Why Thinking Must Move Beyond AI Agents
If you were to believe the hype, AI agents are an end game in themselves – offering the promise of autonomous workers, digital employees and a new organisational layer capable of radically improving processes and delivering multiple business benefits. While there is some validity in this view of AI capabilities, industry observers believe the hype cycle is accelerating faster than many organisations’ ability to implement the systems needed to make these assumptions reality.
The agent delusion, mistaking capability for readiness
AI vendors promise autonomy. Organisations buy the idea. Once agents are integrated with existing processes and systems, reality bites. This highlights the error many organisations make – agents aren’t the revolution and will only highlight deficiencies in existing processes unless fundamental structural change takes place.
Large numbers of businesses have reported experimenting with AI agents, but very few are seeing tangible benefits. The RAND Corporation reports suggestions that more than 80% of AI projects fail, around twice the number reported for comparable non-AI IT projects. The findings, drawn from qualitative interviews with data scientists and engineers, shouldn’t be taken as a precise percentage, but can be taken as evidence of a high failure rate.
Four critical elements for AI adoption success
RAND assigns the blame to multiple issues: misaligned purpose ensures that leaders and technical teams don’t necessarily agree on the problem the project is meant to solve. This ensures that success is rarely clearly predefined and accountability never lands. Inadequate or poorly structured data is also cited. Organisations significantly underestimate the quality, access rights and governance discipline AI agents require. For many organisations, this only become apparent once resources have already been committed.
The importance of infrastructure and integration is also often overlooked. Organisations find that progressing an AI project from a promising demo to a production-ready system proves to be more difficult and costly than previously estimated. And there’s also the technological blind spot to consider. Too many organisations focus on the technology, not the problem, meaning they chase the latest tools and systems, not the business outcome they hope will be delivered.
What this highlights is that it’s system inefficiencies that lie behind the lack of agent effectiveness, not intelligence. Intelligence alone can’t produce satisfactory outcomes, yet vendors continue to sell the promise of autonomy without informing organisations that existing inconsistent workflows, systems and process designs need to be significantly improved or re-engineered to take full advantage of the technology.
It’s about systems, not agents
Agents can do all sorts of clever things: browse, book, compare and transact – but they need logical structure to be effective. Without a logical structure in place, even the smartest technologies fall short. Commercial websites provide a good example of where poor structure undermines both accessibility and usefulness. Research discussed in W3C Semantics for the Agentic Web workshops suggests more than 70% of websites fail to support reliable AI agent navigation due to illogical layouts missing semantic landmarks and poor HTML structure. Stanford HAI research adds weight to this observation, finding properly constructed and accessible websites achieve around 78% agent task success.
In enterprise applications, structural and data integrity (or more pertinently the absence of it) creates even bigger problems. A PEX Report states 52% of professionals cite data quality and availability as a significant barrier to AI adoption. Gartner also predicts that 40% of AI projects may be cancelled by 2027 due to inherent operational inefficiencies and a predicted poor ROI on current AI investments.
Industry blind spots
Agents aren’t intelligence, they’re execution nodes. They aren’t strategy either, they’re more analogous to plumbing. Without the right pipes – machine-readable workflows, stable interfaces, governed data flows, clear ownership and predictable execution environments – systems become blocked and unusable.
Some organisations have got the plumbing right: 57% deploy agents for multi-stage workflows and 16% for cross-functional processes, according to Accenture. Their Pulse of Change study also found that 81% of European business leaders plan to tackle more complex, system-level use cases in 2026. This illustrates that when the correct foundations are in place, organisation recognise that AI adoption can deliver significant cross-organisational benefits.
What is system-level readiness?
Yet, without the correct structure, AI adoption will always struggle to make the desired impact. System-level readiness needs to incorporate a robust semantic structure that delivers predictable, machine-interpretable environments populated by data that’s consistent, effectively governed and agent actionable. Process clarity is critical. Organisations must ensure that processes contain no hidden steps, no departmental knowledge silos, and that employees define operations and systems that work under pressure, not just in a demo situation. If the system is correctly configured, an AI implementation has a better chance of being successful; if it isn’t, an AI agent will simply amplify the system’s inadequacies.
A hypothetical use case
An enterprise deploys agents to automate procurement. The agents fail spectacularly due to inconsistent data fields, undocumented exceptions and fragmented workflows. After redesigning and standardising the systems and putting in place effective information governance, the agents are suddenly achieving the expected outcomes. In such a case, the agents didn’t get smarter – it’s simply the case that the system, and more importantly the information contained within it, stopped being stupid!
Less hype, more architecture is what the industry is crying out for
Agents themselves aren’t the future: they’re the catalyst for a restructuring of the systems and process improvements needed to get the very best value out of them. Only when systems themselves are intelligently structured will true autonomous capability be realised. Enterprises need to stop fetishising AI and start building the systems to allow it to work effectively. The winners won’t be the ones deploying the most agents, they’ll be the ones with the most coherent systems in place.
–
John Samuel, Founder, Seventh State.