The AI arms race is rigged: demos win, operations lose
Enterprise AI adoption failures arise when organizations buy headline-grabbing agents and copilots that automate only a small slice of work while ignoring the harder 95 percent of data, workflow, and change needed to make those agents reliable at scale.
The gap between vendor AI promises and agentic AI reality is no longer a rumor; it is now admitted in public by vendors themselves. In hospitality software, one provider recently described the current market as an “AI adoption arms race” and highlighted a viral case where a company accidentally spent about USD 500 million (approx. RM2,300,000,000) on AI tokens in a single month, while another example showed Uber’s leadership conceding that AI costs were outrunning productivity gains. Those are not outliers; they are the visible edge of a pattern where AI implementation costs and failure modes sit off the balance sheet until they explode. Yet the slide decks keep centering glossy agent demos. When the same vendor then acknowledges that the agent is only 5 percent of the work, the entire marketing narrative starts to look like misdirection rather than guidance.
Agentic AI reality: the 95/5 problem vendors do not price
The most important truth in today’s enterprise AI story is this: the chatbot, automation, or AR overlay is the easy 5 percent; the unified data, context, and governance foundation is the hard 95 percent that makes or breaks value.
In a recent hospitality AI session, an innovation strategist framed an “Architecture of Intent” built on three pillars: clear intent, a unified data layer, and deep understanding of each property’s rules and workflows. Even with those in place, he stressed, an AI agent still is not built; the agentic layer sits on top and accounts for only a small fraction of the effort. This is the blunt agentic AI reality: demos show the 5 percent, the 95 percent is invisible infrastructure work that announcement-driven coverage rarely examines. Meanwhile, leaders operate in an environment “where artificial intelligence is everywhere, but meaning is nowhere,” saddled with fragmented systems and disconnected data that increase friction rather than reduce it. Unless buyers price that 95 percent explicitly, they all but guarantee that AI implementation costs will outstrip benefits long before any promised automation appears.
Hidden AI implementation costs: when failures cost more than features
If AI-led marketing were honest, every pitch would start with the failure budget. Instead, enterprises discover the real AI implementation costs only after the damage shows up in token bills, outages, and customer churn.
One hospitality leader cataloged a Pizza Hut franchisee whose AI-driven delivery platform collapsed operations and destroyed more than USD 100 million (approx. RM460,000,000) in business and enterprise value because the system ignored the reality of a DoorDash-dependent operation and went live without proper adaptation, training, or support. His verdict was uncompromising: “AI that works in theory but fails in practice is still a failure, and someone has to pay for it no matter how steep the cost.” Another quote worthy of a procurement dashboard is: “Adoption failures are common and expensive.” Add the USD 500 million (approx. RM2,300,000,000) token fiasco and Uber’s admission that AI costs are outpacing productivity, and a pattern emerges: enterprises are burning more value on badly scoped AI than they gain from the features themselves. The real scandal is that almost none of this risk appears in glossy vendor diagrams.
Where enterprise AI starts to work: embedded, industry-specific workflows
The few bright spots in enterprise AI are not generic copilots; they are deeply embedded, industry-specific AI workflows that grind away at repetitive, high-volume tasks like invoice intake and operating cost settlement.
A property management ERP provider offers a useful counterpoint to the arms race. Its SAP S/4HANA-based strategy for housing and real estate weaves AI directly into invoice intake, incoming mail processing, operating cost settlement, and day-to-day property operations, while acknowledging that customers still face the hard work of ERP migration, data quality, and cloud architecture. At its 2026 summit, the firm’s CEO framed early AI investments as a response to skilled labor shortages, arguing that customers gain when AI is integrated directly into the property management system; a headline announcement was an AI-optimized operating cost settlement process that “significantly” simplifies work for housing companies. Its AI assistants—branded for housing, property managers, and commercial real estate—are designed to automate routines and make data usable inside the ERP rather than in a detached layer. This is what credible industry-specific AI workflows look like: boring, repetitive, and grounded in transactional systems.
Why embedding beats bolt-ons—and what enterprises should do next
The core lesson from both hospitality and property ERP is that real enterprise AI success comes from embedding intelligence into existing workflows and data models, not from bolting on a generic system-of-action layer that sits above the ERP and hopes to orchestrate everything from afar.
In property ERP, the AI portfolio rides on an SAP S/4HANA template built from more than two decades of sector experience, with deployment options that include private cloud; meanwhile, organizations are racing the 2027 end of mainstream support for ERP 6.x, knowing that a move to SAP S/4HANA secures maintenance until 2040. Research cited in this context shows how AI hype is already warping ERP strategies: 55 percent of organizations have deployed SAP S/4HANA or its cloud variant, yet only 34 percent have completed the transition, and 43 percent now cite AI announcements as the primary external factor shaping ERP decisions, ahead of the looming maintenance deadline at 39 percent. The right response is not more demos; it is blunt due diligence. Rewrite AI evaluation criteria around the 95 percent foundation, price the failure modes before a line of code is shipped, and demand industry-specific workflows, not abstract agents.
