A once-overlooked layer of the artificial intelligence market is attracting significant investment: the infrastructure that sits between organisations and the AI models they use.

AI routers and gateways provide access to models from multiple providers and help businesses route requests among them based on factors such as cost, speed, reliability and task complexity. A routine request might be routed to a smaller, cheaper model, while a more demanding task could be directed to a more powerful frontier model. Sensitive information might be restricted to approved systems, while other models might be appropriate for coding, language or industry-specific tasks.

Recent transactions underline the growing value of technology that helps organisations manage those choices. Palo Alto Networks completed its acquisition of AI gateway company Portkey in May, and NVIDIA acquired GPU orchestration software provider Run:ai in 2024. The transactions involve different businesses and different layers of the technology stack, but point to a common development: as enterprise AI environments become more complex, companies are placing greater value on the tools that coordinate and govern them.

But the issue is not only which tools an organisation can access, according to Professor George Shinkle, School of Management and Governance at UNSW Business School, and Minnie Singh-Murphy, Executive Manager, Westpac AI Team, and a former academic at AGSM @ UNSW Business School. It’s also whether people, workflows and systems can keep learning as those tools change, and they say organisations will need to build that capability for adaptation and strategic resilience alongside their AI systems.

“The challenge here is: how do we get people and organisations to learn and adapt faster?” Prof. Shinkle said. “It’s about how we make the systems more resilient to deal with this pace of change, which we anticipate is going to continue at a very high rate.”

More choice does not always mean more flexibility

The appeal of a multi-model approach is clear. AI models are developing quickly; their capabilities vary, and their costs can vary greatly. But more choice can also make it harder for organisations to build expertise and improve the way they work.

“Traditionally, organisations take an action, then they improve it and get more efficient – they are learning in that process,” Prof. Shinkle said. “But if you constantly change what the actions are, you disturb the learning process, and the organisation can get more confused rather than more efficient,” he added. “It doesn’t necessarily create flexibility because the confusion impinges on what you might believe would be flexibility.”

Learn more: AI is changing work – now, work has to change

Ms Singh-Murphy suggested the answer lies in making learning portable. When judgement, evaluation habits and the lessons of experimentation sit with people and shared practice rather than inside a particular tool, an organisation can change models without starting again. The challenge is to work out where flexibility is valuable and where the organisation needs to commit. An organisation might benefit from being able to change models for a particular task, for example, but changing the underlying system too frequently may undermine the learning that makes it more effective.

Artificial intelligence may reduce the cost of keeping some technology options open, but organisations still need to commit to the foundations that allow those options to be used effectively. These include data quality, cybersecurity, evaluation processes, workflow integration, governance and a workforce capable of continuous learning. Moving early can also generate valuable organisational learning, as companies learn where AI creates value, where their data and processes are weak, what employees need and which governance arrangements work in practice.

The benefit comes from making a bounded commitment, learning from it and retaining the ability to revise course as the evidence changes. Prof. Shinkle describes this as a form of flexible commitment. “The real challenge with organisations in this space is being resilient and adaptable, changeable to a world that is fast-paced,” he said. “The issue is: how do we keep up with it?”

The organisation chooses the boundaries

In a large organisation, employees may interact with an AI-enabled system rather than select an individual model. Ms Singh-Murphy described the harness as the software layer that connects the user and the model to approved data, tools, and capabilities, while applying the organisation’s configured permissions and controls.

This can allow a common system to support different roles and tasks. An employee using AI for summarisation may require different models, tools, permissions and controls from someone using it in a software development workflow. “The value is not simply in providing access to more models. It lies in enabling the organisation to match capabilities to particular roles and tasks within clearly defined security, data and governance boundaries,” Ms Singh-Murphy said.

Domain-level expertise remains essential, she added. Access to a more capable model does not, by itself, enable someone to frame the right problem, assess the quality of the output or apply it responsibly in context. “Powerful tools do not eliminate the need for domain expertise. When users lack the knowledge required to interpret, challenge and apply the outputs, even sophisticated tools may be used ineffectively or inappropriately,” she said.

The implication is that learning should be matched to the role as closely as access is. The capability that ultimately determines value sits with the people framing the problem and judging the output, not with the model itself.

This places limits on the idea that AI can simply automate technology selection. The system may be able to decide which model is best suited to a task, but the organisation still needs to define the task, the acceptable risks, and the outcomes it aims to achieve.

The economics of the right tool

The most capable AI tool will not always be the most appropriate one. As with any organisational change, leaders implementing AI-related changes must weigh costs and capabilities alongside the intended task, information sensitivity, operational risk, and the consequences of error.

Ms Singh-Murphy said model selection should consider the economics of the intended use case, rather than assuming that the most capable frontier model will always deliver the greatest value. Model selection is, therefore, not simply a question of cost and performance. It is also a matter of security, privacy, governance and accountability.

Subscribe to BusinessThink for the latest research, analysis and insights from UNSW Business School

Ms Singh-Murphy said access to AI capabilities should be proportionate to the user’s role, the intended task and the sensitivity of the information involved. In an enterprise environment, capability should be introduced within approved security, privacy and governance boundaries and supported by role-specific learning that develops alongside access.

Growing investment in routing and gateway technology reflects this focus on matching capability to roles, and such technology may help organisations put those choices into practice. Prof. Shinkle said the appeal of a router is that it can help organisations make decisions that are becoming too dynamic for traditional processes to handle easily.

The basic decision process has not changed; organisations still need to assess technology against their objectives, constraints and risks. But the speed of change has made that assessment more difficult. “Validating what you’re doing, which has been the traditional organisational process, becomes really difficult, because things are moving so quickly,” Prof. Shinkle said. 

“If a human’s doing the validation, then by the time you get the analysis complete, there’s frequently a new software release out that changes things. It’s that pace of change where AI routers can help with those decision processes. But you still need humans to decide where you want to push the organisation’s objectives, and where you are willing to accept the risk.”

Learning is the capability that lasts

The ability to work across multiple models also creates new demands for employees and leaders.

The first is foundational AI literacy, Ms Singh-Murphy said: everyone needs a working knowledge of what AI systems can do, where they may be unreliable and how their outputs should be assessed.

It should combine conceptual understanding of AI beyond generative AI tools with appropriately governed practical experience, enabling people to understand how AI systems work, evaluate their outputs and recognise their limitations. This requirement extends to executives sponsoring AI initiatives. Leaders need sufficient technical understanding to question proposed uses, test assumptions, recognise limitations and exercise informed oversight.

The second is role-specific capability. Just as access to AI should be matched to the role, so should the learning that comes with it, grounded in the tasks, data and risks people work with.

The third is continuous learning. Formal programs alone cannot keep pace with the release cycle of AI models, so learning also needs to happen in the flow of work, through short refreshers, shared practice and the lessons of experimentation.

"AI touches every part of an organisation in terms of systems, workflows and capability"

MINNIE SINGH-MURPHY

Ms Singh-Murphy said technology deployment alone is insufficient. Organisations also need people who are open to learning and resilient, as AI systems, underpinning technologies, and generative AI tools change, able to evaluate alternative approaches and prepared to revise their practices as the evidence and technology evolve.

That makes change management an important part of AI adoption. Organisations may be introducing systems whose capabilities, costs and risks shift before employees have fully adjusted to the previous version.

Both Prof. Shinkle and Ms Singh-Murphy agreed that resilience is a key underlying concept in managing the AI transition – resilient systems, workflows and, especially, people. The organisations best placed to adapt will be those with employees who are curious, capable of learning quickly and able to apply judgement in unfamiliar situations.

Experimentation needs guardrails

The need to learn quickly does not remove the need for control.

One way to enable experimentation is to provide employees with access to AI tools in a secure environment via sandboxes. But even sandboxes can pose risks if their boundaries are poorly designed, as the July attack, in which AI agents being tested by OpenAI escaped their evaluation sandbox and breached Hugging Face’s systems, showed. Sandboxes also need to do more than contain risk. Their value lies in turning what one team learns into knowledge that the whole organisation can use.

Learn more: The strategic impact of AI on business transformation

Ms Singh-Murphy said organisations need to build responsible use by design into the way AI is introduced, with foundational AI literacy for employees and sponsoring leaders, and ongoing learning linked to access. Effective education, she said, depends on credible expertise across the technology, responsible use and role-specific capability.

She also highlighted the importance of transparency and appropriate explainability, particularly where AI informs consequential decisions that must be understood, reviewed or defended.

Ms Singh-Murphy said experimentation and responsible AI should not be treated as competing objectives. An innovation mindset encourages curiosity, experimentation and lifelong learning. A responsible approach to AI integrates critical judgement into that mindset, considering not only what is possible, but what is appropriate, under what conditions and with what consequences.

Prof. Shinkle has been working on the idea of “minimum viable policies” – a limited set of clear rules that protect the organisation while leaving room for employees and AI-supported processes to adapt. “The critical thing is for organisations to decide where they will need to have very tight control – let’s say, because of regulations – and then to do these things perfectly,” he said. “You put in a tighter governance system and allow less experimentation there, and then you can allow more in other places.”

Building systems that can change

Personal data, proprietary information and regulation can all limit where a new system can be used. Organisations also need workflows, governance arrangements and people capable of making use of the technology.

Ms Singh-Murphy said the systems and workflows around AI need to remain iterative, with the organisation learning and adjusting as it goes. Access and education should develop together, rather than every employee simply being given the most capable system available.

“AI touches every part of an organisation in terms of systems, workflows and capability. A substantially better AI model is a small part of that equation,” she said. “Conceptually, with technologies such as a router, you may simplify that strategic choice within the organisation, and then it may become a dynamic capability.” Such technologies, however, do not remove the need for leaders to establish the objectives, selection criteria, controls and accountability that govern AI decisions.

But routers and gateways may create a new form of dependence: an organisation may become less reliant on a particular model while becoming more reliant on the system that chooses among them. Routing rules, security policies, monitoring data and evaluation histories may eventually become embedded in the AI system. Leaders, therefore, need to understand what would happen if they had to move those functions elsewhere, and how easily important workloads could bypass the system.

Learn more: Tomorrow's decisions need yesterday's wisdom and today's AI

This is where learning becomes a strategic safeguard. An organisation that understands why its routing rules, evaluation criteria and controls exist can carry that knowledge to a new system. One that has simply handed those choices to the system cannot.

Ms Singh-Murphy said the strategic goal is to invest in capabilities that will outlast any single model and in people's capacity to keep learning, while leaving enough room to adapt as technology changes. In her view, the advantage will go to the organisations that invest in education to use AI well, rather than simply those with the broadest access to the technology. 

As Prof. Shinkle concluded: “AI can inform choices; people need to own the strategic commitments and consequences.”

In summary and next steps

Organisations that pair AI access with role-specific learning, clear governance boundaries and an understanding of why their routing rules exist may be better placed to adapt than those that rely on access alone, according to Prof. Shinkle, who provided a number of considerations for industry practitioners:

1. Boards and CEOs. Directors and executives may benefit from building enough technical understanding to question proposed AI uses, while retaining responsibility for the objectives and risk appetite that guide them.
2. CIOs and CTOs. Technology leaders could consider making bounded commitments that allow for revision, including an assessment of how routing rules, monitoring data and evaluation histories would transfer if the organisation changed systems.
3. CHROs and L&D leaders. People and learning leaders may find value in developing AI access and education together, spanning foundational AI literacy, role-specific capability and continuous learning in the flow of work.
4. Risk, compliance and legal teams. These teams could consider a limited set of clear rules that apply closer control where regulation requires it, with transparency and explainability where AI informs consequential decisions.
5. Finance and procurement leaders. These leaders may wish to assess AI model selection against the economics of each use case, as the most capable model will not always deliver the greatest value.
6. Innovation and team leaders. Team leaders could consider how sandbox boundaries are designed and how lessons from experimentation might be shared, so that learning in one team can inform practice across the organisation.

Minnie Singh-Murphy’s contribution reflects her personal professional perspective. The views expressed are her own and do not necessarily represent the views or positions of Westpac Group.