3 min read

In Age of Empires you don't lose by attacking late, you lose by advancing an Age without an economy

Before moving an AI agent to production, check five things (trace, evaluation, cost, owner and shutdown), because Gartner predicts that more than 40% of agent projects will be canceled before the end of 2027 due to costs, unclear value and inadequate risk controls.

In Age of Empires you don't advance to the next Age because you feel like it. You need the buildings of the Age you're in and the resources in the bank. Until that exists, the button is off, and you don't argue about it with anyone.

In product there's no button. You can say you're in production whenever you feel like it.

The four Ages of an agent

With agents, there are also four Ages.

The four Ages of an AI agent
AgeWhat it looks like
DarkIt runs on your machine and you're the one demoing it.
FeudalOne real use case, one team, people who know you.
CastleSeveral use cases, used by people who didn't build it.
ImperialThe organization runs on it every day and it can't be turned off without warning.

The mistake isn't attacking early. That's a decision, and you own it. The mistake is declaring yourself in the Imperial Age with Dark Age villagers: you announce the Age and don't build the economy that supports it.

What you pay for

In June 2025, Gartner said that more than 40% of AI agent projects will be canceled before the end of 2027. The three reasons it gives aren't technical:

  • Costs that spiral.
  • Unclear business value.
  • Inadequate risk controls.

All three are economics. And the distance between saying and having is huge.

  • 75%of leaders say they've already adopted agentsForrester, June 2026
  • 17%of organizations have agents deployedGartner CIO Survey, 2026
  • 11%have agentic systems ready for productionDeloitte, December 2025

Three different measurements, the same year.

A real Age advance

On August 5, 2026, Salesforce announced that the U.S. Department of Defense authorized Agentforce 360 at Impact Level 5. The first deployment is the Army Human Resources Command, which handles more than 1,500 cases per day.

They couldn't before, and it wasn't because of the product. An authorization was missing. That's literally the game's greyed-out button: you don't advance an Age until the controls exist.

The Age-advance checklist

Five boxes, before going to production.

  1. Trace

    You can reconstruct what the agent did in a specific run, step by step.

    Passes if you can answer why it did that without guessing.

  2. Evaluation

    There's a set of cases with expected answers that runs before every change.

    Passes if someone can break it on purpose and the set catches it.

  3. Cost

    You know what happens to the bill if usage goes up tenfold.

    Passes if it's a number, not a look of surprise.

  4. Owner

    There's a person who answers when it fails on a Sunday.

    Passes if they have a name.

  5. Shutdown

    You know how to turn it off and what breaks if you do.

    Passes if it's already been turned off once on purpose.

The trap

Rushing is a valid strategy. Lying to yourself about rushing isn't.

What doesn't hold up is rushing while convinced you're booming.

Frequently asked questions

What checklist should I run before moving an AI agent to production?
Five boxes: trace (reconstruct what it did in a run), evaluation (a set of cases that runs before every change), cost (know what happens to the bill if usage goes up tenfold), owner (a named person who answers when it fails) and shutdown (know how to turn it off and what breaks).
How many boxes can fail before I'm not ready?
If three of the five are missing, you're not ready to advance an Age. You can still go out as a bet, but you have to say it out loud.
What does Gartner say about AI agent projects?
That more than 40% will be canceled before the end of 2027, because of costs that spiral, unclear business value and inadequate risk controls.
How many people really have agents in production?
According to three measurements from the same year, 75% of leaders say they've adopted agents, 17% of organizations have them deployed, and 11% have production-ready systems.
Is going to production fast a mistake?
No. Attacking early is a valid decision if you make it knowingly. The mistake is believing you're in an advanced Age without having built what supports it.

Sources

Let's talk

If you have an agent that's about to go to production and you want to review it with the checklist, tell me how it looks today.