I started working directly on AI operating problems inside a large consumer business in 2023.
In 2024, I moved into a newly created AI Transformation role at another consumer platform. The executive who created the role believed AI Transformation would eventually become a standard function inside companies.
At the time, that was mostly a prediction. There wasn’t an established career path, a common job description, or much agreement about what the function should own.
The titles are becoming more common now. But I don’t think the creation of another job category is the most important change ahead.
I think AI Transformation is an early version of a capability that will change what it means to be a COO.
As AI begins performing work across functional boundaries, the COO becomes responsible for something new: making sure the decisions about people, systems, information, authority, and risk still add up to one coherent operation.
That doesn’t mean every AI team should report to the COO. It means someone has to design how the pieces work together.
The question isn’t who owns AI
Companies tend to approach AI ownership the same way they approach other executive responsibilities. They look for the function with the strongest claim.
Technology leaders own the systems. Data leaders own the information. Legal and risk leaders establish boundaries. Finance governs investment and value. HR helps define how jobs and responsibilities change. Functional executives remain accountable for their business outcomes.
All of those claims are legitimate.
That is why asking one executive to “own AI” doesn’t resolve very much.
An AI system operating inside a customer workflow may use company data, recommend a commercial decision, communicate with the customer, update the system of record, create a support obligation, and expose the company to financial or reputational risk.
No single function owns every part of that work.
The more useful question is: Who owns which decisions, and who makes sure those decisions fit together?
That second responsibility is where the COO becomes central.
AI can cross the boundaries the company still depends on
Most companies are organized into functions. Work is not.
A customer outcome may depend on Marketing identifying an opportunity, Sales making a promise, Product delivering the experience, Finance recognizing the revenue, and Support handling whatever happens next.
The org chart assigns responsibility for each part. It does not automatically create ownership of the combined outcome.
People have traditionally carried much of that coordination. They search for information, ask another team what a policy means, copy details between systems, route unusual cases, remember who needs to approve something, and follow up when the next step doesn’t happen.
A lot of this work is nearly invisible because capable employees keep doing it.
AI can remove some of that coordination. A system can gather context, compare records, prepare a recommendation, route the next step, update a tool, and notify the right person.
But as soon as it does, the company has to make decisions that were previously buried inside human effort.
Which information can the system trust? What can it change? Which rules apply? When should it stop? Who receives the exception? What evidence does that person get? Who is responsible when the workflow produces the wrong result?
Those decisions may involve several executives. Someone still has to make them form a workable operating model.
Authority has to be designed into the work
Traditional software usually follows logic and workflows established in advance. AI systems can respond to less structured information, generate intermediate steps, and select among possible actions.
That makes authority part of workflow design.
At an enterprise software company, I mapped more than 100 go-to-market workflows and hundreds of underlying tasks. The original question was where AI could create the most value.
Once we looked closely at the work, the harder problem became visible.
The workflows crossed teams and systems. Important context lived in people’s heads. Similar processes were performed differently across functions. Some employees had the information but not the authority to act. Others had the authority but had to wait for someone else to reconstruct the context.
“Automate go-to-market” was not a useful instruction. Even “automate this workflow” was often too broad.
Authority had to be decided task by task.
A system might be allowed to gather account information without approval, prepare a recommendation for review, update a low-risk field automatically, and stop before sending a customer message or changing a commercial commitment.
Each step could require a different level of evidence, oversight, and permission.
Engineering can implement those controls. Legal can identify prohibited actions. Security can govern access. The functional leader can define the outcome.
The COO’s job is to make sure the full workflow can operate, not simply that each function completed its portion of the design.
Human involvement needs a purpose
“Keep a human in the loop” sounds responsible. It isn’t specific enough to build around.
Which person is involved? What are they expected to notice? What evidence do they receive? Can they disagree with the system? Can they change the result? Do they have enough time to perform the review well?
If an employee is asked to approve an AI-generated decision but can’t understand the basis for it, they aren’t providing meaningful oversight. They are absorbing accountability for a system someone else designed.
A multiyear field study of employees held responsible for AI-generated decisions found that workers actively interpreted and reshaped outputs before presenting them to customers, managers, and colleagues.
That work shouldn’t be treated as friction around an otherwise successful system. It is evidence that the workflow still depends on human context and judgment.
A review step is useful only when the person can make a better decision because they are there.
That may mean giving the employee access to supporting evidence. It may mean giving them authority to override the result. It may mean routing only uncertain or consequential cases to them instead of asking them to inspect everything.
Human presence is not the same as human authority.
The COO will need to understand the difference.
Attention becomes an operating constraint
AI can increase production without increasing useful capacity.
A system can create more recommendations, reports, alerts, customer messages, and decisions than employees can reasonably inspect. Work that disappears from one part of the process may reappear in review queues, escalations, corrections, and investigations.
The company may look more automated while its employees become more distracted.
I saw the other side of this while rebuilding contract-to-revenue reconciliation at an enterprise software company. The process had required about a week of manual Finance work each month.
In that workflow, 89% of contracts reconciled automatically end to end. That did not eliminate Finance’s role. It changed where Finance’s attention went.
Instead of manually reconstructing every agreement, the team could concentrate on missing information, unusual terms, conflicting records, and the cases where financial judgment was actually useful.
That distinction matters.
If employees must inspect hundreds of predictable results to catch a few important problems, the workflow is consuming expertise without benefiting from it. If clean cases can move through known boundaries and uncertain cases arrive with the right context, automation creates real capacity.
The COO already plans headcount, budget, and technical capacity. Human attention will need the same care.
Exceptions show where the operating model breaks
As AI handles more predictable work, employees may receive a higher concentration of unusual, ambiguous, or consequential cases.
Those exceptions are easy to treat as cleanup. I think they are much more valuable than that.
An exception shows where the system reached the edge of what the company understands well enough to automate.
The policy may be unclear. Necessary information may be missing. Two systems may disagree. A customer’s situation may not fit the company’s assumptions. The system may have reached the boundary of its authority.
Repeated escalations can reveal a broken process. Consistent overrides can expose a bad rule. Missing information can point to an ownership problem. Customer complaints can show that an efficient workflow is producing the wrong experience.
The exception queue may become one of the COO’s clearest views into the real operation.
That requires more than closing each case. The company needs to capture what happened, why the system stopped, what the employee decided, and whether the same situation should be handled differently next time.
AI Transformation can build that learning loop. The COO can make sure the company acts on what it reveals.
Performance has to mean operating performance
Many companies measure AI progress through licenses, usage, pilots, or the number of automations launched.
Those measures show that activity occurred. They don’t show whether the operation improved.
The COO needs measures connected to the work: cycle time, errors, rework, cost, customer outcomes, employee capacity, and recovery when something goes wrong.
The tradeoffs behind those measures should remain visible.
A faster process may produce more errors. Greater oversight may reduce risk while consuming employee attention. More automation may improve consistency while handling unusual cases poorly.
Those aren’t technology tradeoffs alone. They are operating decisions.
The COO is already responsible for deciding which tradeoffs the company can accept. AI increases the number of places where those choices have to be made explicitly.
AI Transformation brings product discipline to the operating model
I don’t think every COO needs to design prompts, evaluate models, or build agents.
They do need a capability that can connect AI strategy to the way work actually happens.
AI Transformation can study the current workflow, find where coordination is consuming people’s time, identify which decisions require judgment, and determine where AI, conventional automation, process redesign, or a basic query fits best.
It can help define human and system responsibilities, establish boundaries for authority, coordinate technical dependencies, build feedback loops, and measure whether the business outcome improved.
That is product discipline applied to the operating model.
The product is not the agent. The product is the operation that people and AI perform together.
AI Transformation may sit in the COO’s office, but the capability has to extend across the company. Its value does not come from owning every AI initiative. It comes from giving the organization a repeatable way to decide where AI belongs and what has to be true before it can operate responsibly.
The COO owns coherence
The COO has traditionally been responsible for making the company work as one operation.
AI does not remove that responsibility. It changes what the COO is operating.
The company is no longer made up only of people, teams, processes, and conventional software. AI systems are beginning to interpret information, recommend decisions, communicate with customers, update records, and take action across functions.
Every one of those systems can work as designed while the combined operation still fails.
The data can be correct but unavailable at the right moment. The model can perform well while the workflow sends its output to someone who can’t act. Each function can fulfill its responsibility while no one owns the customer outcome. Automation can save time for one team by creating more work for another.
Coherence is the work of making those local decisions produce a dependable whole.
The COO will increasingly design an operation in which people and AI systems can act together without losing clarity about authority, evidence, accountability, or the outcome the company is trying to produce.
AI Transformation may begin as a function reporting to the COO.
Its larger contribution will be giving the COO a new way to see and design the company: not as an org chart connected by software, but as one operating system made up of people, judgment, information, and machines with carefully bounded authority.
That is the hybrid company.
Building it may become the defining work of the COO.