What 2016 NLP taught me about LLM agents in finance

By Ariel Berger · October 19, 2026

Category: llms-and-nlp

What 2016 NLP taught me about LLM agents in finance

I built AI for accounting from 2016. Here are five lessons from that work that apply to LLM agents in finance today.

Key takeaways

The problem Vendors call LLM tools autonomous agents but rarely say how accuracy is measured or what happens when the model is wrong.
Core insight Measure accuracy on the posted entry, pair every model with a verifier, test across repeated runs, and write the governance first.
Practical outcome Ask any vendor for posted-entry accuracy, the verifier, repeat-run results and an action log before you buy.

I started building AI for accounting in 2016, four years before anyone called it "generative." Our models read financial documents and turned them into journal entries. Today I watch the industry rebuild that idea with large language models, and I keep seeing the same mistakes we made. Here is what I think the old work teaches about the new tools.

1. Measure the posted entry, not the model

Large open ledger book with a bold checkmark stamped on an entry line, with faint fading symbols floating above.

At Task Sheriff we used supervised models for NLP and computer vision, and we measured results that mattered to customers: about 2x better data accuracy and about 5x less manual work. Looking back, the lesson I would give any team is this. A field read correctly is not an entry posted correctly. The unit of accuracy should be the final entry, judged by a person who owns the ledger.

Today's vendors still mix these up. One AP vendor's own materials cite 99% for key-field extraction and roughly 90% fewer "steps" for coding. Those are two different claims, and neither is accuracy of the posted entry. If a vendor cannot tell you which one it measures, assume the weakest.

2. A model needs a verifier

I believe the best systems in finance pair a language model with something that checks it. Xero says its JAX assistant lets the LLM handle the conversation while separate models do the calculations. NetSuite's 2026 bank reconciliation suggests a match and waits for a person to accept it. Both put a check between the model and the ledger.

I hold the same view about the machine-learning systems we built in 2016: a classifier is only as safe as the checks around it. I think "autonomous agent" is mostly a label. Look at what happens when the model is wrong.

3. Reliability across repeats is the real problem

Averages hide the risk. The APEX-Accounting benchmark (arXiv 2607.27189, July 2026) reports a best mean score of 56.4%, but no model passed more than 2.6% of tasks on all eight attempts. I read only the abstract, and the tasks are synthetic, so treat the numbers as a signal and not a verdict. The signal matches my experience: a system that is right 9 times in 10 is still wrong often enough to need a safety net, because accountants run the same task thousands of times.

4. Your customers' corrections are the product

The data moat in accounting is not the pretraining. It is the stream of corrections: the entry the accountant changed, and why. Whoever sits on the transaction flow, such as the ledger, the card, or the payables tool, sees those corrections. A tool bolted on the side does not. This is my inference and not a measured result, but it is why we built Task Sheriff to be personalized to each customer, and customers reported about 10x less repetitive work.

5. Write the governance first

At DeckRobot, before ChatGPT made the topic fashionable, I wrote AI governance guidelines for Fortune 500 clients: compliance, brand safety and human oversight. My view since then: agree who approves what, and what gets logged, before you ship. Finance has less tolerance for error than most functions, so I would write the audit trail before the prompt.

What I would do now

If you buy or build an AI accounting tool this year, ask for four things: accuracy measured on the posted entry, the verifier, results across repeated runs, and a log of every action. A vendor that answers all four has done the work. I wrote this from the founder's chair, not from a lab. At Sage I later founded the team that built the conversational AI (NLU, NLP, LLMs) behind Sage Co-Pilot, and the same questions came up there. In my experience the people and the technology from Task Sheriff fed directly into that work and into Sage's AP automation. Sage can speak for its own products. I can speak for what we learned.

The jump from rule-based NLP systems to agentic AI in enterprise settings is rarely as clean as the demos suggest, and Jerry Ting's experience building contract AI at scale offers a grounding counterpoint to the current hype. His hard-won perspective on what actually breaks in production - and how enterprise constraints shape what's deployable - maps closely onto the lessons this article draws from early finance AI work. Worth watching alongside the points raised above.

The jump from narrow AI tools to agentic systems in enterprise settings involves hard-won lessons that rarely surface in research papers. Jerry Ting's experience building contract AI at scale offers a practitioner's perspective on what that transition actually looks like inside large organizations - the organizational friction, the trust gaps, and the deployment realities. It maps well onto the financial AI evolution this article traces.