↓ Skip to main content

From AI Pilots to an Operating Model for Agents

C-level corner · Executive brief

Your organisation already uses AI agents, mostly ones employees found for themselves. What is missing is a way to turn experiments into governed services. The fastest route is not more agents: it is one purpose-built agent for your most painful workflow, built with identity, limits and testing from day one, as the template for the next.

The business problem
#

Every enterprise is experimenting with AI agents. Few are in production. The familiar pattern: a general-purpose assistant is connected to a handful of tools, pointed at company data, impresses in a demo, then fails the move to production. In the technical article’s analysis, the cause is rarely the model. Pilots fail on four things:

  • Security. Each connection logs in on its own terms, not as the user, and the audit trail is scattered across systems.
  • Reliability. The same request can produce different steps on different runs.
  • Compliance. Nobody can easily show which data the agent touched and on whose authority.
  • Cost. Overhead that is invisible in a pilot becomes a line item at scale.

Meanwhile employees already use AI assistants on their own, outside any governance: the article calls shadow AI the fastest-growing category in every enterprise. And stalling has a price: Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, with inadequate risk controls one of the three reasons it gives.

What changes
#

Treat agents as an operating model change, not a technology project. The question shifts from “how do we connect an agent to everything?” to “how do we build an agent that does this one thing reliably, securely and at scale?”

  • One job per agent. A reconciliation agent, an audit agent, a supply chain risk monitor, each with instructions, tools and limits designed for that job alone.
  • Capabilities fixed in advance. What is not in the agent’s code cannot be run: a technical constraint, not a hope that the model behaves.
  • Run like any other production workload. Its own identity, access limited by the rules you already enforce, limits on where it can send data, and one audit trail.
  • Tested before it ships. Each agent must pass a test suite for its business scenarios before release.

A simple maturity scale helps: Level 1 is individuals experimenting with no governance, Level 2 an approved list of tools, Level 3 business-critical workflows run by purpose-built agents with full identity, access control, audit and evaluation, Level 4 many such agents working together under governance the platform enforces. In the article’s view, most organisations sit at Level 1 or 2.

What it’s worth
#

  • Speed to value. In the article’s view, one purpose-built agent for your most painful workflow, deployed with proper identity, governance and evaluation, delivers more than a hundred generic agents running on developers’ machines.
  • Faster approval for the next project. Build identity, limits and audit once and each new agent reuses them, so the security review becomes a checklist, not an investigation. That is the argument of Keep Shipping.
  • Cost you can see. Sending all AI traffic through one central gateway lets you attribute spend to each agent, team and workflow, set quotas and flag anomalies. AI spend becomes a governed operating cost, not an opaque line on the cloud bill.
  • Lower risk. Fewer connections and credentials handed around mean fewer places to attack, and one audit trail answers the regulator: who asked, what was done, which data was used.

Decisions for leadership
#

  1. Choose the one workflow that hurts most, and fund a purpose-built agent for it instead of a broad assistant rollout.
  2. Set the bar for production: every agent has its own identity, a fixed set of tools, one audit trail and a test suite it must pass before release.
  3. Bring shadow AI into the open: take stock of the assistants and agents already in use, and decide which move onto an approved path.
  4. Invest in the internal skill to build agents, not only in buying tools. I’d treat that capability as the asset that compounds.

Questions to ask your team
#

  • Which of our agents are in production, and which have been in pilot for more than a quarter?
  • For each agent: on whose authority does it act, and what can it reach?
  • Can we list every tool an agent can use, and could that list change without a review?
  • Where would we look to answer “what did this agent do last Tuesday?”

Technical background, for your architects

Figures come from the linked articles and their sources; check them against your own situation. Snowflake-related articles explore the art of the possible and are not official Snowflake solutions or endorsed by Snowflake.

Share this article