The AI conversation has been obsessed with size for the last few years.
Larger models. Larger datasets. Larger budgets. Bigger claims of what’s about to come.
It’s not hard to see why. Large language models (LLMs) have changed the way that businesses think about content, customer service, research, coding, internal knowledge, and productivity.
They’ve opened up possibilities that not long ago would have seemed unrealistic.
But as more and more companies move from testing AI to using it on a daily basis, a more pragmatic question is starting to emerge.
Does every job really need the biggest model?
I don’t think it does.
In fact, I think one of the next big competitive advantages in AI will come from companies that know when to think smaller. With no less ambition, but with more precision.
That’s where Small Language Models, or SLMs, come in. While they may not get the same kind of attention as the largest models, they could be one of the most useful elements of an enterprise AI strategy.
Because in business, the best tool isn’t always the most powerful one. It is the one that fits the job.
The AI Conversation Is Maturing
The first wave of generative AI was all experimentation. Businesses wanted to see what it could do.
- Would it write?
- Summarise?
- Analyse customer queries?
- Create code?
- Help with research?
It was a hands-on way for people to experience the potential of AI.
But experiment is not transformation.
The real challenge comes when the AI is put into real workflows.
- Is it accurate enough?
- Can I afford it?
- Is it safe?
- Can it work at scale without new risks or unnecessary costs?
This is where the “bigger is better” mentality starts to fall down.
Most businesses don’t need an AI that can answer every question imaginable. They need AI that can solve specific problems in specific contexts.
Customer support teams need to understand intent and route enquiries. Compliance teams need to look through documents and find risk. A marketing team needs to make sense of search behaviour, content gaps, and audience intent.
These are not imaginary AI problems; they are real business challenges.
These challenges need the right solution, not the biggest tool.
A Better Question For Leaders To Ask
One of the most useful questions a leadership team can ask is a simple one: how much intelligence does this task truly need?
Some tasks need deep reasoning, wide context, and a powerful general-purpose model. Some don’t. Lots of business tasks are narrow, repetitive, and well-structured. Creativity is less important to them than consistency. They need speed, not sophistication. They need control more than they need breadth.
Here are five things that businesses need to look at to make the right choice:
- Complexity of the task: Is it open-ended reasoning or a repeatable process?
- Data sensitivity: Does it involve customer records, contracts, financial data, or internal IP?
- Usage volume: Will it be used once or will it be used thousands of times across the business?
- Speed: Is it a workflow that requires an instant answer?
- Risk: Is the likelihood of poor output causing legal, financial, or reputational damage?
These questions help leaders break free from model hype and move towards better decision-making.
If a task is low risk, high volume and clearly defined, a smaller model could be the smarter commercial choice. For ambiguous, high-risk, or strategically important tasks, a larger model and human oversight may be needed.
The point is to select the right model because the business case requires it.
Cost Will Need Greater Discipline
Many AI pilots are small in size, which makes them look very impressive.
Once AI begins to be deployed across departments, cost becomes tougher to ignore. Each query, output, integration, and workflow comes with a price tag. Something that looked affordable in a pilot can become expensive when it’s used by hundreds or thousands of people every day.
This is why smaller models are important.
If a smaller model can reliably classify enquiries, extract information, summarise documents, or search internal knowledge, then you may not need a larger model for the same task. It is a bit like asking your top strategist to deal with every basic admin request. They could but that doesn’t mean they have to.
The better approach is to use bigger models where they are actually needed and smaller models where they are enough.
That’s not taking shortcuts. It’s rational resource management.
As repeated McKinsey research shows, many organisations are adopting AI, but few have reengineered their workflows to capture value at scale. That’s a big gap. The companies that close it will not only be the ones with access to powerful technology. They will be the ones who redesign processes, costs, and decision-making around where AI actually adds value.
Data Control Can’t Be An Afterthought
Data is another reason businesses should pay attention to SLMs.
Many companies want to deploy AI more broadly but are understandably cautious. Their business involves customer data, contracts, financial information, employee records, commercial plans, and intellectual property. The question here is not only if AI can help but also where that data goes.
Sometimes, smaller models give organisations more options. They can be easier to run in private environments, on internal systems, or closer to where the data is created. That can be useful when building AI systems around the sensitivity of the work.
Smaller doesn’t necessarily mean safer, of course. Testing, monitoring, and governance are required even for a small model. It can still go wrong.
But it can help companies build AI more proportionally. A model used to summarise public marketing content does not need the same level of control as a model used to support legal review or financial decisions. Governance should be proportional to risk.
We need more of that kind of thinking in business.
Not The Flashiest But The Most Useful Use Cases
Some of the most valuable uses for smaller models aren’t always the most exciting. They’re often the practical, slightly unglamorous use cases that remove friction from everyday work.
Smaller models can help sort queries, understand urgency, summarise conversations, and route people to the right team in customer service. That is way more useful than another generic chatbot that keeps customers in a loop.
To be compliant, they can help review standard documents, compare policies, or flag missing information. They can help to interpret maintenance notes or equipment logs in manufacturing. In HR, they can help employees understand policies, onboarding materials, and internal processes.
There is a clear role for more focused models in marketing and search as well. Looking at digital growth, search, and content strategy, businesses are already dealing with a more fragmented discovery journey.
Viewers are not just using search engines. They use AI assistants, answer engines, social platforms, comparison sites, and specialist communities.
Smaller models can help with content classification, intent mapping, topic gaps, and information organisation to enable better decision-making. This may not sound as appealing as a general AI system that can do everything, but it is exactly the kind of focused application that can generate real operational value.
A business doesn’t always need AI that can do it all. More often than not, they need AI that can do a single thing very well, over and over again, at no extra cost or risk.
The Risk Is Letting Teams Do Their Own Thing
Smaller models are easier to deploy and so can spread quickly through an organisation. This creates a risk of model sprawl, where different teams start using different systems without central visibility.
Unless businesses know which models are in use, what data they can access, how they are monitored, and who is accountable for their outputs, they do not have an AI strategy. It is a collection of experiments.
They need to know what models are in play, what they are supposed to do, what data they have access to, how their performance is validated, where human review is necessary, and who owns the risk.
That may sound operational, but it is strategic. AI can’t scale properly without this kind of structure.
You don’t want to slow teams down. It should be to make sure that innovation can happen safely and repeatably.
The Future Is Hybrid
LLMs are still incredibly valuable. They are useful for complex reasoning, creative work, research, strategy, coding, and ambiguous problems where the answer is not clearly defined.
But they should not be the default for everything,
The future is hybrid. Businesses will use large models where they need reasoning and breadth. Where they need speed, focus, cost control, and data discipline, they will use smaller models. They will use humans where judgement, accountability, and context matter the most.
NVIDIA has highlighted the importance of small language models in scalable agentic AI systems, especially in scenarios where tasks can be routed, monitored, and optimised across different model types. That direction seems to be important.
It indicates that the future is not going to be one big model sitting at the centre of every workflow. It will be a more intelligent network of models, systems, and human decision points.
That’s where AI maturity really begins to show. Not in the number of tools a business has adopted, but how well those tools fit into the way the business actually operates.
Thinking Smaller Is Maturity
There’s a tendency in technology to conflate scale with ambition. More extensive platforms, larger systems, and more extensive transformation projects.
However, some of the best business decisions are about restraint.
Not using the biggest model when you know you shouldn’t. That’s not a lack of ambition. It is a sign that a company has grasped what it is trying to do.
The businesses that win the next phase of AI will not be the ones that use AI everywhere just because they can. They will be the ones who match the right model to the right task, with the right controls, the right cost structure and the right human oversight.
The first phase of generative AI paid off curiosity. The next phase is rewarded by judgement.
That means knowing when to use an LLM, when to use an SLM, when to build a hybrid system, and when not to use AI at all.
For business leaders, this may be the most important change of all. The AI strategy is moving from access to insight.
The next advantage of AI will not be in thinking bigger by default. It’ll come from thinking smaller, sharper, with far more intent.
