When it comes to AI, bigger isn’t always better

There is growing concern about trust in AI as the technology is adopted by more people. Large language models (LLMs) continue to face persistent challenges with hallucinations and inaccurate outputs.

LLMs are probabilistic systems trained to give answers, even when the correct answer is unclear or unknowable with the context provided. Humans are more likely to admit that they do not know an answer, especially when there is a financial or reputational consequence at stake, while on the other hand, LLMs are designed to act confidently, no matter what.

Most leading models fall within a 20 to 27 percent range of hallucination rate, making it a persistent and unresolved challenge across current AI systems, because it’s not just an architectural problem, it’s a contextual one.

Enterprise AI can’t afford hallucinations

While consumer AI is typically optimized for scale and creativity, enterprise AI must optimize for consistency and precision.

There are many high-stakes industries where getting the answer wrong can have detrimental and long-lasting effects. For example, in healthcare, legal and finance, the margin for error is zero, and one single hallucination can lead to serious consequences. A wrong supplier name, a misread total, or a compliance misstep isn’t a quirky model behaviour, it’s a liability. Typically, in business, it’s not just the one issue that is the concern, it’s the compounding of issues and scale.

In enterprise AI tools, just bolting a general-purpose LLM onto a workflow and hoping for accuracy is a dangerous gamble. Frontier models can change overnight, resulting in a workflow that was 92 percent accurate on Monday but, by Tuesday, produces entirely different results, may be under export controls, or may refuse to process some items. When your product has dependencies on something not designed for the job, you may not get the accuracy you need or the cost you expect because you’re effectively renting your house, and the cost of the rent can change at any time. It’s fast to build with frontier AI models, but you can usually tell when there is no accuracy claim: ‘AI can make mistakes, we may or may not train on your data…’

Before generative AI, we lived in a world of deterministic code – there were bugs, but you could reason over the system. As we move into a world of generative AI and purely probabilistic systems, things are going to behave differently. Looking forward, enterprises are aiming to achieve a balance between the two. A blend of deterministic logic, specialised models, frontier systems and the correct grounding context with human supervision. Building and maintaining this orchestrated symphony at scale—while ensuring absolute trust—is a non-trivial challenge.

Enter SLMs: Faster, cheaper and more accurate

We need to move away from a one-size-fits-all approach to AI, or even a one-model system, and this is where small language models (SLMs) come into play. SLMs are specialised, domain-specific language models designed for a purpose.

By training on narrow, high-quality datasets, these models operate in a world focused on accuracy. Because they leverage highly targeted, niche datasets, compared to the internet and world knowledge which LLMs are trained on, SLMs are inherently leaner, faster and cheaper to run. Consequently, their logic is easier to reason about, allowing them to guarantee a much higher degree of accuracy in their output. Unlike LLMs, they aren’t trying to be clever; they’re trying to be correct.

A report from Gartner found that LLM response accuracy declines when tasks require specific business context. As a result, they predict that by 2027, smaller, context-specific models will see usage volumes at least three times greater than those of general-purpose LLMs. It’s critical to state that the difference lies in the focus and the distribution of data these models have seen.

Early adoption of AI was driven by experimentation and productivity gains. The next step is for AI systems to influence operational decisions, and that shifts the standards required for trust. For CIOs and technology leaders, we need to build systems that operate consistently under real-world conditions – ones that maintain performance over time and will withstand regulatory and customer scrutiny.

Building trusted systems

The most effective enterprise AI tools will combine both LLM and SLM models – using LLMs for orchestration and SLMs for deterministic verification. By using multiple specialist models built for precision, organizations can build AI systems with accuracy they can stand behind and that can run efficiently from a cost perspective.

1. Getting work done right requires AI and humans to work together

AI adoption isn’t a binary choice between models; true operational resilience comes from utilizing the best elements of different architectures. When large and small models are combined, large thinking models can dispatch and reason while smaller models can act and verify, and the whole system can adapt to improve itself.

Models only get better when there are feedback loops through signals and context. On the surface things should look simple and feel like magic, but under the covers, the complexity is often many layers deep.

In financial document processing, for example, large thinking models can handle complex reasoning and user/company/accounting preferences, while the smaller models handle values, tax and line items. The key is that it is not a one-size-fits-all, and it’s not just the models; it’s the embeddings of similarity for processing and the context and guidance that shapes the outcome.

2. Creating best in class

To build a best-in-class system, you need data, compute and a feedback loop. We achieve this by training in-house models on our billions of documents for fast and accurate extraction, then leverage frontier-level models for reasoning. Relying solely on frontier-level models for extraction would tank our accuracy and skyrocket costs. This balance of accuracy comes at the price of generability and reasoning, but leveraging reasoning post-processing provides a top-tier solution, especially when that reasoning is looking to mimic the user’s preferences.

3. The human feedback loop

Human and AI oversight must be built into frontier systems from the beginning, not treated as a fallback when something goes wrong. There are always edge cases that have “it depends” answers – sometimes the answer is we do not know, but without any auditing or sampling, it’s impossible to know how well you did. AI and complex systems can drift, and accuracy is heavily dependent on the distribution of data, so continuous sampling and monitoring are critical components to these systems.

Human users and reviewers in the accounting world are those who are accountable for outcomes, and they provide the signal of feedback to the models and weights. On our systems, we sample over 100,000 documents every month to ensure accuracy is a measurable metric, not a hope.

In our system, when a bookkeeper accepts, edits, or rejects AI suggestions, that data is fed into an evaluation loop too. The result is a continuously improving system that understands the unique financial context of each of the 700,000 SMBs in our system. Through these feedback loops and rigorous evaluations, outputs remain accurate and dependable over time.

The AI hype

The market is flooded with AI tools and demos that trivialise complex systems. Yet many of these crumble when they encounter the complexity and nuances of real-world workflows, leading to a new tagline: “AI can make mistakes”.

This era of AI will not be defined by who is building the biggest model or who has the best demos. It will be defined by those delivering the biggest sustainable impact – the true overnight success stories, 15 years in the making.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?