Feature
eBook
Designing an AI-Native Contact Center
Key Takeaways
- AI agents fail more often because of organization design than model capability.
- Multi-agent systems demand explicit process clarity most organizations don't have.
- Without observability, evaluation, permissions and escalation infrastructure, agentic systems become untrustworthy.
Multi-agent workflows are the new corporate aspiration: fleets of AI workers coordinating across systems, turning messy organizational work into clean automation.
But enterprise AI is stuck in a contradiction. Vendor demos get better every week. Models improve. Case studies promise cost savings and durable moats.
Meanwhile, most organizations can’t make the value repeatable. Pilots stall. Projects get canceled. MIT reports that 95% of generative AI pilots fail to produce repeatable value. And Gartner predicts that over 40% of agentic AI projects will get canceled by the end of 2027 due to poor cost, risk and value management. Scale remains rare.
Across operators actually embedded in these efforts, the reason is consistent: companies jump to agents before they can answer two fundamental questions:
- What does success look like?
- What needs to change to get there?
The five blockers below are what prevent orgs from defining success in operational terms.
Table of Contents
- What 'Agent' Actually Means Right Now
- Why Do Pilots Stall Before Agents Are Even Possible?
- The 5 Blockers That Stop AI Agents From Working
- A Boring Playbook That Actually Works
What 'Agent' Actually Means Right Now
An AI agent is not a chatbot or a simple automation.
In practice, an agent is a system that can perform work across steps with delegated autonomy: retrieving context, calling tools, generating artifacts, triggering downstream workflows and responding to outcomes — often with defined checkpoints for human oversight.
A multi-agent workflow is a coordinated system of these agents, each narrowly scoped to part of a larger process. A client request might be ingested by one agent, researched by another, drafted by a third and routed into production systems by a fourth.
But this autonomy comes with a cost. Agents only function inside organizations that can clearly define processes, constraints and escalation paths — conditions many teams have never needed before.
Steven Howell, AI change architect at Humble Advisory, frames it this way: “AI isn’t Salesforce. It’s databases.” Databases don’t deliver value on their own. Organizations have to redesign how work is done around them.
Common Questions About AI Agents
Why Do Pilots Stall Before Agents Are Even Possible?
Most organizations try to implement AI the way they implemented traditional SaaS. That’s a core mistake.
LLM-backed agents are not deterministic software with bounded behavior. They are probabilistic systems with broad capability and weak natural constraints. Treating them as plug-and-play tools creates risk long before it creates value.
Howell uses a historical analogy to explain the moment we’re in. In the late 1800s, textile mills adopted electricity — but saw little increase in productivity. Owners kept the same machines, layouts and workflows, simply swapping steam for electric power.
The gains from electricity only arrived after redesigning from the fundamentals: flexible layouts, individual motors, nighttime operations and localized failures rather than line-wide changes. The mistake wasn’t adopting electricity too early. It was assuming it would create value without redesigning the work around it.
AI today looks like electricity did then: awkward, expensive and frequently misunderstood as magical.
This helps explain why modern AI success stories are rare. The system design at this stage includes IP logic, handoffs, context management and verification strategies. Even if multi-agentic systems are simply “wrappers” around foundational models, the wrappers themselves deliver the majority of value. Those are the hardest-won assets, and the easiest for competitors to copy if exposed.
As McKinsey puts it, realizing agentic value requires reinventing how work gets done from the ground up. Agents and AI are a forcing function that demands an honest evaluation of what an organization actually does and what it should be doing.
Related Article: Multi-Agent AI: The Next Phase of Enterprise Automation
The 5 Blockers That Stop AI Agents From Working
Across embedded transformation efforts, the same pattern repeats. Organizations don’t fail at multi-agent workflows because of model capability. They fail because five organizational blockers make agentic systems hard to sustain in production.
Most companies hit at least two or three at once.
Blocker 1: No Shared Reality About How Work or AI Is Actually Used
Executives often say, “Our teams are already using AI.” What they usually mean is that licenses have been deployed. What remains unclear is who is proficient, how AI is being used or where it’s creating value.
Kyra Atekwana, who leads AI transformation work at Section, starts most engagements with an AI maturity deep dive designed to replace top-down narratives with bottom-up data. “People see Copilot usage and assume proficiency,” she said. “But most organizations don’t even have definitions for what proficient actually means.”
These assessments surface a consistent surprise: shadow AI.
“People say, ‘We use the company tool, but it’s behind, so I use my personal account,’ or ‘I’m routing data through workarounds because I’m being held back.’ That tells you what’s actually happening.”
The irony is that slowing down here — resisting the urge to chase quick wins — often surfaces the fastest path to impact. Nearly every organization already has someone using AI exceptionally well. Finding those people, understanding what they’re doing differently and scaling that behavior is often the shortest route to real returns.
Webinar
Feb
25
Content Strategy Leaders Live: Scaling for Speed, Complexity and AI in High Tech
A candid roundtable on how high-tech leaders are rethinking content at scale.
Register
Webinar
Feb
26
Do More with Less: Modernizing the Cloud Contact Center for 2026
Learn how to leverage cloud platforms without adding a single hire to personalize every customer interaction.
Register
Webinar
On demand
Cut the Noise: Deploying AI That Actually Moves the Needle
Learn how to turn AI experimentation into concrete revenue operations.
Watch Now
Webinar
On demand
Ditch the Desk Phones: How Modern Teams Drive AI-First Communications
Find out how one team finally pulled the plug on a legacy phone system. And built something smarter.
Watch Now
Webinar
On demand
Rebrand. Migrate. Optimize. How to Do It All (Without Slowing Down)
Cresta leveled up site speed, design flexibility and marketer sanity (in record time). Find out how.
Watch Now
Webinar
On demand
From Manual to Magical: How AI Transforms CX Teams
Learn how to replace manual support processes with automation that actually delivers.
Watch Now
Blocker 2: No Legible Workflow Processes or Documentation
Even when leaders understand who is using AI, they often don’t know how work actually gets done.
Agents don’t inherit tacit knowledge. Every assumption, boundary, input and output has to be explicit. Organizations that succeed by relying on undocumented institutional knowledge quickly encounter issues when they need to provide tangible context for agentic actions.
Howell describes two paths companies take when pursuing advanced AI adoption:
- Technology-First: Teach people how LLMs work, then reason outward.
- Process-First: Write down the workflow in painful detail with the people doing the work, then decide what should be automated, augmented or left human.
The second path “almost always wins,” meaning the best solution might often not be “use AI.”
Arjita Shrimali, AI Positioning & Sales Enablement Strategist at Signal & Scale, uses a lightweight framework to force this clarity:
- People: Who touches this workflow, end‑to‑end?
- Process: What actually happens, including systems and handoffs?
- Problems: Where are the bottlenecks and failure points?
- Data: What exists now, what’s missing and how hard will it be to get?
“This is a sanity check on ROI and feasibility,” Shrimali explained. “It keeps teams from taking a spaghetti approach — grabbing the nearest use case and hoping it sticks. You can build much faster with AI. But that makes it more important you’re building the right thing.”
Blocker 3: No Production-Grade Infrastructure
Getting to a compelling agent demo is no longer hard. Turning that demo into a trusted system is where efforts stall.
Multi-agent workflows introduce requirements most innovation teams aren’t prepared for: fine-grained permissions, observability, explicit memory management, evaluation and regression testing, escalation paths and governance frameworks. Agents can touch various disciplines, creating challenges for heavily siloed organizations.
Shrimali pointed to the standard failure mode: “Even with perfect process mapping, implementation has limits. Getting to 80% is easy. That last 10-15%? That’s the ‘Dory loop’ [the forgetful Pixar fish]. You’re constantly spinning.”
Without this last 20%, organizations can’t trust agent behavior, which can then make verifying their efforts more cumbersome than human ownership. And even if it works, if the specific handoffs and evaluations remain unclear, it’s less of a multi-agentic system and more of a magic trick.
Blocker 4: Incentives Punish Disclosure
Even when systems are sound, many agent initiatives fail because incentives are misaligned.
Employees want AI to remove low-value work. Leadership often wants AI to increase output without changing expectations, evaluation criteria or compensation. The result is predictable: employees stop sharing workflows and personal automations, and experimentation moves underground.
Atekwana described the opposite pattern with her company's internal AI usage, “It’s strongly encouraged to share what worked with AI — and what didn’t — publicly. That visibility creates momentum.”
Successful organizations make AI usage visible, rewarded and safe to share. They treat experimentation as operational work, not extracurricular risk-taking.
When AI use is rewarded with more meaningful work rather than higher quotas, value compounds instead of stalling.
Blocker 5: Leaders Outsource Judgment
This blocker alone can invalidate progress on all the others.
Executives mandate AI adoption while remaining personally distant from the tools and workflows they sponsor. They used ChatGPT “a few months ago” but haven’t used it regularly since, or see it as another hype bubble waiting to burst. The result is shallow understanding, unrealistic expectations and unclear risk ownership.
“This is for you too,” Atekwana said. “Telling people to use AI without using it yourself is incredibly damaging.”
Leaders do not need to code. But they do need lived exposure: personal workflows, real failures and the discomfort of working with systems that do not guarantee outcomes.
As Howell puts it: “You can’t outsource understanding non-deterministic systems.”
When leadership participates, scope becomes realistic, risk tolerance explicit and investment decisions improve. When they don’t, agent initiatives devolve into mandates that teams quietly work around.
Related Article: 5 Questions Every Leader Should Ask Before Building AI Solutions
A Boring Playbook That Actually Works
If you’re leading AI — or buying it — this is the least exciting path with the highest odds of success:
Baseline Reality (Blocker 1)
Run a maturity/proficiency assessment. Map who’s actually using AI, what they understand about risk/policy and where shadow AI is happening.
Make the Work Legible (Blocker 2)
Map one workflow with the people who do it. Break it into granular steps. Decide what should be AI, what should be non-AI automation and what should remain human.
Build the Substrate (Blocker 3)
Invest early in permissions, observability, evaluation, escalation paths and governance.
Align Incentives and Learning Loops (Blocker 4)
Make AI use visible, rewarded and safe to share. Scale internal wins before inventing new use cases
Require Leadership Participation (Blocker 5)
Leaders should have their own workflows and firsthand intuition for failure modes. “Everyone uses AI… except me” is not a strategy.
AI is a forcing function that exposes whether an organization actually understands how it operates. Multi-agentic systems, then, aren’t something you buy. They’re a capability you earn.