Every major technological shift begins with experimentation: A new capability emerges, companies rush to adopt it, and for a while, activity is mistaken for progress. The number of tools deployed, licenses purchased, or people using the technology becomes the easiest way to demonstrate momentum.

We are entering that phase with AI agents.

Agents will fundamentally change how companies operate. They will not simply help people complete existing tasks faster. They will increasingly perform entire units of work, collaborate with humans and other agents, and become a meaningful part of how organizations deliver value.

However, deploying agents is the easy part, the harder and much more important question is whether they are producing work that actually matters and how to optimise it.

The future of work is agentic

I have spent most of my career building and scaling technology companies. Every time a company grows, the same thing happens: the organization becomes harder to understand.

More people create more coordination. More systems create more fragmentation. And agents, will be no different. Actually, I strongly believe that the agentic workforce will amplify this dramatically.

A single person may soon manage the output of multiple agents and those agents may generate more work in a few hours than a traditional team previously produced in days. They will operate across different tools, models, vendors, and workflows. Some work will be performed entirely by humans, some collaboratively with AI, and some autonomously by agents.

This does not make humans less important. On the contrary, it makes understanding the relationship between human judgment and agent execution much more important and critical than ever.

The future organization will need to determine which work should remain human, which should be augmented, and which can become fully agentic.

Adoption tells you almost nothing

Today, most companies measure their AI transformation through adoption.

They track how many employees use AI, how many licenses have been activated, how many prompts have been submitted, or how much agent-generated work has been produced.

These numbers may tell you that something is happening but they do not tell you whether it is working.

An agent can produce an enormous amount of output while creating very little value. It can complete a task quickly but introduce defects that require days of human intervention. It can reduce the time spent producing work while dramatically increasing the burden of reviewing it. It can appear inexpensive until rework, coordination, and quality costs are taken into account.

When production becomes nearly unlimited, volume stops being a useful definition of productivity.

The questions that matter are different:

• What does a unit of agent work cost?

• What meaningful impact is it producing?

• How does its quality compare with human work?

• How often does it need to be corrected?

• Which agents perform best for which types of work?

• Where is human involvement adding the most value?

Without objective answers, companies cannot build an agentic strategy.

Before transforming the organization, learn from it

There is enormous pressure on leaders to develop an AI strategy immediately. But the most valuable first step may be much simpler: do nothing, and learn.

This means resisting the temptation to redesign the organization before understanding what is already happening inside it, and some of the most important questions you should be answering are:

• Which teams are improving?

• Which use cases create measurable impact?

• Where is quality deteriorating?

• Which models work best for which tasks?

• What should become autonomous, and where is human judgment still essential?

This is the data with which you should shape the transformation. The transformation should not be built on assumptions.

This is especially important because every organization will arrive at a different answer. There will be no universal ratio of humans to agents and no single model that performs best across every function. The right operating model will depend on the work, the context, the quality required, and the economics of producing it.

What the agentic workforce is missing

Every important organizational resource eventually gets a system of record.

Companies have systems of record for customers, employees, financial transactions, and product activity. These systems create a shared, trusted understanding of what is happening and make it possible to manage the organization accordingly.

The agentic workforce does not (yet) have one: Signals are scattered across models, AI assistants, agent platforms, development tools, project systems, and vendor dashboards. Each system reports its own activity, usually through the metrics most favorable to that particular tool.

What companies need is an independent layer that brings human, assistant, and agent work together and evaluates them against the same outcomes:

• That means separating human output from agent output without losing the relationship between them.

• It means attributing quality issues and rework to the work that caused them.

• It means comparing cost, impact, complexity, and quality across teams, agents, models, and vendors. A

• nd it means understanding not only what was produced, but whether it moved the organization forward.

This is the infrastructure we have been building at Pensero.

We started in engineering because software is where the agentic transformation is happening first and fastest. Our platform evaluates the magnitude and complexity of completed work—effectively creating objective story points after the fact—and connects that work to delivery, quality, cost, and collaboration.

That foundation makes it possible to measure agents side by side, understand how they work together with humans, and identify the combination that produces the best outcomes.

Once this is clear, as a company you will be able to answer the critical questions that will define the next decade:

• Which work should become agentic?

• Which agents should be deployed where?

• What should remain human?

• When does adding another agent increase output, and when does it simply create more review work?

• Which vendor delivers the best quality-adjusted return?

• Where does the organization need better models, better orchestration, or better human judgment?

In an agentic world, understanding what that intelligence is actually accomplishing will be the key differentiator to succeed.