Skip to content
WOSB · HUBZone · MBE Certified UEI E2XCPB9DPCF4 CAGE 9SEC5

Beyond Methodology · Blog

Why AI Demands a New Organizational Operating Model

October 6, 2026 · 15 min read · Robin Sims-Allen

Why AI Demands a New Organizational Operating Model

Most organizations no longer have an AI adoption problem.

They have an AI value problem. The tools are already everywhere. Employees use copilots to draft documents and summarize meetings. Marketing teams generate content in minutes. Engineers ship code with coding agents. Customer service teams test virtual assistants that never sleep. On paper, the company looks modern.
But when the board asks a simple question, the answers get vague. Where is the money? Which costs actually came down? Which processes now run differently? Why do so many promising pilots still sit where they were a year ago?

The examples are striking. One small sustainability venture with four people serves 20 enterprise clients. One marketplace moved from 50 deals per adviser to running 3,000 at once, not by hiring 60 times the staff but by building agent analysts for the volume work while people handle the high-stakes conversations.

Those are not efficiency gains.
They are different business models.
And they are only possible because the organization was built to use AI, not merely to tolerate it.

The shift is not one change but six<p>at once, touching how work, teams, decisions, governance, knowledge, and talent are organized.
Today’s operating models were built for a different kind of work
Every operating model encodes an assumption about who does the work.

The industrial era built functional hierarchies because work was physical, repetitive, and needed tight supervision. The digital era introduced cross-functional and agile teams because software made faster iteration possible. In both cases, the structure assumed that people performed nearly every step of the process.

That assumption is now breaking.
AI does not just digitize a step. It can read documents, classify requests, draft responses, write code, recommend actions, and increasingly carry out multi-step tasks on its own. In other words, it can do work that used to require human judgment at every point.
McKinsey describes this shift as the largest organizational paradigm change since the industrial and digital revolutions. It calls the emerging model the agentic organization, where people and AI agents work side by side at scale, at near-zero marginal cost.
That is a different category of change from previous technology waves. When companies adopted cloud computing or mobile apps, the org chart stayed recognizable. Decision rights stayed where they were. The technology served the structure. With AI, the structure has to serve the technology.

If an agent can complete in minutes a task that once moved between four teams, the four handoffs are no longer a process. They are friction. If a small team can supervise hundreds of agent-run tasks, traditional spans of control stop making sense. If a model can recommend a decision instantly, a weekly approval committee becomes the slowest part of the system.

Those are operating model questions, not IT questions.
Pilot purgatory is an organizational symptom
Many companies are stuck in what people now call pilot purgatory.

The proof of concept works. The demo impresses leadership. A small team reports encouraging results. Then the initiative needs production data access, a security review, integration with core systems, a change to how frontline teams work, and a budget line that no single department wants to own.
That is usually where momentum fades.
It is tempting to blame the technology. The model was not accurate enough. The vendor was not mature enough. The data was not clean enough. Sometimes that is true.

More often, the pilot was never designed to survive contact with the organization.
BCG’s research supports this. Its 2024 study of more than 1,000 senior executives found that around 74 percent of companies were struggling to achieve and scale value from AI. It also found that roughly 70 percent of implementation challenges came from people and process issues, about 20 percent from technology, and only around 10 percent from the algorithms. That finding became BCG’s 10-20-70 rule: put roughly 10 percent of effort into algorithms, 20 percent into technology and data, and 70 percent into people and processes.
Most companies spend their energy the other way around.

They debate model choice, vendor selection, and platform architecture in great detail. Then they treat process redesign, role changes, and adoption as tasks for later.
Later rarely arrives with enough funding or attention.
The result is a growing portfolio of clever experiments and a stubbornly unchanged organization.
Value comes from redesigning work, not inserting tools
The most consistent finding across recent AI research is also the least glamorous.
The organizations that capture value redesign workflows. The ones that do not mostly insert AI into the workflows they already have.
McKinsey’s 2026 survey found that nearly three-quarters of AI high performers report fundamentally redesigning workflows because of AI, up from 55 percent the year before. Among all other respondents, the figure is about one-quarter.
That difference matters more than any model benchmark.

Inserting AI into an existing process usually produces local gains. A report gets written faster. A reply is drafted sooner. Code ships more quickly. Those improvements are real, but they often disappear into the surrounding process. The request still waits in the same queue. The approval still needs the same three signatures. The handoff still happens by email. Redesigning the workflow asks a different question.
If AI can do this part of the work reliably, what should the whole process look like now?
That question often reveals that entire steps can be removed, that checks can happen once instead of three times, or that work no longer needs to pass through four teams before anyone acts on it.

McKinsey’s September 2026 research on operating models adds an important nuance. It found that organizations using AI to automate existing workflows, without pushing for reinvention, actually report lower clarity and weaker skills than companies at an earlier stage of adoption. Automation without redesign can make an organization more confused, not less.
The lesson is uncomfortable but clear.

A single AI use case inside an unchanged process is a faster cog in a slow machine.
When agents become teammates, roles and decision rights must be redrawn
In AI-native companies, the language around AI has already shifted.
Leaders no longer talk only about copilots for personal productivity. They talk about agents with defined responsibilities, escalation paths, and places on shared task boards. McKinsey’s research describes one company where a single product idea submitted by an employee is picked up by ten specialized agents covering product definition, technical feasibility, revenue recognition, and legal compliance, producing a full requirements package within hours.

That is not a tool. That is a team structure.
And team structures need clear answers to old questions.
Who decides.
Who checks.
Who is accountable when something goes wrong.

In many traditional organizations, those answers become blurry once AI is involved. The business assumes technology owns the model. Technology assumes the business owns the process. Risk assumes it will review things before launch. Operations assumes the tool will simply work. Each assumption is reasonable. Together, they leave a gap exactly where accountability should be.

This is also where the phrase human in the loop often becomes too vague to be useful.
A human in the loop only works if that person has the time, context, authority, and skill to challenge the machine. If someone is asked to approve hundreds of AI outputs a day with no real ability to investigate them, the loop is decorative. It satisfies a policy without protecting the business.

McKinsey’s work on the agentic organization suggests people increasingly move above the loop rather than sitting in it. Instead of completing every step, they set goals, supervise agent teams, manage trade-offs, and own outcomes.
That requires deliberate design.

McKinsey’s operating model research found that organizations reinventing work with AI favor standardized processes with clear protocols and metrics, precisely because clarity makes it easier to define where decisions sit and where human judgment is required. The same organizations push decision rights down, so more choices can be made without escalation.
More clarity about where decisions belong.
More authority close to the work to actually make them. Trust has to come before autonomy Most organizations handle new technology with a familiar pattern.
The business moves first. Governance arrives later. Then everything slows down while controls are retrofitted onto something that was never designed to carry them.
With AI, that pattern is especially costly.

AI-native companies take a different approach. They treat autonomy as something a system earns, not something it is granted on day one. McKinsey’s research on AI-native companies describes this as progressive autonomy: AI generates, people judge, and the system is given more freedom only when it has proven it can be trusted in that specific context.
One founder put it simply. Run the process manually until the pain forces automation, then automate one step at a time.
The same research shows why this matters. One company let an email agent handle a contract negotiation thread, and the agent immediately drafted a reply accepting changes that actually needed more negotiation. A person caught it before it went out. The company switched the feature off at once.
That is the real lesson about AI governance.
It is not a final approval gate. It is a design discipline that runs through the whole workflow.

Each step up gives AI more room to act, but only after it has proven reliable at the step below. People stay accountable at every level.
In practice, that means risk, legal, security, and compliance shape a use case from the start rather than reviewing it a week before launch. It means clear tiers of oversight, so a low-risk internal summarization tool does not face the same process as a customer-facing decision. It means defining in advance where human approval is mandatory and encoding it into the workflow. And it means monitoring after go-live, because AI systems can drift in ways traditional software does not.

McKinsey’s agentic organization research goes a step further. It argues that governance can no longer be a periodic, paper-heavy exercise. As agents work continuously, governance needs to become real time and embedded in the work, with people holding final accountability.
Governance bolted on at the end slows everything down.

Governance built into the operating model is what allows speed to last.
The bottleneck is knowledge, not the model
Many AI debates inside companies revolve around which model to use.
AI-native organizations spend far less time on that question. They have learned that the ceiling on their AI is usually set by what their agents can find, not by how clever the model is.

McKinsey’s research on AI-native companies makes this point directly. When an agent cannot answer a question, the cause is often that the answer was never written down or sits somewhere the agent cannot reach. The real gap is knowledge infrastructure: unrecorded meetings, scattered documents, and expertise that lives only in people’s heads.
That turns data and knowledge into an operating model question.

Who owns the source of truth for each type of information? Who decides what an agent is allowed to read? Who is responsible when an agent acts on outdated content? Who funds the cleanup work that benefits five departments but sits in none of their budgets?
Without clear answers, data governance becomes a permanent bottleneck.
AI-native companies tend to answer those questions pragmatically. Rather than waiting years for a perfect single data lake, many build lightweight connectors that make information queryable wherever it already lives, whether in shared documents, messaging threads, or meeting transcripts. They record and transcribe meetings by default. They treat their internal knowledge base as a product that needs to stay fresh.

The payoff is real. In one example from McKinsey’s research, a knowledge agent that indexes code, internal documentation, and team conversations lets new hires get up to speed in days.
The risk is real too. A stale knowledge layer leads agents to give confident but outdated answers, and user trust erodes quickly.
That is why knowledge ownership cannot be left to chance. In an AI-native organization, it is part of how the company is run.
Centralize the platform, decentralize the work. One of the first structural questions every company faces is where AI should sit. Should there be a central AI team that builds and governs everything? Should each business unit run its own efforts? Or should the organization land somewhere in between?

There is a case for early centralization. In a McKinsey review of gen AI use at 16 large institutions, about 70 percent of organizations with highly centralized operating models had moved use cases into production, compared with only about 30 percent of those with a fully decentralized approach. Scarce talent, early architecture choices, and risk standards are easier to manage from one place.
But AI-native companies show the limits of that model.

McKinsey’s interviews with AI-native leaders found a consistent pattern. A central platform team owns governed access to models, architecture, security guardrails, and shared infrastructure. Business teams own the problems to be solved and how to solve them. As one leader put it, it makes no sense for one person to oversee ten business units they do not understand.
The same research describes two opposite failure modes.

Decentralization without a platform leads to chaos, with unapproved tools and shadow automation that later have to be shut down. Centralization without focus leads to overreach, such as trying to build one agent that can do everything instead of shipping smaller, specific solutions.

The answer to both is the same: a real platform, governed centrally, with clear boundaries, that lets business teams build freely on top of it.
That is a hybrid answer.
In complex organizations, it usually is.
The question is not centralized versus decentralized. It is deciding which parts of the operating model must be consistent across the company and which must sit close to the work.

A realistic example of the shift
Imagine a mid-sized company that wants to use AI in customer support.
The first pilot goes well. An AI assistant reads incoming tickets, pulls relevant account history, and drafts replies for agents to review. Support staff say it saves several minutes per ticket. Leadership is pleased.
Then the company tries to scale it.

The security team asks which customer data the assistant can access and how that access is logged. Legal asks who is accountable if the assistant gives a customer incorrect information. Support managers notice that faster drafts are now piling up in the same review queue, so resolution times barely move. Product teams realize the assistant keeps citing an outdated help article because nobody owns the knowledge base. Agents are unsure whether they are now writers, editors, or something else entirely.
None of these are model problems.
They are operating model problems.

The company that moves forward is the one that steps back and redesigns support as a whole. It decides which request types the AI can resolve on its own, which need a human to approve, and which go straight to a specialist. It assigns a clear owner for the knowledge base. It gives agents a new role focused on complex cases and quality review. It sets measures based on end-to-end resolution, not draft speed. And it gives the AI more autonomy only as accuracy is proven, one request type at a time.

This kind of redesign is not theoretical. McKinsey describes a global financial services company that chose to redesign enterprise workflows from the ground up rather than deploy AI as disconnected use cases. Its teams tested copilots inside the redesigned processes, saw rapid productivity gains, and expect cycle times for priority processes to fall by more than half. In another example, an electronics retailer first invested in data governance, engineering standards, and an internal AI factory before scaling AI across e-commerce, pricing, and fulfillment. That transformation contributed to an estimated 35 percent increase in EBITDA in its first year.
The technology made those results possible.
The operating model made them real.<br />Roles, skills, and incentives have to move too.
AI changes what people do. That means it also changes how people should be hired, developed, managed, and rewarded.
This is the part of the operating model most companies postpone the longest.

Job descriptions still assume work that AI can increasingly handle, such as first drafts, data gathering, routine analysis, and standard triage. When those tasks shrink, the role does not vanish. It changes shape. The analyst becomes a reviewer. The coordinator becomes an exception handler. The manager leads a mix of people and agents.

If no one redesigns those roles deliberately, people end up doing their old job plus checking the AI. That creates fatigue, not productivity.
The workforce question is also more pressing than it was a year ago. McKinsey’s 2026 survey found that 39 percent of respondents expect AI to reduce their organization’s total headcount over the next year. Yet only 14 percent reported an AI-related decline over the past year, less than half the share who had expected one.
That gap between expectation and reality is telling.

Workforce change from AI is slower and messier than leaders predict, largely because the operating model has not changed enough to absorb it.

McKinsey’s operating model research points to two shifts that help. Organizations reinventing work with AI are more likely to use skills-based talent models, where people are staffed and developed based on capabilities rather than static job titles. They are also more likely to use hybrid rewards that recognize both individual contribution and team outcomes, which discourages each function from protecting its own territory.

AI-native companies add one more layer: adoption as a flywheel rather than a rollout. McKinsey’s interviews describe four reinforcing habits. Leaders use AI visibly themselves. Teams share what works so successes spread. Adoption is measured openly. And hiring screens for curiosity and willingness to experiment with AI, even in nontechnical roles.

When any one of those habits is missing, AI tends to spread in pockets and never becomes an organizational capability.
What leaders should do in practice
Leaders do not need to redesign the entire company before using AI.
They do need to stop treating the operating model as something that will sort itself out later.
That starts with a few honest questions.

Are we measuring AI by the number of tools deployed, or by the processes that now run differently?
For each AI initiative, do we know exactly who owns the outcome?
Are risk, legal, and security shaping the design, or only approving it at the end?
Can our agents actually reach the knowledge they need, and who keeps it current?
What will people in the affected roles do differently, and have we told them?
Those questions usually reveal where the operating model is holding AI back.

From there, the practical moves are fairly clear. Pick one or two domains where AI can change the end-to-end process, not just one task. Redesign those workflows with business, technology, risk, and operations teams together. Centralize the platform, standards, and guardrails, but give business teams real ownership of execution. Grant AI autonomy step by step as trust is earned. Redesign roles, skills, and incentives at the same time as the technology, not after it.

Governance cadence matters too. McKinsey’s research found that organizations reinventing work with AI rely less on multi-year rolling plans and more on quarterly outcome reviews or dynamic reforecasting, so they can scale what works and stop what does not.
Not every AI initiative deserves to survive.
An operating model that can redirect resources quickly is worth more than a portfolio that keeps every pilot alive out of politeness.
The operating model is the AI strategy
Most companies do not lack access to AI.

The same models, platforms, and vendors are available to nearly everyone. That means the technology itself is rarely a lasting advantage.
The advantage comes from how quickly and safely an organization can turn that technology into a different way of working.
That is an operating model capability.

It shows up in how decisions are assigned, how workflows are designed, how governance is embedded, how knowledge is owned, and how people are organized and rewarded. It shows up in whether a successful pilot becomes the new standard way of working or quietly fades away.
That is why AI demands a new organizational operating model.
Not because the old one was badly designed.
Because it was designed for a world where people did all of the work, and every structure, control, and role assumed that.

AI changes that assumption.
AI-native organizations have already accepted it. The rest still have a choice: redesign deliberately now, or keep layering new tools on old structures and wonder why the value never arrives.

Leave a Reply

Your email address will not be published. Required fields are marked *

Ready to Understand What's Holding Progress Back?

Schedule a Bottleneck Analysis and gain clarity on the challenges affecting your organization's ability to move forward.

Book A Consultation