Architecture · AI control plane

When every employee has an agent, routing is the decision.

Cisco is putting an AI agent in front of all 90,000 of its employees this month. Its finance chief described the agent's real job: pick the cheapest model that works, keep the data on premise. The model was never the decision. The layer that governs every model is.

Field briefing · HB-2026-0015 The Hikari Blue team · Operator notes from Austin 5 min read

This month Cisco puts an AI agent in front of all 90,000 of its employees. The headcount is not the signal. What the company says the agent does is.

The common read of enterprise AI is a race for access. Buy the frontier model. Give every employee a seat. Count the logins. The companies already running agents at scale have moved past that read.

What Cisco actually deployed

Cisco's chief financial officer, Mark Patterson, described the agent's first job in plain terms. It routes each request to the most efficient model available. It runs on the company's own infrastructure, for control over cost and data (Fortune, July 2026). On frontier models he was blunt. The agent is "not going to burn a whole bunch of tokens with frontier models." The same stack already drafts 80 to 90 percent of the finance team's earnings narrative.

Set that against the base rate. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027. The reasons it names are escalating cost, unclear business value, and weak risk controls (Gartner, June 2025). Model quality is not on the list.

The decision was never the model

The thesis deserves to be stated plainly.

The enterprise AI decision is no longer which model to buy. It is whether you own the layer that routes, prices, and records every model you run. Hikari Blue · operator note

Cisco named that layer without naming it. A control plane sits between the employee and the model. It reads the request. It picks the cheapest model that clears the quality bar. It caps the spend. It keeps the data inside the perimeter. It records the action. Hand every employee a frontier chatbot instead, and you buy the most expensive path to the least control. A summarization task does not need a frontier model. Paying frontier prices for it is a capital allocation error, repeated a million times a day.

This is a management problem, not an engineering one

MIT Sloan frames the answer as a continuous, life-cycle discipline. Decision boundaries, escalation protocols, and threshold values get defined before the agent runs, not after the incident (MIT Sloan Management Review, September 2025). An agent that acts for the firm needs an owner, a limit, and a record. Without those three, the CFO cannot attribute the cost, the CISO cannot bound the data exposure, and the board cannot answer a regulator.

In a bank or an insurer, that record is not optional. The same layer that controls cost is the one that produces the audit trail a supervisor will ask for. Cost governance and regulatory governance are the same wire.

So the line between leaders and laggards is not model choice. Laggards buy seats and measure adoption. Leaders build the control plane and measure three things: cost per task, the share of work served by non-frontier models, and the portion of agent actions carrying a full provenance record. Cisco's route-to-cheapest, keep-it-on-premise posture is the second pattern, in the open. The first pattern is what Gartner is counting toward cancellation.

The question to bring to the next operating review

Do not ask which model your teams use.

Ask who owns the layer that routes it, caps its cost, holds its data, and records what it did.

If the answer is a different vendor for every team, you do not have a control plane. You have a bill, and no record. The agent is the interface. The control plane is the asset.

The Hikari Blue team · Austin, July 2026

More from the Newsroom

See all articles in the Newsroom →

Map your exposure before the board does.

Thirty minutes with an operator. No slides.

Direct call with one of the partners. We listen, we structure, and we tell you what the control plane would look like (model routing, cost governance, data perimeter, action trail) for your specific stack and regulatory perimeter.

Book a 30-minute working session