Most organisations have spent two years convinced that AI performance comes down to the quality of the question. Ask smarter prompts, and AI generates better results. That logic drove the surge in prompt engineering courses, certifications, and consulting engagements that swept through corporate budgets from 2023 onwards.

But times are changing. Gartner forecasts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. Gartner has also predicted that over 40% of those agentic AI projects will be cancelled by 2027, citing escalating costs, unclear business value, and inadequate risk controls. The common thread running through these challenges is poor context management: agents fed fragmented data, deployed on broken workflows, and operating without the institutional knowledge they need to act well.

This recognition is driving a structural shift in how leading organisations think about AI. London Business School (LBS) identifies context engineering as the defining competitive discipline of the current moment: the deliberate design of the information, structure, and governance that allow AI systems to perform reliably in a specific organisational setting. In 2026, as LBS puts it, the battle will be about who has the richest and most defensible context for their AI model. 

Why prompt engineering is no longer enough

Amit Joshi, Professor of AI and Strategy at IMD, watched this pattern play out and reached a clear conclusion: the whole premise was the wrong problem to solve. "A couple of years ago, there was this massive focus on prompt engineering. People were thinking this is going to be the key that really unlocks the secrets of generative AI,” he recalls.

Since then, a large number of organisations have shifted from simply using web-based tools to embedding them within their organisations. “The moment you set something up inside your own sandbox, all of a sudden, you control the data that this thing is accessing. The dependence on the prompt significantly reduces, because you now have a model that has access to your very specific contextual data, which kind of overwhelms the simple quality of the prompt you might need to generate a good answer,” he says.

The second shift, Prof. Joshi argues, is that the models themselves have matured. He draws a parallel to the evolution of internet search. Twenty-five years ago, an effective Google search required Boolean syntax, inverted commas, and upper-case operators. "What do we do today?” he asks.

“This thing has gotten so good that even if you make a spelling error in your Google search, it'll give you the correct search result. Something like this is happening with prompt engineering as well. The LLMs themselves have gotten better, so they don't need the same level of handholding that they did even two years ago. Which is not to say that a quality, well-thought-through, detailed contextual prompt is useless, but the value it brings has significantly shrunk."

Why AI agents are breaking down inside organisations

Prof. Joshi sees two failure modes playing out across organisations at every level of maturity. The first is the most basic.

"They just don't have enough context, which is a fancy way of saying that their data pipelines, data cleanliness, and data access are just not up to speed. If that is not up to speed, or it's not accessible, or it's not clean, it's going to be the classic issue, either of not working, or it's going to be garbage in, garbage out,” he says. “That's issue number one.”

The second failure mode is more insidious, because it afflicts organisations that have already addressed their data problems. "Even with more sophisticated organisations that do have access to the data, what tends to happen is there is a tendency to essentially take AI, take agents, and slap them on existing processes. The processes that we use in our organisations are obviously not designed for AI. AI might actually generate efficiency within each task in the process, but the process remains the same. So the net gains are often either incredibly small or don't deliver the ROI,” Prof. Joshi explains.

Learn more: How AI is changing work and boosting economic productivity

His warning to organisations that believe they have their data in order is pointed. "If you don't have quality data, if it's not clean, if it's not accessible, you might think it's there, but if it's sitting on 17 different systems that don't speak to one another, don't show up to the party. Because what's going to happen is you might make a lot of other investments, and a little too late down the line, you might figure out that you just don't have enough data to actually make this work."

The system problem behind agentic AI failure

Yenni Tim, Associate Professor in the School of Information Systems and Technology Management at UNSW Business School, identifies the same breakdown from a systems perspective. The problem, she argues, is organisational fragmentation, and it surfaces in agentic deployments as what she terms "context fragmentation" and "context rot."

"Isolated systems managed by different teams, ambiguous relationships between data and the systems consuming it, divergent configurations: all of that makes effective agent implementation harder before any AI model question enters the picture,” she says. “The governance gaps sitting on top of these fragmented systems aren’t new, but they’re now becoming a real constraint. Some legacy systems are hard to see into, the architecture doesn’t always line up, and agents are adding new governance requirements that the current setup just isn’t built to handle."

The downstream consequence is what she calls the "pilot-to-production gap". Small, focused experiments with agents work in isolation, and without system-level support, they do not scale. As Director of the UNSW PRAxIS Lab, she has tested this directly, comparing one higher-capability agent given broader access and authority against multiple lower-capability agents working alongside human staff on the same internal IT support function.

An interesting finding was that the mixed configuration resolved tickets more efficiently and more successfully. The reason? Context. “The less capable agents were embedded inside the existing review workflow, so the humans around them carried the institutional knowledge that the agents needed to resolve a ticket well, things like past tickets, team conventions, and system quirks. The single more capable agent had broader autonomy but no good way to access that context. Capability did not close the coordination gap,” she says. “The right context did."

From prompt engineering to context engineering: what changes operationally

Prompt engineering is, at its core, an individual skill, as it optimises a single interaction. Context engineering operates at a different level of the organisation entirely. A/Prof. Tim identifies three dimensions that shape context engineering at the system level: technological, informational, and social. The work for business leaders is to think about how those three dimensions combine into a coherent digital representation of the organisation.

"Of course, the technology needs to be in place," she said. "But how are the informational flows feeding the technologies? What do the retrieval, update, and evaluation pipelines look like? Are people supported to work in the required workflows and environments? Prompt engineering optimises one interaction. Context engineering shapes the conditions for all of them."

"What organisations really need to realise is what part of their data is going to be useful for decision making, what part of their data is truly proprietary"

AMIT JOSHI

Prof. Joshi frames the operational shift in terms of where control now sits. When organisations were using public chatbots, the only lever they had was the quality of the prompt. Once a model is embedded inside the organisation's own environment and given access to its own data, the equation changes. "You have a model that has access to the outside internet data, but it also has access to your very specific contextual data, which kind of overwhelms the simple quality of a prompt you might need to generate a good answer,” he explains.

Context as a defensible strategic asset

If context engineering is the new competitive frontier, the question for executives is what makes it a durable source of advantage. A/Prof. Tim's perspective focuses on treating context engineering as an organisational capacity, and she identifies three components that must work together. 

The first is assets: infrastructure, data, and tooling. The second is abilities: people with system-level understanding who can act as architects across the technological, informational, and social layers. The third is what she calls "actualisation": the workflows, ownership, and accountability routines that put assets and abilities to work day to day.

"Accumulating more data, I think, is the easy part and the least defensible. The core is in the combination. Competitors can buy similar infrastructure, but they can't easily replicate the architectural choices, the cross-domain people, and the routines through which the organisation curates and renews its context."

Learn more: Why boards are asking the wrong question about AI strategy

Prof. Joshi approaches the same question through the lens of strategic resource theory. Proprietary data, he argues, only creates genuine value if it passes a basic set of tests: is it copyable, is it accessible to competitors, is it available from public sources?

"What organisations really need to realise is what part of their data is going to be useful for decision-making, what part of their data is truly proprietary. What will not be available to competition, what is not available in public sources and may not be available for a reasonable time in the future, is not copyable,” he says. “The classic tests that we assign to strategic resources need to be applied to your data to understand how much strategic value it's going to add."

He also cautions against the trap of accumulating volume for its own sake. "Companies need to be wary of simply adding more data and pretending that's context. This needs to be relevant and clean. Even with the best LLM tools today, it can't be too big, because sometimes even the really sophisticated LLMs get confused as to which part of the data to pull from,” he observes.

Morgan Stanley is one firm that put this principle into practice. Rather than point an assistant at the open internet, its wealth division curated hundreds of thousands of pages of its own research and used retrieval to ground the tool in that material, with each use case tested before release, a person reviewing the output, and the data held under a zero-retention arrangement. As a result, more than 98 per cent of adviser teams adopted the assistant, and document access rose from 20 to 80 per cent.

Context-as-a-service: commodity layer or strategic core?

The emerging "context-as-a-service" model has attracted attention as a potential new layer of the enterprise technology stack. Both professors see context-as-a-service filling a genuine market need, particularly for organisations that lack the internal capability or resources to build context infrastructure from scratch. Prof. Joshi identifies small and medium businesses, as well as organisations in heavily regulated or traditionally non-technology sectors, as the most likely early adopters.

"For firms that have the capability and the resources to create the context in-house and completely build that stack in-house, they will choose to do that,” he says. “But context-as-a-service might absolutely be an interesting business model for those who, for regulatory reasons, for strategic reasons, industry reasons, have never really gotten their hands dirty with this kind of stuff. I do think context-as-a-service will have enough of a market to survive."

He also expects context-as-a-service providers to operate independently from the large language model companies. "The Anthropics, the OpenAIs of the world, I suspect they might choose not to get into this, because they would say, our strategic focus ought to be on providing the intelligence, and getting the context to plug into that intelligence, we're going to leave to you."

For organisations with the resources to build in-house, A/Prof. Tim's view is that delegation carries real strategic risk. Proprietary workflows, execution models, and institutional knowledge are only visible in day-to-day operations, and it is these that agent systems need to perform well over time. "To benefit from AI agents in the long run, companies should see context capacity as their own agency. The learning loops and the governance of agents over time will shape the organisation's whole AI trajectory, and not all of that can be delegated,” she suggests.

Moving from experimentation to production

McKinsey research puts the scale of the challenge in sharp relief. While 62% of organisations are experimenting with AI agents, only 23% are scaling them, and in no single business function have more than 10% of organisations achieved scaled deployment. The gap between experimentation and production is exactly the terrain both professors focus on in their work with organisations.

For Prof. Joshi, the starting point is an honest internal mapping of where data lives, how clean it is, and how accessible it actually is. The second step is strategic clarity. "The issue I often see with companies is the classic 'let a thousand flowers bloom' kind of philosophy, which in many ways is lazy, because it doesn't force you to prioritise,” he says.

“The problem is that you've got a thousand little POCs (proofs of concept) and experiments, none of which are scaling, all of which are costing money, but none of which is really getting an ROI for your organisation. So once you know your data landscape, you need a North Star: what are a few narrow areas in which, as an organisation, I'm going to focus? Is it mainly about internal efficiency? Is it mainly about customer experience? Is it going to be pricing or supply chain? Have a North Star, and then within those North Stars have very, very narrow projects that you can start developing, and for those, you look for the right data and the appropriate tools."

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

A/Prof. Tim's approach starts with the same discipline: identify where in the organisation an agent would genuinely add value, then build the scaffolding around that choice. She defines scaffolding as accountability structures, processes, and technologies for curating, governing, and evolving AI agent use cases over time.

"Organisations need to ask: where does the relevant context live? How is it kept current? Who governs it? What workflows surround the agent? What outcomes do we measure? Many frameworks are emerging, and most have something to offer. But just like agents need proper context to perform, the unique organisational context of why agents are being implemented here, in this firm, for this work, also matters,” she explains. “The leaders pulling ahead are the ones asking that question first: what's our context?"

The organisations that do this work now will not need to catch up later, predicts A/Prof. Tim, who notes that the next wave of AI advantage will belong to those who have built the data discipline, the governance routines, and the cross-domain people that give their agents the context to act well.