For much of the Generative AI boom, progress was easy to associate with scale.
Larger models. More compute. Bigger context windows. More ambitious use cases.
That equation is starting to change.
As organizations put Generative AI into everyday products and business processes, the question is becoming less about how much AI capability they can access and more about how intelligently they can deploy it.
Three developments are accelerating that shift:
This creates a different challenge for technology leaders.
The most advanced model will not automatically be the best model for every task. A business may need frontier reasoning for one workflow, a smaller specialized model for another and a retrieval-based system connected to internal knowledge for a third.
Enterprise AI is therefore becoming an optimization problem.
The organizations that build the strongest AI capability will increasingly be those that can match the right model, infrastructure, data and controls to the right business problem.
A successful proof of concept can be surprisingly simple.
A team can connect a model to a dataset, create an interface and demonstrate a useful capability relatively quickly.
Production changes the requirements.
Gartner reported in 2025 that 60% of GenAI pilots fail to enter production. Its research also highlights value, security and governance, skills and adoption among the challenges organizations face when trying to scale GenAI across enterprise applications.
Moving beyond that point requires organizations to think about everything surrounding the model.
This is where LLMOps, RAG and guardrails become important.
LLMOps, or Large Language Model Operations, provides the processes for deploying, monitoring, evaluating and maintaining LLM-based systems after they go live. Gartner's 2025 MLOps framework similarly emphasizes operationalizing models in a way that supports continuous value and manages risk over time.
RAG, or Retrieval-Augmented Generation, gives models access to relevant information from trusted sources before generating a response. Instead of relying entirely on information learned during training, an enterprise assistant could retrieve the latest internal policy, product documentation or customer information before answering.
Guardrails provide another layer. They help validate model inputs and outputs, apply business rules and reduce the likelihood of inappropriate or unsafe behavior. Gartner's 2025 guidance specifically recommends implementing guardrails around GenAI inputs and outputs as part of a holistic approach to deployment.
Together, these capabilities make AI more manageable.
They also demonstrate why enterprise AI is becoming an operating capability rather than a series of model deployments.
That matters commercially. Gartner found that organizations conducting regular assessments of AI system performance and compliance were more than three times as likely to report high GenAI value as organizations that did not.
Operational discipline is not simply about reducing risk. Done well, it can improve the value organizations extract from AI.
Tenth Revolution Group helps organizations develop the cloud, data and AI capability needed to turn AI concepts into sustainable business solutions, from the specialists who build and operate systems to the training and transformation expertise needed to support adoption.
Model capability continues to advance quickly.
But another competition is taking place underneath it: making AI cheaper and more efficient to run.
Inference is central to that challenge.
Inference is what happens when a trained AI model processes a request and produces an output. Every employee query, generated summary, automated decision or AI agent action consumes computational resources.
At small scale, that may be relatively easy to absorb.
At enterprise scale, millions of interactions can change the economics considerably.
The encouraging development is that efficiency is improving quickly.
Gartner forecasts that by 2030, inference on a one-trillion-parameter LLM will cost GenAI providers more than 90% less than in 2025. It expects improvements in semiconductors, infrastructure, model design, chip utilization and inference-specialized silicon to contribute to that decline.
But falling unit costs do not automatically mean falling AI bills.
Gartner also estimates that agentic models can require between five and 30 times more tokens per task than a standard GenAI chatbot. As AI performs more complex work and usage increases, overall consumption can rise faster than individual token costs fall.
That changes the optimization challenge.
Instead of asking only "Which model performs best?", leaders increasingly need to ask:
These questions move AI architecture closer to commercial strategy.
New generations of GPUs and AI accelerators will continue improving performance.
But hardware alone will not determine the economics of enterprise AI.
Software and model optimization are becoming equally important.
One example is quantization.
AI models normally represent their internal parameters using numerical values at a particular level of precision. Quantization reduces that precision, lowering the amount of memory and computing power needed to run the model.
The objective is simple: preserve enough model quality for the task while making inference substantially more efficient.
McKinsey's 2026 analysis of inference economics found model optimization to be the strongest near-term lever it assessed. It estimates that moving from FP16 to INT4 precision can improve inference throughput by two to four times, although lower precision can introduce accuracy trade-offs that need to be managed carefully.
This is important for business leaders because it demonstrates that AI cost optimization does not necessarily mean using less AI.
It can mean designing AI better.
Model selection, quantization, workload routing, infrastructure utilization and application architecture can all influence the cost of delivering the same business outcome.
That creates an opportunity for organizations to make AI economics part of solution design from the outset rather than attempting to control costs once adoption has already scaled.
Not every enterprise task requires a frontier LLM.
That is becoming increasingly important as smaller models improve.
Stanford's 2025 AI Index illustrates how quickly that progress has happened. In 2022, the smallest model exceeding a 60% score on the MMLU benchmark had 540 billion parameters. By 2024, Microsoft's 3.8-billion-parameter Phi-3-mini crossed the same threshold, representing a 142-fold reduction in model size.
That does not mean a smaller model can replace a frontier model in every scenario.
It means organizations have more options.
Small Language Models, or SLMs, can be particularly attractive when a task is narrow, repetitive, latency-sensitive or needs to run closer to where data is generated.
A specialized internal assistant, classification workflow or high-volume customer process may not require the same reasoning capability as complex research or strategic analysis.
Gartner expects routine, high-frequency tasks increasingly to be routed toward smaller and domain-specific language models, noting that these can perform better than generic approaches at a fraction of the cost when they are aligned to specialized workflows.
For enterprise leaders, this introduces a useful principle:
Use enough model for the problem, rather than the most model available.
The same move toward choice is happening in model ownership.
Enterprise AI does not need to become a binary decision between proprietary and open-source models.
Increasingly, organizations are using both.
McKinsey's 2025 research with the Mozilla Foundation and Patrick J. McGovern Foundation found that more than 50% of respondents reported using open-source AI across the data, model and tools layers of their technology stack. 76% expected their organizations to increase their use of open-source AI over the following several years.
The economics are one attraction.
60% of decision-makers surveyed reported lower implementation costs with open-source AI compared with similar proprietary tools, while 46% cited lower maintenance costs. Proprietary tools retained an advantage in speed, with 48% associating them with faster time to value.
That trade-off helps explain why multimodel strategies are becoming more attractive.
An organization might use a proprietary frontier model where advanced reasoning and rapid deployment matter most, an open model where customization and control are priorities and an SLM for high-volume specialized tasks where speed and cost matter.
The goal is not to choose a single winner.
It is to create an architecture capable of choosing intelligently.
There is another side to this flexibility.
More models create more decisions.
Organizations need to understand where models come from, how they have been evaluated, which data they can access, how they are secured and whether they are appropriate for a particular business process.
Open-source AI brings its own considerations. McKinsey's survey found respondents identified cybersecurity at 62%, regulatory compliance at 54% and intellectual property infringement at 50% among concerns associated with AI tools.
The Heidrick & Struggles 2025 Global CISO Survey reinforces how closely AI capability and security are now connected. 57% of CISOs identify AI, machine learning and data analytics among the top areas of expertise they need to build or maintain over the next three to five years. 60% are actively seeking people with expertise at the intersection of AI and cybersecurity, with another 34% assessing their requirements.
This is why guardrails cannot be separated from model choice.
As organizations introduce proprietary LLMs, open models and SLMs into the same environment, they need consistent evaluation, monitoring, access controls and governance across all of them.
The model portfolio may become more diverse.
The standards surrounding it need to become more consistent.
The first phase of enterprise Generative AI encouraged organizations to explore what models could do.
The next phase will require more deliberate choices about where and how they are used.
For business leaders, several principles stand out.
Frontier models remain valuable, but they should not become the automatic choice for every workflow. Evaluate performance, latency, control and cost against the outcome required.
Hardware, inference stacks, quantization and model routing can materially change AI economics. Cost optimization should begin during solution design.
RAG, evaluations, monitoring, guardrails and LLMOps help turn model capability into a dependable business service.
Open-source models and SLMs can provide greater flexibility, customization and efficiency where the use case supports them.
More model choice should not create fragmented standards. Security, data governance, monitoring and accountability need to apply regardless of which model sits underneath an application.
Enterprise AI is becoming more capable, but capability alone is no longer the whole story.
The bigger opportunity is learning how to apply that intelligence selectively.
Organizations that can combine powerful models with efficient infrastructure, trusted data, strong operations and intelligent model selection will be better positioned to scale AI without allowing complexity or cost to scale at the same rate.
That is where the next phase of enterprise AI value will be created.
Tenth Revolution Group helps organizations combine talent, training and transformation to build cloud, data and AI capabilities that are efficient, scalable and ready for long-term value.