The advantage goes to organizations that can redesign work as fast as AI changes it.

CHAPTER 01

The first phase gave AI intelligence. The next gives it a role in execution

The first phase of enterprise AI was largely about access: give people capable models, connect those models to useful information, find the tasks where they help. That phase produced real gains and a familiar pattern. A product manager turns an hour of writing into ten minutes. Someone in procurement builds an agent that collects information before a supplier call. A team automates an approval that used to bounce between inboxes.

Each improvement is genuine. Follow the work from beginning to end, though, and the picture is often less dramatic. The specification is drafted faster and still arrives at the same handover. The supplier information is gathered automatically and somebody still discovers too late that a required field was missing. The approval takes seconds instead of a day, while the decision it protects stays where it has always been for reasons nobody has revisited.

The next phase is different in kind. Once software can gather context, use tools, make bounded decisions and trigger actions, the question is no longer whether an AI can perform a task. The organization has to decide how the work should be designed around that capability, and who is permitted to act on its behalf.

Analysts are converging on the same point. Gartner describes enterprise AI moving from assistive intelligence toward outcome-focused workflows, in which people increasingly supervise systems that execute within policy and identity constraints rather than performing every procedural step by hand. The useful shift in that language is from AI features to execution authority: who or what may act, under which permissions, against which systems, with what accountability. Gartner, April 2026

CHAPTER 02

Making the old process faster can preserve the wrong process

Take a product-introduction process. A customer requirement arrives, product development interprets it, sourcing starts collecting supplier information, quality checks the evidence, and eventually somebody has enough confidence to approve the next step.

The obvious AI move is to leave that process intact and make each part faster. Let AI prepare the first specification. Let an agent collect missing supplier information. Let another system check documents before the request reaches an approver.

That creates real value. It also leaves the most important assumption untouched: that the current process is still the right process.

Once AI becomes capable enough, the boxes on the process map deserve the same scrutiny as the work inside them. An approval may exist because somebody historically had to collect and interpret evidence by hand. If that evidence can now be assembled and checked continuously, the decision point may belong somewhere else entirely. Sourcing, quality and product development may each gather similar information because their systems evolved separately; a redesigned process may not need three collection moments. A handover may exist because expertise once sat with one specialist, while the future process can carry the context with the work and involve that specialist only when judgment is genuinely required.

This is the point where AI stops being a productivity layer and becomes an operating-model question. A company can become extremely efficient at executing a process that no longer makes sense.

CHAPTER 03

Redesign becomes an operating capability, not a program

Traditional process redesign is episodic. A transformation program maps the current state, defines a future state, implements the change and hands the process back. The result may then live untouched for years.

AI does not respect that rhythm, because its capabilities move faster than the transformation cycle.

A task that was sensibly manual last year may now be well suited to AI assistance. Later, the person stops performing the whole activity and instead reviews the judgment-heavy part while an AI teammate gathers the evidence. Once that happens across several connected tasks, the process moves with them. A handover disappears. Information has to arrive earlier. A control changes position. A role that used to execute the work starts managing the exceptions. The process has not become faster; its shape has changed.

Gartner calls this continuous work redesign and predicts that organizations establishing it as a core capability will be twice as likely to sustain AI-driven work transformation by 2028. Gartner, September 2026

The prediction is less interesting than the reclassification. Redesign stops being a project and becomes something an organization does as part of operating. And that only works if design and execution stop living in separate places, because execution is the evidence for the next redesign.

CHAPTER 04

The durable object is the work, not the agent

There is a strong temptation, once agents become capable, to organize the future company around the agents themselves. Create one for product information, one for supplier requests, one for quality evidence, one for reporting, then build a registry and a control plane around the growing fleet.

That may be useful infrastructure. It is a fragile operating model.

Agents will change. Models will change. Some agents will disappear entirely when a platform absorbs their capability; others will split into smaller skills. An organization that has defined itself around today's agent catalogue will keep rebuilding its operating model around the technology.

The work is more durable. A task still has an expected outcome. It still requires certain information. It still sits somewhere in a value stream. Somebody or something is responsible for it. There are dependencies, quality criteria, deadlines, approvals and evidence requirements. How that task is executed can change completely without losing the organizational meaning of the task itself.

Today it may be manual. Then assisted. Then coordinated across AI, systems and people. Eventually an approved AI teammate may perform most of it inside clearly defined limits. At Pilaro we call that movement the Skill Shift, and the point of naming it is that it should be a decision rather than a drift.

So the stable question is not "Which agent owns this?" It is "How should this work be performed now?"

CHAPTER 05

The harness around the model carries the organizational requirements

The same logic shows up in the technical architecture.

For a while, model choice dominated the conversation. As agents become operational, more of the enterprise value moves into the software surrounding the model. The Australian Signals Directorate describes this as the agentic AI harness: the layer that lets a model interact with data, tools and systems while handling security, governance, reliability and the operational realities of execution. Australian Signals Directorate, September 2026

Models are increasingly interchangeable. A better one arrives, the economics shift, a customer has a different provider policy. The organizational requirements around the work do not disappear when the model does. The harness determines what context is provided, which tools are exposed, how identity is handled, when a human must approve and what evidence is kept afterwards.

Above that sits a question no single implementation answers: once different teams run different agents, who decides which AI capabilities exist, how autonomy is delegated, how failures are handled and how the organization learns rather than leaving every team isolated. Thoughtworks calls this the organizational harness. Thoughtworks, July 2026

Gartner is grouping much of the market convergence underneath this under Business Orchestration and Automation Technologies, and warning that fragmented orchestration becomes its own operational problem as the automation estate grows. Gartner Magic Quadrant, September 2026 Gartner, September 2026

The vocabulary is still unsettled — business orchestration, agentic orchestration, control plane, execution plane, process intelligence, enterprise harness. The architectural direction is not. Organizations are accumulating people, agents, automations, APIs and AI-enabled applications that can all perform pieces of work, and somebody has to keep those pieces connected to the business outcome they are supposed to serve.

CHAPTER 06

Execution evidence is what closes the loop

The moment a redesigned way of working starts running, it produces something more valuable than status updates.

Consider the difference between knowing a task is complete and knowing how it was completed: which executor performed it, what context was available, what information was missing, where a person intervened, how long the work waited, which exception changed the path.

Most management systems sit some distance from the work they are meant to steer. A KPI moves and the organization reconstructs the explanation afterwards, by collecting updates, comparing reports and asking process owners what is happening underneath the aggregate.

The better connected execution becomes, the shorter that distance gets. A margin indicator starts moving in the wrong direction. Instead of stopping at the number, the organization can follow it into the value streams that contribute to it, see that one phase is repeatedly delayed, trace the delay to a small group of tasks and find that those tasks are all waiting on the same incomplete information source.

At that point the right action may have nothing to do with deploying another agent. It may be to fix an information requirement upstream. That is the connection between execution and steering, and it is why steering should increasingly mean noticing what deserves attention while there is still time to change the outcome.

CHAPTER 08

The next frontier is not inside the company

Almost every discussion of AI-first operating models stops at the company boundary. That boundary is where most of the remaining waste actually lives.

In physical-product value networks, the expensive failures are rarely internal inefficiencies. They are the missing declaration, the specification that meant something slightly different at the supplier, the compliance evidence that arrives after the packaging is printed, the second and third request for information somebody already sent. A supplier, a manufacturer and a retailer can each become impressively AI-first and still hand work to each other over the same wall.

The organizations that pull ahead will redesign work across that wall: shared understanding of what a task must produce, evidence that travels with the work instead of being re-requested, and AI participation that is governed on both sides of the relationship rather than improvised at each end.

That is harder than internal automation. It is also where the next order of magnitude sits.

CHAPTER 09

This is the territory of AI-first execution and steering

Seen together, the new vocabulary describes layers of one problem. Business orchestration keeps an end-to-end process moving. Agentic orchestration puts AI into that flow. Harnesses make sure it can act without bypassing identity, policy or system boundaries. Process intelligence makes the flow observable. Continuous work redesign keeps the whole thing from hardening into the next fixed machine.

The missing connection is not another layer. It is the relationship between all of these capabilities and the work they are supposed to serve.

That is the territory we describe at Pilaro as AI-first execution and steering.

The starting point is deliberately not the agent. It is the work. Pilaro begins with the tasks people actually perform, the outcomes those tasks must produce, the systems and information they depend on and the handovers between them — the real work, not the documented process. That matters because so much operational knowledge never reaches a process diagram: the local spreadsheet, the check somebody added after a problem years ago, the colleague you know to call. Redesigning from the official process alone automates the documented organization rather than the real one.

From there the work is redesigned into a value stream where people, AI teammates, systems and external participants each take the part that makes sense. Execution is not the end of that exercise; it is what tests it. Tasks arrive prepared with skill, context and quality checks, an AI Work Coach guides execution and surfaces gaps, people review the evidence and decide, and what happens during execution becomes the input for steering and for the next redesign. Pilaro capabilities How Pilaro works

Maximum autonomy is not the target. Work can stay manual, become assisted, move into orchestrated execution, or eventually be delegated when the information, governance and risk profile support it. The important part is that the shift is intentional and visible rather than the byproduct of an employee finding a better tool.

And the execution layer belongs around the systems an organization already runs, not in place of them. ERP, PLM, PIM, quality and planning systems stay authoritative for their records. What has been missing is the layer connecting the work moving across them to people, AI participation and the business outcome.

CHAPTER 10

The compounding advantage is learning speed

Enterprise AI will keep producing impressive adoption numbers: more agents in production, more tasks automated, more employees using models daily. Those numbers tell us the technology is spreading. They say much less about whether an organization is becoming better at operating with it.

An organization can automate a task that should no longer exist. It can accelerate one department and create waiting in the next. It can deploy sophisticated agents on top of information too poor to support reliable execution.

The durable capability is learning how the operating model should change, and being able to change it again next month. That requires an unglamorous discipline: seeing the real work rather than the documented process, making explicit choices about where people, AI and systems participate, keeping enough evidence to know what actually happened, and connecting those patterns back to the objectives the business already cares about.

Do that and the improvements compound. Every run of the process teaches you something about the next one. The gap between noticing a problem and redesigning around it shrinks from quarters to weeks. That is a different kind of advantage from being 30% faster at drafting documents, and it is much harder for a competitor to copy, because it is not a tool they can buy.

The first phase of enterprise AI gave people access to a new kind of intelligence. The current wave is giving that intelligence tools, memory, identity and the ability to act. What follows is an organizational question: where all of that capability belongs inside the way a business actually operates, and across the network it operates within.

The companies that answer it well will not simply have more AI. They will have a better mechanism for changing how work gets done as AI changes what is possible.

You do not need a transformation program to find out whether that is true. You need one important process, the people who actually perform it, and the willingness to let what happens during execution tell you what to change next.

Ask Marit. She already knows which process it should be.