AI is already transforming verticals like law (Harvey), coding (Cursor), and customer support (Sierra). Accounting has yet to see its breakout despite substantial VC and incumbent investments. The lag isn’t due to lack of opportunity. Accounting lacks the immediate value generation from LLMs that is catalyzing their rapid adoption in other domains. Accounting presents unique challenges that call for more sophisticated engineering scaffolding around LLMs. But once addressed, AI's impact on accounting should be commensurate with other verticals.
Challenge 1: Language vs. Numbers
LLM capabilities ushered in capabilities in processing and generating natural language that seemed hopelessly out of reach only a few years ago. These capabilities resulted in near instant value generation across a variety of verticals. In domains like law, where the work is almost entirely language-based, LLMs provided immediate utility with minimal additional tooling. Similarly, in customer support, LLMs excel at summarizing, understanding, and routing natural language tickets. In software development, which in many ways is a more structured form of natural language, LLMs deliver significant value nearly out of the box. Accounting, by contrast, is rooted in structured numerical data rather than unstructured language which LLMs excel at. Applying LLMs directly to structured accounting data doesn't offer the same immediate benefit.
Challenge 2: Hallucination is Unacceptable
LLMs are inherently probabilistic, which makes them susceptible to hallucinations in their raw output. In some domains, like image or content generation, this behavior can be beneficial, as creativity stems from hallucinations. In accounting hallucinations are disqualifying. There is no tolerance for fabricated journal entries or imbalanced trial balances. Any application of LLMs in accounting must address the Siren song of hallucinations.
Challenge 3: Embedded Ambiguity
Financial data may appear deterministic, it's numerical, after all, but in practice, it embeds business context that isn't explicitly captured in the numbers themselves. Consider choosing the correct GL account for a transaction or determining how a contract should be translated into a revenue schedule. These decisions often cannot be resolved purely through algorithmic analysis of raw data. Human-in-the-loop design is required, and the optimal paradigms for human-AI interaction are still evolving.
Accounting’s Structural Fit for AI
And yet, accounting problems have natural affinities with AI. The data is quantifiable, which means outputs can be explicitly validated and refined by the model. In other words, the models have an objective they can optimize towards (these explicit objectives are missing in many of the verticals mentioned above). So how do we make LLMs effective with financial data? Some of the key requirements are: (a) translation layers that convert natural language and tabular data into verifiable deterministic systems which don’t hallucinate; (b) infrastructure that allows the AI to iteratively validate and correct its outputs; and (c) interfaces that support effective human-AI collaboration.
Engineering AI for Accounting at Doyen
At Doyen, we're addressing some of the core challenges of applying AI to accounting data (see here for a recent Doyen AI post on AI-native products). For example, our platform translates natural language instructions into structured business rules and generates deterministic code to execute financial data migrations. We also apply uniquely developed technology to test the generated code within deterministic systems, ensuring correctness and consistency. Another area we are investing in is human-AI interfaces which enable humans to validate, analyze, and reconcile financial data. These are a few examples of how we're tackling the complexities of accounting workflows with AI. We're continuing to build this infrastructure in collaboration with leading billing, ERP, and CAS platforms to ensure broad compatibility and robust performance across real-world accounting systems.
AI hasn’t yet transformed accounting. But it will. Stay tuned.