Cisco's AI Agent Cost Trick: Route, Don't Reason

Cisco began rolling out a personal AI agent, MyAgent, to its roughly 90,000 employees on August 27. The headline is the scale. The detail worth your attention is the cost design: reporting on the rollout says about 50-60% of requests go to open-weight models, another 20-30% are handled by plain software automation, and only a small remainder is sent to a frontier model. In other words, the company that can afford the most expensive model for every request deliberately does not use it for most of them. If Cisco is routing, a 40-person business paying frontier-model prices to summarize a calendar invite has a budget problem it hasn't noticed yet.

The Mistake: One Model for Every Job

Most small-business agent deployments start the same way. Someone picks the best available model, wires it to email, the CRM and a spreadsheet, and lets it handle everything that comes in. It works, so nobody revisits it. But look at what those requests actually are: classify this email, pull an order status, fill a template, check whether a field is empty. The vast majority are routine. A frontier reasoning model doing routine work is like paying a senior consultant to file receipts. It gets done correctly, and you overpay for every single one.

This is also why per-token price cuts have not lowered many AI bills. When every request goes to the top tier, spend tracks volume, and volume only goes up as agents get more useful.

What Cisco's Split Actually Teaches

The three-way split is the lesson, and it maps cleanly onto any workflow:

• Tier 1, plain automation (roughly a quarter of requests): anything with a fixed rule. Status lookups, routing by keyword, data copy between systems. No model needed, so cost is close to zero and behavior is perfectly predictable.
• Tier 2, a small or open-weight model (over half of requests): classification, extraction, summarizing, drafting from a template. Cheap, fast, and good enough when the task is well bounded.
• Tier 3, a frontier model (a small remainder): ambiguous requests, multi-step planning, anything where a wrong answer is expensive. This is where you spend on purpose.

Why the Automation Tier Is the One People Skip

The interesting number is not the open-weight share, it's the 20-30% that never touches a model at all. Teams building agents tend to reach for the LLM first because it is flexible. But a request that follows a rule every time should be a rule. A workflow in n8n or Make that checks an invoice status is faster, free of hallucination risk, and costs fractions of a cent to run. Moving deterministic work out of the model also makes the agent easier to audit: fewer model decisions means fewer decisions to explain.

A Realistic Scenario

A Wizeb client, a 35-person equipment rental company, ran a single frontier-model agent across its shared inbox: quotes, availability questions, invoice queries and damage reports, roughly 1,400 messages a month. Their model bill had climbed past $2,100 a month. We logged a month of traffic and sorted it. About 30% were availability and invoice-status questions answerable directly from their booking system. Another 45% were standard quote requests and simple classifications a small model handled with the same accuracy in side-by-side tests. Only around 15% needed real judgment, mostly damage disputes and unusual multi-day bookings, and the rest were spam and auto-replies filtered out before reaching any model. After we added a rules layer, a small-model tier and an escalation path to the frontier model, the monthly bill dropped to about $640, response times on routine questions went from 40 seconds to under 10, and the damage-dispute answers were unchanged because they still went to the strongest model.

How to Build Your Own Routing Layer

1. 1Log a month of real requests. Don't guess the mix; sample a few hundred and label each as rule-based, bounded, or needs judgment.
2. 2Move rule-based work into plain automation first. It is the cheapest win and the most reliable.
3. 3Test a smaller model on the bounded tier against your current outputs, using your own examples, and keep the ones that hold up.
4. 4Define explicit escalation triggers: low confidence, high dollar value, a customer complaint, or any action that cannot be undone goes up a tier.
5. 5Keep human approval on external actions, as Cisco does, and review the routing mix monthly so drift toward the expensive tier gets noticed.

The Caveat: Routing Has Its Own Failure Modes

Routing is not free. A misrouted request, a hard question sent to a small model, can produce a confident wrong answer, and that is worse than a slow correct one. That is why the escalation rules and the monthly review matter as much as the savings. Build the safety net before you chase the last dollar, and measure cost per resolved request, not cost per token, so a cheap wrong answer never looks like a win.

How Wizeb Approaches This

Every agent Wizeb builds starts with a traffic audit and a tiered design: rules where rules work, small models for bounded tasks, and a frontier model reserved for the calls that justify it. Our AI agent development (wizeb.com/services/ai-agents) and workflow automation (wizeb.com/services/automation) teams build the routing layer, the escalation triggers and the monitoring together, so cost goes down without accuracy going with it. Cisco has the scale to build this in-house over years. You can have the same design pattern in weeks.