“The amount of infrastructure needed for an agent is meaningfully higher than for a chatbot. For every human you might have 10, 100 or on the aggressive side 1 000 agents… They just keep working and that consumes a chunk of [compute].” – Jeetu Patel – President and chief product officer at Cisco

Enterprise AI is colliding with a hard economic constraint: systems that act on behalf of humans, rather than simply converse with them, generate continuous demand for compute, memory, storage, networking and security that scales non-linearly with adoption.¹ What began as an experiment with a few chatbots embedded in customer support flows has evolved into fleets of autonomous or semi-autonomous agents woven into operational systems, provoking a structural rethink of infrastructure capacity planning, architecture and cost allocation.¹ As organisations discover that a single knowledge worker can be surrounded by dozens or hundreds of always-on software entities, the initial enthusiasm for ubiquitous AI is running directly into power limits, data-centre constraints and budget ceilings.¹⁰

From conversational interfaces to autonomous execution

The crucial distinction is architectural. A typical chatbot is a reactive interface: it receives a text or voice prompt, runs a model invocation, retrieves data from a knowledge base or simple API, and returns a response.^2,5,8 It behaves largely as a query layer on top of existing systems, often operating in a single-turn or short multi-turn context without independent goals, long-running state or direct write access to critical applications.^2,8,14 Infrastructure demand in this pattern is relatively bursty and predictable: each user query maps to a bounded unit of computation and network traffic, which can be statistically smoothed across a population of users.

Agents invert this relationship between interface and execution. An AI agent is typically defined as a goal-driven, autonomous system that can perceive context across multiple data sources, reason about what to do next, select tools, and execute multi-step workflows across business systems.^5,8,14,23 Rather than stopping at an answer, agents act: they update records in line-of-business applications, orchestrate workflows in CRM or ITSM platforms, trigger network changes, or coordinate with other agents to complete composite tasks.^5,8,17,21 The conversational surface, if it exists at all, becomes only one of many possible triggers. The core of the system is an execution loop that persists until the goal is completed or escalated.²⁰

This difference in behaviour drives a step change in infrastructure needs. A chatbot largely consumes compute at inference time, with minimal persistent state and shallow system integration. An agent, by contrast, is bound into the fabric of enterprise systems and may hold long-lived context, subscribe to event streams, maintain memories and embeddings, and repeatedly invoke models and tools until objectives are satisfied.^5,8,17 That shift from “answer once” to “keep working” is what turns a marginal AI workload into a structural infrastructure commitment.

Why one human can imply hundreds of agents

The claim that a single human may be associated with 10, 100 or even 1 000 agents captures an emerging pattern in enterprise design.¹ Knowledge workers are increasingly surrounded by specialised AI entities: a scheduling agent, a document drafting agent, a data-pull agent for analytics, multiple monitoring agents for infrastructure, and domain-specific agents embedded in each SaaS tool. Each of these has a narrow remit, but together they form an ecosystem of micro-work units continuously reacting to events, logs, tickets and user behaviour.^5,17,21

From an architectural perspective, this resembles the move from monolithic applications to microservices, but with an added dimension of autonomy and continuous reasoning. Instead of a single large conversational system, enterprises are decomposing tasks into networks of agents with defined roles and tool sets, sometimes coordinated by higher-level orchestrator agents.^9,21,24 That design enables agility and specialisation, yet it multiplies the number of active components that must be powered, connected, secured and observed. Each agent may appear lightweight in isolation, but at the scale of tens of thousands of employees, the number of concurrently active entities can rise by two or three orders of magnitude.

Importantly, many of these agents are not user-facing at all. Infrastructure teams are experimenting with agents that monitor network telemetry, propose or implement configuration changes, execute test suites, and validate rollouts using digital twins of production networks.²¹ Security teams are piloting detection and response agents that ingest logs, enrich alerts, and autonomously contain incidents within pre-defined guardrails.^27,28 In these environments, the mapping of “agents per human” reflects not only personal productivity tools but also operational automation embedded deep in the stack.

The continuous compute drain of agentic workloads

Because agents keep working in the background, infrastructure load is no longer driven solely by explicit user queries. Agents subscribe to message queues, listen to event buses, and schedule periodic checks; they may maintain rolling embeddings of new documents, update knowledge graphs, or re-evaluate risk scores as fresh data arrives.^9,17,21 Each step involves model invocation, data movement, or both. Even if individual actions are small, the aggregate forms a constant baseline of demand that persists 24/7.

For cloud or on-premise operators, this means the cost function moves from primarily variable, usage-driven spend towards a blend of fixed and semi-fixed commitments. If N represents the number of human users and k the average number of agents per user, the potential number of agents is k \times N. Yet the relevant quantity for infrastructure sizing is the set of concurrently active agents, which depends on trigger frequencies, task durations, and coordination patterns. A naive but illustrative view is that if each agent has an activity duty cycle \rho (fraction of time spent computing or transferring data), then expected concurrent load scales with k \times N \times \rho. Even modest values of k and \rho can generate a substantial continuous baseline when N is in the tens of thousands.

Network demand behaves similarly. Agentic systems that act across hybrid environments must traverse data-centre fabrics, campus networks and WAN links to reach telemetry sources, SaaS APIs and edge devices.^7,10,19 Jeetu Patel has described how emerging agent workloads drive persistent high-volume east-west and north-south traffic, contributing to what he terms a network supercycle as enterprises upgrade switching, routing and optical capacity to cope with AI-driven data flows.¹⁹ These patterns differ markedly from classic web or batch analytics workloads, with more continuous, many-to-many flows and tight latency expectations for closed-loop control.

Agentic AI as an infrastructure business

The strategic implication for infrastructure providers is straightforward: agents monetise hardware. Every move from static chat interactions to autonomous workflows increases the mix of workloads that are both compute-intensive and long-lived, lengthening hardware refresh cycles and justifying investment in high-performance networking, accelerators and storage.^7,10,19 Vendors positioned across the stack are racing to articulate platforms specifically designed for agentic operations, with unified views of networking, security, compute and observability that treat agents as first-class operational actors.^3,10,12,18

One expression of this is the emergence of platforms that explicitly support “humans and AI agents running critical infrastructure together” and provide a single control plane for both traditional systems and autonomous components.^3,12,18 These platforms aim to normalise agents as operational peers: they authenticate, authorise and log agent actions; expose natural-language interfaces for defining new workflows; and integrate with existing observability stacks to track performance and anomalies in agent behaviour.^3,10,12 This positioning reflects a belief that the long-term value in AI will accumulate less in standalone chat applications and more in integrated operational systems where agents co-manage infrastructure and business processes.

Economic tension: cost, usage and budget shock

For enterprises, the same dynamic poses uncomfortable questions about cost and governance. Early adoption of generative AI often focused on text chat interfaces, where spending could be bounded via rate limits, per-seat pricing and clear attribution of usage to teams.¹ The transition to agents challenges these controls. An agent that can trigger actions in ITSM, CRM or ERP systems may also quietly trigger costs: more API calls to third-party platforms, more log ingestion into SIEM tools, higher storage for generated artefacts, and, crucially, higher inference and orchestration compute.^1,21,27

Reports of organisations pulling back on expansive AI deployments due to spiralling cloud bills capture this emerging reality.¹ Once agents are embedded into daily operations, turning them off is not as simple as disabling a chatbot widget. They become entangled in workflows and SLAs. Finance teams, meanwhile, discover that AI line items are not merely “experimental” but have become semi-fixed operating costs. The ratio of spend to value becomes harder to measure when hundreds of agents operate in the background, each doing small, distributed pieces of work whose direct contribution to revenue or cost savings is difficult to isolate.

Vendors of foundation models and APIs have responded with tiered pricing, volume discounts and specialised tokens for specific use classes, but the basic arithmetic is unchanged: sustained autonomy consumes sustained resources.^1,23 Internally, CIOs and CFOs are being forced to adopt more granular cost-allocation models, tracking which business units are responsible for which agent fleets, and tying deployment approvals to explicit ROI hypotheses in terms of reduced headcount, faster cycle times, or risk reduction.^1,9,21

Why agents are harder to host than chatbots

From an engineering standpoint, agentic systems impose stricter requirements across multiple dimensions of infrastructure. First, they demand more sophisticated state management and storage, as agents need to remember prior context, plan over long horizons, and coordinate with other entities. This often implies vector databases for embeddings, graph stores for relationships between entities, and durable logs for auditability.^5,8,17,21 All of these add to storage and I/O requirements compared with a stateless chatbot backed by a simple knowledge base.

Secondly, agents need deeper integration with identity and access management. Because they can execute actions that affect real systems, they must be governed by policies defining which tools they can call, what data they can read or write, and under what approval conditions.^9,10,27,28 This adds complexity to security architecture: agents require service identities, rotating credentials, fine-grained permissions and sometimes per-action human approvals, all of which must be enforced consistently across hybrid and multi-cloud environments.

Thirdly, the network and compute layers must be designed for low-latency, high-reliability closed loops. Agent workflows that modify infrastructure or process financial transactions cannot tolerate unpredictable delays or frequent timeouts. This drives demand for high-bandwidth, low-loss fabrics inside data centres, intelligent traffic engineering across WANs, and tight coupling between observability systems and control planes so that anomalies trigger automated mitigation rather than manual tickets.^7,9,10,19 These characteristics go well beyond what is needed to serve sporadic chatbot traffic.

Power, sustainability and the physical edge of the agent boom

Behind these logical architecture concerns lies a physical constraint: power. Large-scale AI deployments are increasingly bottlenecked not by chips alone but by the availability and cost of electricity to feed data-centre clusters and edge compute nodes.^7,10,16,19 When each human user implies potentially hundreds of always-on agents, total energy consumption can rise sharply, especially if models are large or poorly optimised. Leaders in the field have warned that planning for power is a first-order requirement for CIOs considering agentic AI at scale, rather than an after-thought once use cases have been defined.^1,10,16

This constraint interacts with geography and regulation. Regions with limited grid headroom or stringent environmental policies may face harder trade-offs between expanding AI capacity and meeting sustainability commitments. Enterprises are therefore exploring techniques to bend the resource curve: model distillation and quantisation to reduce inference cost; adaptive scheduling that defers non-urgent agent activity to off-peak times; on-device or near-edge inference for local tasks; and more precise scoping of agent authority so that they do not perform unnecessary or redundant work.^9,10,21

Security and the need for agent-aware defences

Security concerns compound the infrastructure challenge. Agents that autonomously operate across networks and applications introduce new attack surfaces and failure modes. If compromised, an agent with write access to critical systems could cause damage at machine speed. Even without malicious interference, mis-aligned objectives or prompt injection attacks can lead agents to take unintended actions. Security architectures are being retooled to treat agents as high-value, high-risk entities that must be monitored, constrained and protected.^10,27,28

Several patterns are emerging here. One is the fusion of security controls directly into networking and AI infrastructure, so that traffic to and from agents can be inspected, segmented and policy-controlled without relying solely on application-level safeguards.^27,28 Another is the deployment of “security agents” that watch other agents, analysing behaviour for anomalies, enforcing guardrails and escalating suspicious activity for human review.^27,28 This meta-agent layer, however, introduces yet more continuous workload, reinforcing the original observation that agentic ecosystems increase the total infrastructure footprint.

Debates and objections: are hundreds of agents per human inevitable?

There is a live debate about whether the projected density of agents is a necessary outcome or an artefact of early design choices. Critics argue that current enthusiasm for fine-grained agents may be over-engineering: instead of dozens of micro-agents per user, organisations could converge on a smaller number of more capable, multi-role agents, reducing orchestration overhead and infrastructure load.^5,14,17 Others suggest that existing automation and rules-based workflows can absorb a substantial share of tasks without requiring full agent autonomy, using agents only at decision points where judgment and flexible reasoning are genuinely required.^5,9

There are also concerns that agent proliferation could outpace human ability to understand and govern system behaviour, leading to opaque networks of interacting entities whose collective impact is difficult to predict. In response, some practitioners advocate for stricter criteria to qualify a component as an agent: it must demonstrate clear, measurable business value, operate within narrow and auditable boundaries, and be subject to regular decommissioning reviews to prevent uncontrolled growth in the agent population.^9,17,21

Yet even more conservative designs acknowledge that the direction of travel is towards greater autonomy and deeper integration of AI into operational systems. Once organisations experience the productivity and resilience benefits of agents that can, for example, validate network changes against digital twins before deployment, or automatically draft and triage IT tickets, the pressure builds to expand their remit.^21,24 The resulting increase in sustained compute and network demand may be somewhat optimisable, but not easily reversed.

Why the distinction matters for strategy and policy

Understanding the gap between a chatbot and an agent is not a matter of terminology; it is a strategic planning issue for boards, regulators and investors. Boards need to grasp that AI initiatives framed as “assistive chat” can evolve into critical operational dependencies with recurring infrastructure costs and systemic risk profiles.^1,10,19 Regulators, particularly in sectors such as finance, health and critical infrastructure, must recognise that autonomous systems executing actions require different oversight, audit trails and safety cases compared with systems that merely answer questions.

Investors and market analysts, meanwhile, are watching how infrastructure vendors position themselves around this transition. Companies with the ability to supply not just model capabilities but also the secure, power-efficient, network-rich environments in which agent fleets can safely operate may enjoy durable demand, as agentic workloads lock in customers for multi-year infrastructure refresh cycles.^7,10,19 Conversely, enterprises that underestimate the infrastructure implications of moving from chatbots to agents may find early AI gains eroded by escalating costs and operational fragility.

The deeper story is that the move from conversational AI to agentic AI transforms AI from an application feature into an organising principle for enterprise architecture. Once software entities can act with sustained autonomy, the infrastructure beneath them becomes a strategic asset and a potential bottleneck. The observation that supporting this shift requires markedly more infrastructure than hosting simple chatbots is less a prediction than a description of a redesign already under way in large organisations worldwide.^1,7,9,10

 

References

1. “‘We created a monster’: companies rein in AI usage as costs strain budgets” – https://www.ft.com/content/1d37cc08-e0aa-45a4-a45d-4ad282529314

2. Building AI Ready Infrastructure Across APJC With Cisco – YouTube – 2025-12-08 – https://www.youtube.com/watch?v=EW-sVTms7Pk

3. AI Agent vs. Chatbot: How Businesses Can Benefit from AI … – Slack – 2025-06-24 – https://slack.com/blog/transformation/ai-agent-vs-chatbot-understanding-the-differences-and-business-impact

4. Cisco Unveils Agentic Platform for Critical IT Infrastructure – 2026-06-02 – https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m06/cisco-unveils-agentic-platform-for-operating-and-defending-critical-it-infrastructure.html

5. Thriving in the Era of AI: Essential Innovations with Cisco’s Jeetu Patel – https://www.sixfivemedia.com/content/thriving-in-the-era-of-ai-essential-innovations-with-ciscos-jeetu-patel

6. AI agent vs chatbot: what is the difference? – Make – 2026-04-28 – https://www.make.com/en/blog/AI-agent-vs-chatbot

7. Cisco Pricing 2026: Ultimate Guide for Security Products – 2025-03-07 – https://underdefense.com/industry-pricings/cisco-pricing-ultimate-guide-for-security-products/

8. Artificial Intelligence (AI) Solutions – Cisco – 2024-08-22 – https://www.cisco.com/site/us/en/solutions/artificial-intelligence/index.html

9. AI agents vs. chatbots: Capabilities, limitations, and applications – 2026-02-05 – https://rasa.com/blog/ai-agent-vs-chatbot

10. [PDF] Infrastructure for Agentic AI: Architecting Scalable Secure Multi … – https://www.ciscolive.com/c/dam/r/ciscolive/global-event/docs/2025/pdf/BRKCOM-1126.pdf

11. Navigating the Frontier of Agentic AI – Cisco Blogs – 2026-06-02 – https://blogs.cisco.com/?p=491881

12. AI Agent vs. Chatbot – What’s the Difference? – Salesforce – 2024-09-03 – https://www.salesforce.com/agentforce/ai-agent-vs-chatbot/

13. Cisco unveils agentic platform to run critical IT infrastructure – 2026-06-05 – https://futurecio.tech/cisco-unveils-agentic-platform-to-run-critical-it-infrastructure/

14. Cisco’s AI-Powered Operations with Jeetu Patel – LinkedIn – 2026-02-12 – https://www.linkedin.com/posts/cisco_agenticops-simplifies-and-accelerates-management-activity-7427741494960283648-EkVv

15. Key Differences Between AI Agents, Chatbots, and Assistants – 2026-06-16 – https://business.adobe.com/blog/differences-between-ai-agents-chatbots-and-assistants

16. Cisco unveils cloud control platform for AI agent management – 2026-06-02 – https://www.investing.com/news/company-news/cisco-unveils-cloud-control-platform-for-ai-agent-management-93CH-4722222

17. Cisco Live! 2025: Jeetu Patel Has a Vision for Agentic AI – 2025-06-16 – https://biztechmagazine.com/media/video/cisco-live-2025-jeetu-patel-has-vision-agentic-ai

18. AI Agents vs Chatbots: What Are the Differences? – Elementum AI – 2026-03-30 – https://www.elementum.ai/blog/ai-agents-vs-chatbots

19. Cisco’s AI Agent Strategy Explained: What Partners Need to Know – 2026-06-04 – https://www.youtube.com/watch?v=u0utkKqHJuU

20. Cisco: AI driving a ‘network supercycle’ – Light Reading – 2026-06-03 – https://www.lightreading.com/ai-machine-learning/cisco-ai-driving-a-network-supercycle-

21. AI Agents vs AI Chatbots: What REALLY Makes Them Different – 2026-04-28 – https://www.youtube.com/watch?v=gfXqc_wOJVI

22. Cisco: Multi-Agent AI System for Network Change Management – https://www.zenml.io/llmops-database/multi-agent-ai-system-for-network-change-management

23. Jeetu Patel – Cisco Blogs – 2026-06-02 – https://blogs.cisco.com/author/jeetupatel

24. AI Agent vs AI Chatbot: Key Differences Explained – DigitalOcean – 2024-10-10 – https://www.digitalocean.com/resources/articles/ai-agent-vs-ai-chatbot

25. Internet of Agents – Outshift | Cisco – 2026-06-02 – https://outshift.cisco.com/the-internet-of-agents

26. Cisco Announces Agent Security Solutions at RSA Conference – 2026-03-29 – https://www.linkedin.com/posts/jeetupatel_we-are-squarely-in-the-second-phase-of-ai-activity-7444140542474645505-5rNI

27. Understanding AI Agents vs. Chatbots | Microsoft Copilot – 2025-11-27 – https://www.microsoft.com/en-us/microsoft-copilot/for-individuals/do-more-with-ai/general-ai/understanding-ai-agents-vs-chatbots

28. Cisco Transforms Security for the Agentic AI Era, Further Fusing … – 2025-06-10 – https://investor.cisco.com/news/news-details/2025/Cisco-Transforms-Security-for-the-Agentic-AI-Era-Further-Fusing-Security-into-the-Network/default.aspx

29. How Cisco Protects AI Agents From the World of Cyber Threats – 2026-06-04 – https://cybermagazine.com/news/how-cisco-protects-ai-agents-from-the-world-of-cyber-threats

30. Can someone explain the real difference between an AI chatbot and … – 2025-11-12 – https://www.reddit.com/r/learnmachinelearning/comments/1ovejpu/can_someone_explain_the_real_difference_between/

31. Best AI Agent Infrastructure Platforms for Cisco SecureX – Slashdot – https://slashdot.org/software/ai-agent-infrastructure/for-cisco-securex/