Useful work per dollar — what delivery and analytics should change first

The question

If agentic systems are becoming the unit of work, what should a delivery or analytics lead actually change first?

What the signal is really saying

How to manage AI investments in the agentic era (OpenAI Blog) is easy to skim as another enterprise AI post. The useful idea underneath is sharper: measure useful work per dollar (or per hour of scarce human attention), not activity.

That lands for me because I have already lived the activity trap in analytics. More extracts. More tabs. More tiles. Meetings that reconcile numbers instead of deciding what to do. Agents will recreate that failure mode faster, because they can generate plausible work without improving the call.

What people get wrong

"Add agents" is being sold the way "add a dashboard" used to be sold — as if the tool is the operating model. Teams then track tokens, runs, tickets closed, and demo velocity. None of those answers the only question that matters in delivery or portfolio work: did a decision get better?

Capability is not the scarce resource anymore. Judgment, ownership, and rollback are.

Change the definition of done first

Before you change the model, change what "done" means for an agentic loop:

1. What decision or artifact improved? A scoped analysis, a release risk call, a refreshed metric with an owner — not "the agent ran."
2. Who owns the failure when it is wrong? If the answer is "the platform," you do not have an operating model.
3. What would falsify the hype next week? An eval slice, a golden set, a Monday exception list — something that can fail a release.

Without those, you have expensive autocomplete with a project plan.

This is the same instinct as a credit portfolio review. I do not want thirty charts. I want three decisions: is emerging risk accelerating, where is the engine drifting from the model, what changed that we did not expect. Everything else links out.

The analytics parallel (and why it is not a metaphor)

When two teams disagree on "delinquency," a dashboard becomes a debate club. When two teams disagree on what an agent was allowed to decide, an automated workflow becomes the same debate with a worse audit trail.

So the first artifacts I want before the exciting demo are boring on purpose:

• A one-page scope: what the agent may draft vs what a human must sign
• An eval slice: ten real cases where wrong is costly
• A kill switch: how we disable the loop without a war room

Delivery leads should refuse agent workflows that cannot name a rollback. Analytics leads should refuse agent outputs that cannot name a definition.

How this shows up in what I am building

On Orbit — Portfolio & Radar, the same pressure appears in miniature. Signals refresh daily. Drafts can be generated. None of that is useful unless it ends in a publish decision with an owner — Approve in Slack, commit the Write, redeploy. Automation without a gate is just faster noise.

I am treating adjacent pieces like The US is advancing AI safety through state and federal action as context, not a second thesis. Governance matters; the operating lesson for builders is still ownership, scope, and the ability to stop.

Mistakes I refuse to repeat

• Optimizing for demo wow before the metric has a definition doc
• Letting "useful work per dollar" become another vanity KPI with no owner
• Automating the meeting before the exception list is stable
• Shipping a loop I cannot disable cleanly

What I am not doing yet

Chasing every agent framework. Replacing human review on anything that touches risk language or production change. Treating a vendor narrative as a roadmap.

Takeaway

If agents are becoming a unit of work, do not start with the agent. Start with the unit of useful work: the decision, the owner, the eval, and the rollback. That is how you keep Monday honest — whether the artifact is a Power BI tile or an automated draft waiting for Approve.