AI-Native Workflow Transformation for Enterprise Renewals at Scale

cisco · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise customer-renewal workflows

Cisco needed its renewals agentic system to move beyond accurate chatbot answers into embedded workflows that could support a large recurring-revenue customer-experience organization.

Constraints:

• 18,000 to 20,000 person CX organization
• recurring-revenue renewal workflows
• adoption plateau despite high answer accuracy
• LLM planner nondeterminism
• business workflow personalization

Architecture

Cisco evolved renewals AI from chatbot to teammate by combining LangGraph-style routing, domain-specific subgraphs, deterministic scheduled tasks, long-term memory, confidence-based stops, and human teams that help software complete renewal outcomes.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  supervisor-graph["Supervisor graph"]
  business-planner["Business planner"]
  domain-specific-subgraphs["Domain-specific subgraphs"]
  execution-agents["Execution agents"]
  deterministic-scheduled-tasks["Deterministic scheduled tasks"]
  long-term-memory["Long-term memory"]
  confidence-handoff["Confidence handoff"]
  human-renewal-teams["Human renewal teams"]
  supervisor-graph --> domain-specific-subgraphs
  supervisor-graph --> business-planner
  business-planner --> deterministic-scheduled-tasks
  business-planner --> domain-specific-subgraphs
  domain-specific-subgraphs --> execution-agents
  deterministic-scheduled-tasks --> execution-agents
  execution-agents --> long-term-memory
  execution-agents --> confidence-handoff
  confidence-handoff --> human-renewal-teams
  human-renewal-teams --> supervisor-graph

Components

Data flow

Renewal work flows through a supervisor graph into domain-specific subgraphs, scheduled tasks, memory-backed execution agents, and confidence-based handoffs.

Why this architecture

The graph and deterministic-task design respond to planner drift and the adoption plateau that appeared after accurate chatbot answers were not enough.

Models used: LLM planner · domain-specific agent models

Scale

• Cisco describes an adopted renewals agent with 95% accuracy, concurrent use, and many monthly queries before expanding to the broader CX organization.
  “We have 95% accuracy, a lot of users concurrent, many queries a month, all the right things. But then one thing happened.”transcript, stated
• The first system had 95% answer accuracy, concurrent users, and many monthly queries before adoption plateaued.
  “We have 95% accuracy, a lot of users concurrent, many queries a month, all the right things. But then one thing happened.”transcript, stated
• The renewals workflow ultimately expanded to the whole Customer Experience organization.
  “Why? And we expanded to renews to the whole CX organization.”transcript, stated
• The target organization size is an 18,000 to 20,000 person CX organization.
  “customer experience inside Cisco is an organization about 18 to 20 to 20,000 people depending how you look at this.”transcript, stated

Alternatives considered (roads not taken)

LLM planner chooses every workflow step freely

Rejected because: Cisco constrained required workflow steps with deterministic tasks because LLM planners got creative and skipped or changed steps over repeated runs.

Optimizing for accurate chatbot answers alone

Rejected because: Cisco found accurate answers drove initial adoption but plateaued without deeper workflow embedding.

Adding AI to broken existing workflows

Rejected because: Cisco warns that AI can make broken workflows fail faster rather than fix the underlying process.

Patterns

multi-agent-orchestration (orchestration)

A supervisor routes to specialized renewal, sentiment, adoption, and other domain subgraphs instead of relying on one generic agent.

confidence-thresholding (safety)

AI-native workflows execute autonomously by default but stop for human judgment when confidence scores trigger intervention.

agent-memory (context)

The teammate model includes long-term memory as part of moving from chatbot answers to delegated, personalized workflow execution.

durable-execution (orchestration)

For recurring business workflows, Cisco replaced creative planner behavior with deterministic scheduled tasks that run prescribed steps.

Tradeoffs

When to let an LLM planner decide workflow steps.

Chose deterministic scheduled tasks for prescribed business workflows over letting the planner creatively infer every recurring workflow.

Rationale: The team uses LLM reasoning where useful but pins down business-critical recurring tasks to avoid hallucinated process variation.

Whether to augment existing workflows or redesign around outcomes.

Chose AI-native workflow redesign over bolting AI onto each existing workflow step.

Rationale: Cisco reframed the goal from helping humans operate old software to letting humans provide judgment while software executes outcomes.

Gotchas (what bit them)

High answer accuracy did not guarantee adoption.

Root cause: The renewals agent was optional and outside users' daily workflow, so people used it a few times and returned to spreadsheets.

Resolution: Consolidate tools, use management-led adoption pressure, and evolve the system into an embedded teammate rather than another chatbot.

LLM planners drift away from required business steps.

Root cause: The speaker says models can creatively alter prescribed step sequences across repeated runs.

Resolution: Use deterministic scheduled tasks for workflows where the business needs exact recurring behavior.

Adding AI to a broken workflow can make the failure happen faster.

Root cause: AI augmentation preserves the flawed workflow structure unless the team redesigns around outcomes.

Resolution: Recompose AI-native workflows and evaluate success with ROI-driven business metrics.

Evaluation

Cisco reports answer accuracy and adoption behavior, then shifts evaluation emphasis toward embedded workflow adoption and business outcomes.

Metrics: 95% accuracy · many concurrent users and queries per month · adoption plateau despite accuracy

Outcomes

• Cisco expanded from a renewals chatbot to a teammate-style agentic workflow system across Customer Experience.
• The main reported lesson is that embedded workflow value matters more than answer accuracy alone.

Conference talk by Cisco architect; some adoption and revenue-scale claims are high-level rather than audited metrics.