- Half of accountants have experienced losses to their companies due to mistakes made by large-scale language model (LLM) chatbots.
- Small Language Models (SLMs) are trained on narrow, high-quality datasets. By removing the broader general knowledge that often creates hallucinations in LLMs, these specialized SLMs can focus purely on accuracy.
- SLM runs faster, costs less to run, and guarantees much higher accuracy of the output. They’re not trying to be smart, they’re trying to be right.
- The next era will not be won by those with the biggest models, but by those who build the right architectures around these models. The expert foundation, human audits, and feedback loops are all built on proprietary data and infrastructure invisible to frontier research institutions. This is the result of many years of effort.
The trust gap in AI is widening. This is especially true in the high-stakes world of finance. There is zero margin for error, and a single illusion can lead to financial losses and compliance nightmares. Invoices with incorrect supplier names or extra zeros at the end are not strange model behavior, but a misrepresentation of someone’s account.
Our recent data shows that half of accountants have lost money to their companies due to mistakes made by large-scale language model (LLM) chatbots. This isn’t a story about AI failure; it’s a story about companies using the wrong form of AI for the job at hand.
Bolting a general-purpose LLM into your finance workflow and expecting accuracy is too risky a bet. The bottom line is that hope is not a building.
Frontier AI failure in finance alone
Frontier’s large-scale language models are good at language and inference, but are often insufficient to ensure complete accuracy. For example, if you ask a bot to read 500 invoices and extract the data with 100% accuracy, it will eventually fail.
This is because LLM is probabilistic. LLMs are trained to have an answer no matter what. Humans may say they don’t know the answers to complex questions, but LLMs are designed to act with confidence, regardless of whether their actions are accurate or not. Nothing in their training rewards abstention. The result is plausible speculation, which only widens the trust gap.
Another problem with using a frontier model alone is that the model can change overnight. When a provider ships an update, the weights shift and a workflow that was 92% accurate on Monday becomes something different on Tuesday. If your product masks someone else’s API, you can’t own your own accuracy. You’re renting it, and that rental changes all the time.
This is why, for example, pure LLM accounting tools tend to ship without guarantees of accuracy. There is no player they can protect with certainty. Auditing and sampling output at scale is a daunting task, and without it, you’ll be shipping reliable performance rather than a validated system.
smaller solution
Closing this trust gap requires moving away from a one-size-fits-all approach to AI. Enter the Small Language Model (SLM). It is a localized, domain-specific language model designed for accuracy.
LLM is trained on the entire public internet, which consists of billions of parameters, whereas SLM is trained on a narrow, high-quality dataset. By removing the broader general knowledge that often creates hallucinations in LLMs, these specialized SLMs can focus purely on accuracy.
Because SLM is highly targeted and operates on much smaller local datasets, it runs faster, costs less to run, and greatly guarantees output accuracy. They’re not trying to be smart, they’re trying to be right.
Best of both worlds – tuning the inference layer
Companies often mistakenly believe that AI deployments are an either-or model, when in reality, real operational resilience comes from leveraging the best elements of different architectures.
LLMs offer real value in natural language processing and advanced reasoning, while SLMs offer mastery of specific, narrow topics. Combining these models can truly improve your business, with LLM acting as the interface and inference layer and SLM acting as the deterministic validator.
Communicating these models requires a secure bridge using the Model Context Protocol (MCP). However, the role of the reasoning layer is to provide guidance, not just answers. Unearth the muscle memory of how professionals respond to specific clients and codify it into natural language rules. The system uses these rules to operate specialists.
For example, in the context of financial document processing, the LLM handles complex reasoning and conversations, the SLM handles the calculation and recognition of various tax rules, while the MCP securely routes queries from the LLM to the appropriate expert model. This ensures that the final output is actually grounded rather than probabilistic.
Like any technology, SLM has limitations. To maintain such a high level of accuracy, small language models must be kept narrowly bound. In other words, a set of SLMs is essential to cover all the bases in a complex field such as finance. These SLM suites form what is called an “AI swarm.” These are individual experts whose parameters are often less than 1B in size in various areas such as taxes and receipt extraction.
Creating a feedback loop
Human validation should be a core design feature, not a fallback. In high-stakes financial environments, giving every prediction a confidence score builds trust on a system that recognizes its own uncertainty. If the score falls below a set threshold, the score is returned to the human reviewer.
Trust architecture
Over the next year, the market will be flooded with impressive AI tools and demos, but these tools often fall apart when faced with the realities of complex workflows.
The next era of finance will not be won by those with the biggest models, but by those who build the right architecture around these models. This means that our expert foundations, human audits, and feedback loops are all built on proprietary data and infrastructure that is invisible to most cutting-edge research institutions. This is the result of many years of effort.
The goal is to help professionals spend less time on manual entry and more time on high-impact advisory work. Trust has never been more important in every industry. A specialized connected AI architecture is how we ensure that.
Stephen Edginton is the Chief Product Technology Officer at Dext.
read more
Goodbye Software as a Service, Hello AI as a Service – AI as a Service (AIaaS) is likely to replace Software as a Service (SaaS) in the near future. What this means for your organization
From Generative AI to Agent AI – Now the real transformation begins – Mark Skelton from Node4 explains the transition from generative AI to agent AI and how to approach it within your organization.
Will more AI mean more cyber-attacks? – Nick Martindale explains that increased use of AI within organizations could mean more cyber-attacks. Here’s what you can do
