The artificial intelligence (AI) landscape has seen a surge of interest in systems of execution - autonomous systems powered by agentic AI that are touted as capable of completing complex, multi-step tasks with minimal human intervention.
From managing full customer conversations to orchestrating backend processes, the promise of agentic AI is nothing short of revolutionary. Given the promise, the hype and the enthusiasm around agentic AI are at an all-time high. While game-changing, enterprises fail to fully understand the potential of agentic AI and expect it to manage most of the work, making people redundant eventually, especially in specific areas where the technology can be deployed.
But as Utkarsh Kanwat lays out in his incisive blog, Betting Against Agents (read it here), a dose of realism is warranted. His critique isn’t rooted in cynicism; it’s grounded in production-grade realities that any enterprise deploying AI at scale should consider.
Reach out to discuss this topic in depth.
Three hard truths about agentic AI
Kanwat raises three compelling challenges:
1. Error compounding across multi-step workflows
Consider a realistic example where a sophisticated AI agent has a fairly high accuracy rate of, say, 95%. In a contact center environment, where the final task might involve a multi-step execution orchestrated across different AI agents, the accuracy degrades rapidly. For instance, in a 10-step process, with 95% accuracy per step, the net accuracy falls to close to 60%, which is unacceptable in a business environment. In the Customer Experience Management (CXM) domain, where workflows span identity verification, knowledge retrieval, response generation, compliance checks, and escalation logic, such compounding errors become business-critical risks.
2. Token cost explosion
CXM involves long, memory-heavy interactions. Tracking full conversation histories across channels inflates inference costs, putting a strain on its Return on Investment (RoI) when scaled to enterprise-wide deployments. This can become a challenge for enterprises as they push for more hyper-personalized experiences through AI, wherein context retrieval and preservation become important.
3. Integration friction in real-world systems
The promise of seamless orchestration clashes with the reality of tangled tech stacks, and most enterprises don’t want to be beta testers for error-prone workflows in production. Messy Customer Relationship Management (CRM) implementations, latency-prone Application Programming Interfaces (APIs), tool-specific quirks, and complex compliance requirements all add friction, making the deployment of agentic AI far from plug-and-play. Yes, innovations such as the Model Context Protocol (MCP) protocol aim to address these gaps, but challenges still exist, especially with legacy systems.
Kanwat’s blog urges enterprise leaders to rethink blind faith in agentic AI. His critique, which centers on practical deployment roadblocks, sets the stage for a CXM-specific reality check.
The real opportunity: tool-integrated, task-focused AI agents
Rather than chasing full autonomy, forward-thinking CXM leaders are embracing domain-bounded, assistive agents – targeted applications of AI that improve efficiency without introducing unacceptable risk. These are not moonshot agents replacing entire workflows, but co-pilots enhancing specific steps. And they’re already driving value at scale. Some examples in the contact center space include:
- Call summarization
Platforms such as Dialpad and Zoom now offer real-time AI summaries of customer calls. These systems auto-tag intents, identify the next-best actions, and reduce manual after-call work, enhancing agent productivity without replacing decision-making.
- Real-time knowledge surfacing
Solutions such as NiCE Enlighten and Sprinklr embed AI agents to monitor conversations in real time, surfacing relevant knowledge base articles or policy snippets. This augments agent performance without requiring full control handoff.
- Autodraft, not autosend
Generative assistants in tools such as Intercom and Zendesk draft responses tailored to the customer and context. Human agents remain in the loop to validate and edit, striking a balance between speed and oversight.
- Post-call Quality Assurance (QA) and compliance scoring
Agentic AI is being deployed to automatically assess 100% of customer interactions against predefined compliance and quality benchmarks, something that was previously impossible at scale with manual QA processes.
- Rule-bound process automation with fallbacks
Some Business Process Outsourcing (BPO) providers are piloting AI agents that automate back-office actions such as issuing refunds or updating subscriptions, within strict boundaries and with human oversight on edge cases.
To scale these wins, enterprises need more than individual pilots, they need a modular, governed architecture that enables agents to work across systems while remaining traceable and compliant. Embedding agents into redesigned workflows, backed by clear governance and skilled talent, is key to unlocking broader CXM impact.
Everest Group takeaway: does this mean CXM will never have true agentic AI?
The future of AI in CXM isn’t about handing over the reins to fully autonomous systems of execution. Instead, it’s about embedding lightweight, modular intelligence into workflows, where AI augments agents rather than replaces them.
While the “AI agent will do it all” narrative grabs headlines, the actual enterprise wins are coming from invisible agents doing small things well, with precision, context, and human-aware control. This approach reduces friction, contains risk, and unlocks meaningful gains in both efficiency and experience. True agentic AI in CXM isn’t off the table, but the path is longer and more incremental than many assume. It will require breakthroughs in reliability, contextual reasoning, and enterprise-grade governance integration. For now, the real value lies in assistive, domain-bounded AI that supports human agents within clearly defined boundaries.
CXM decision makers should shift focus from the allure of end-to-end automation to targeted, tool-integrated deployments that enhance specific tasks. Enterprises that prioritize alignment over autonomy, deploying AI where it fits, not where it dazzles, will be the ones delivering measurable value today while building toward future capability.
If you found this blog interesting, check out our blog Why Systems Of Execution Offer The Highest Potential Return From AI | Blog – Everest Group, which delves deeper into the global services industry regarding Systems of Execution.
If you have any questions or want to discuss Systems of Execution in more depth, please contact Sharang Sharma ([email protected]) and Aishwarya Barjatya ([email protected]).
