data. They fail because they follow the form, not the meaning. follow the form, not the meaning Transformer models like FinBERT rely heavily on the grammatical shape of sentences, often more than on their semantic content. Sentences with high syntactic density are statistically more likely to be misclassified. So what’s the risk? So what’s the risk? Misclassified travel expenses
Reimbursements logged under the wrong cost center
Audits that don’t reconcile
Regulatory violations
Financial misstatements Misclassified travel expenses Reimbursements logged under the wrong cost center Audits that don’t reconcile Regulatory violations Financial misstatements All triggered by a sentence that "sounded too professional." The fix: Rewrite to reduce risk The paper proposes a structured rewriting method called the fair-syntax transformation. fair-syntax transformation It reformats expense descriptions into a simplified Subject-Verb-Object (SVO) structure, stripping out misleading grammatical forms. Subject-Verb-Object (SVO) Instead of:"Coordination of transport and accommodation services for vendor engagement" "Coordination of transport and accommodation services for vendor engagement" Use: "The team booked transport and arranged accommodation." "The team booked transport and arranged accommodation." This intervention alone reduced classification errors by 15%. reduced classification errors by 15% It improved code accuracy and reduced false positives. The key insight? AI behaves better when we use simpler grammar. AI behaves better when we use simpler grammar. What this says about AI in finance AI systems don’t interpret, they execute. Their logic is not semantic but syntactic. That means that a well-written sentence can be more dangerous than a numerical error. Who should care? Who should care? CFOs, controllers, audit leads
NLP engineers building ERP integrations
Regulatory bodies developing audit-compliance standards
Any enterprise automating expense workflows
Anyone who’s ever submitted a reimbursement form with too much jargon CFOs, controllers, audit leads NLP engineers building ERP integrations Regulatory bodies developing audit-compliance standards Any enterprise automating expense workflows Anyone who’s ever submitted a reimbursement form with too much jargon The uncomfortable conclusion: Grammar is not neutral. In automated systems, it governs decisions. it governs decisions When syntax misaligns with accounting intent, the system doesn’t correct—it complies. Author of the study: Agustin V. Startari[DOI (Zenodo): https://doi.org/10.5281/zenodo.16322760][SSRN Author ID:https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915][Website:https://www.agustinvstartari.com/][ResearcherID: K-5792-2016 | ORCID: 0009-0001-4714-6539] [DOI (Zenodo): https://doi.org/10.5281/zenodo.16322760 [SSRN Author ID: [Website: [ResearcherID: Ethos Ethos I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.— Agustin V. Startari I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.

Tokenomics by Grammar

Expense Coding Errors Aren’t Random—They’re Rule-Based, New Study Finds

ORCID

SSRN Author Page

Zenodo

Personal Blog

Figshare

Medium

Hcommons

Scholar

Research Gate

Academia

Web of Science

Read My Stories

Too Long; Didn't Read

Make resilience your competitive advantage

The Hidden Risk in ERP Automation: How One Bad Sentence Can Cost Your Company Millions

The Hidden Risk in ERP Automation: How One Bad Sentence Can Cost Your Company Millions

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Untitled Story

A Radical Guide to AI Sovereignty

The Noonification: Use This 7-Step McKinsey Framework to Solve Any Problem (1/10/2023)

The Noonification: A Taxonomy of Inclusiveness (1/11/2024)

The Noonification: What is the InfiniteNature-Zero AI Model? (11/19/2022)

10 Ways AI Has Changed Our Lives

A Radical Guide to AI Sovereignty

The Noonification: Use This 7-Step McKinsey Framework to Solve Any Problem (1/10/2023)

The Noonification: A Taxonomy of Inclusiveness (1/11/2024)

The Noonification: What is the InfiniteNature-Zero AI Model? (11/19/2022)

10 Ways AI Has Changed Our Lives

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps