Agentic AI: From Hype to the Hard Questions
Enterprise AI agent adoption is the process of embedding autonomous, tool-using AI systems into real business workflows so they can reason over organizational data, execute tasks across systems, and support or automate entire processes rather than only handling simple chat interactions.
While big tech pours billions into agentic AI deployment, the story on the ground is more complicated. Salesforce’s Futures VP Mick Costigan distills customer anxiety into three blunt questions: How good will agents become, how do we bring them into the organization, and what do humans do once they arrive? Those questions expose the real business AI barriers: capability uncertainty, messy implementation realities, and unresolved human roles. Meanwhile, AI agents are already slipping into customer touchpoints. A shopper can message a store on WhatsApp and get stock status or alternatives within seconds, showing how fast expectations are shifting. If enterprises treat agents as upgraded chatbots rather than as a new operational layer, they will drown in pilot projects that never become dependable systems.
Challenge One: The Capability Gap Between Demos and Dependable Systems
The first obstacle is not that models are weak; it is that they are progressing faster than enterprises can absorb. Costigan notes model capability feels like it is at “the beginning of an exponential rate” of improvement. Yet real-world AI agent implementation challenges lie in everything surrounding that intelligence. Agents must operate inside complex environments with trusted context, controlled access to data, and guardrails that make their outputs auditable.
Salesforce’s response is what it calls an “agentic harness”: the equivalent of brakes, steering and dashboards around the engine of an AI model. In Costigan’s own analogy, early LLMs were like an 1886 motor wagon—an engine bolted to bicycle wheels and a garden bench. An enterprise agent without a harness is the same thing: an impressive demo that no responsible leader would put on a busy road. To turn capability into organization-wide value, companies must invest as much in this harness layer—data access, context handling, safety rules—as they do in chasing the latest model release.
Challenge Two: Implementation Is an Organizational Problem, Not a Prompt
The second challenge is the brutal reality of implementation. Even highly capable agents still need to be wired into real data, tools, processes, systems of record, permissions, governance frameworks, and user interfaces. That work is slow, political and technical all at once. Costigan is blunt: getting data right, access to tools right, interfaces right, and productivity gains that scale across the organization is the hard part.
This is where the hype around enterprise AI agent adoption collides with decades of legacy systems. Enterprise agents must plug into CRMs, ERPs, identity systems and compliance workflows if they are to move beyond toy use cases. Big tech’s moves show the direction of travel. Meta’s Business Agent, announced at its Conversations conference in London, answers customer questions, qualifies leads, manages bookings and processes transactions directly in WhatsApp. That is more than a chatbot; it is a transactional bridge between attention and action. If enterprises fail to build their own integration and governance layers, they risk ceding that bridge—and with it, the customer relationship—to platforms they do not control.
Challenge Three: Redefining Human Work Before Agents Redefine It for You
The third challenge is the human question: what do people do when agents can reason, act and automate work across business functions? AI agents already show how quickly they can reshape expectations. A customer’s WhatsApp message now triggers immediate answers, inventory checks and recommendations, delivering faster responses and more seamless experiences. But inside the enterprise, roles, skills and accountabilities often remain frozen in a pre-agent world.
According to Costigan, if AI takes on more work, companies must rethink jobs, skills, workflows, accountability and the human role in decision-making. In practice, that means redesigning processes so humans set goals, handle exceptions, oversee governance and provide judgment, while agents handle repetitive orchestration. Ignoring this shift creates a vacuum where no one is clearly responsible for outcomes produced by AI—a governance risk as much as a cultural one. The agentic enterprise is not only a technical architecture; it is a new social contract between people, systems and the data that links them.
What Enterprises Should Do Now: Build Harnesses, Not Science Fiction
The temptation is to treat agentic AI deployment as a race to the flashiest demo. That is a mistake. The global market for agentic AI is projected to grow from USD 10.9 billion (approx. RM50.1 billion) in 2026 to USD 182.9 billion (approx. RM841.3 billion) by 2033. With that kind of money at stake, every vendor story will sound world-changing. Yet Costigan’s advice is more grounded: pursue near-term ROI from practical use cases while experimenting with frontier models that might justify deeper restructuring of the business.
For Salesforce, the future of enterprise AI will be shaped by the systems around models: harnesses, governance layers, workflows, partner ecosystems and platforms that make AI usable in real businesses. Meta’s shift from advertising into the transactional moment, off the back of US$200 billion (approx. RM919.6 billion) in revenue, shows how platforms want to insert AI agents directly into the customer relationship. Enterprises that do not build their own harnesses will find themselves dependent on those platforms, with little say over how work is automated or who owns the data. The path forward is clear: stop arguing about whether Skynet is coming and start building the brakes, steering and dashboards that make AI agents safe, accountable and worth trusting at scale.
