A lot of AI work begins with the tool.
A team sees a new capability, identifies a task it could perform, and starts looking for a place to use it. That approach can produce a working automation. It doesn’t necessarily produce a better business.
I think AI transformation leadership starts from a different premise: AI is an execution capability, not a strategy. The work is to design an operating system in which people, processes, information, and technology produce better outcomes together.
That changes the first question. Instead of asking, “Where can we use AI?” I’d ask, “What outcome are we trying to improve, and what is getting in the way today?”
Sometimes AI will be part of the answer. Sometimes the process needs to be simplified first. Sometimes the information is scattered across systems, ownership is unclear, or people are compensating for a policy that no longer makes sense. And sometimes the answer is a basic rule, query, or automation that is more predictable and less expensive than AI.
Choosing among those options is the work.
Start with the work, not the technology
Imagine a company wants to use AI to speed up customer issue resolution.
The visible task might be writing replies. But when you look more closely, the delay may come from somewhere else. Employees could be searching several systems for account history, interpreting inconsistent policies, asking another team to approve exceptions, or trying to determine who owns an unusual case.
Generating a response faster won’t resolve those problems. It may simply produce an answer before the organization has the information or authority to act on it.
Before automating a workflow, I want to understand how the work actually moves:
- What outcome matters?
- Where does the work slow down or fail?
- Which steps require judgment?
- Which steps are mostly searching, copying, comparing, routing, or updating?
- What information do people need, and can they trust it?
- Who owns the decision?
This is why operational design has to come before scale. AI amplifies the system around it. A clear, reliable process may become faster and more consistent. A fragile process may become harder to inspect while producing mistakes at greater speed.
A useful test is simple: if this process ran ten times faster, would we be comfortable with output?
If the answer is no, scaling the tool is premature.
Design the human and AI roles together
Good AI transformation doesn’t begin with a target for removing people from a workflow. It begins by deciding where different kinds of work belong.
AI can be useful when the work involves speed, volume, consistency, summarization, or finding patterns across a large amount of information. People are still needed where context, judgment, accountability, relationships, and exceptions matter.
That division should be intentional.
In the customer issue example, AI might gather relevant account information, summarize prior interactions, identify the applicable policy, and draft a response. An employee might decide whether an exception is appropriate, account for details the system can’t see, and own the communication with the customer.
The exact mix will depend on the risk and the work. A draft for internal review is different from an autonomous decision that affects someone’s access, money, employment, or care. “Human in the loop” is not specific enough. Leaders need to define what the person is reviewing, what evidence they receive, what authority they have, and when the issue should be escalated.
Most importantly, accountability remains human.
AI may recommend or execute a decision, but it can’t own the outcome. If no one can answer who is responsible when the system is wrong, the workflow is not ready.
Treat attention as part of the design
Adding human review can make an AI workflow feel safer. It can also create a new operational burden that no one has measured.
Every review queue, validation step, exception, and approval consumes attention. If employees must inspect hundreds of low-risk outputs to catch a small number of meaningful problems, the workflow may be technically automated while the organization becomes more distracted.
Human attention is one of the most limited resources in a business. It should be spent where judgment changes the outcome.
That means review should be designed around risk. High-impact or uncertain cases may need careful oversight. Routine cases with strong evidence may need sampling, threshold-based review, or monitoring instead. The system should make it easier for a person to see why something was recommended, not force them to reconstruct the reasoning from scratch.
The right question isn’t simply whether a human is involved. It is whether the workflow uses human attention well.
Make the tradeoffs visible
There is no AI implementation without tradeoffs.
Faster responses may be less accurate. More oversight may reduce risk while adding delay and cost. Greater consistency may leave less room for unusual but legitimate cases. A higher-quality model may be too expensive for the volume of work.
These aren’t failures to eliminate. They are decisions to make openly.
For each workflow, leaders should be able to explain what the organization is optimizing for, what it is willing to give up, and where the limits sit. Those choices should reflect the consequences of being wrong.
A low-risk internal summary and a customer-facing eligibility decision should not share the same accuracy target, approval process, or monitoring standard. Applying one governance model to every use case usually creates one of two problems: too little control where the stakes are high, or so much control that low-risk improvements become impractical.
Good governance creates clear boundaries without pulling every workflow decision away from the people who understand the work.
Security standards, risk categories, evaluation methods, and shared platform capabilities can be centralized. Workflow design and improvement should remain close to the teams responsible for the outcome. That balance gives the organization consistency where it matters and flexibility where local knowledge matters.
Build learning into the workflow
Launching an AI-enabled process is the beginning of the operating work.
Models change. Business rules change. Customer behavior changes. The information feeding the system changes. A workflow that performs well at launch can become less useful without producing an obvious technical failure.
Every AI-enabled process needs a way to learn from what happens after deployment.
That includes measuring the business outcome, capturing errors and exceptions, watching for changes in performance, and reviewing whether the human workload is moving in the right direction. It also means separating model problems from process problems. A poor result may come from the model, but it may also come from outdated information, unclear instructions, conflicting policies, or missing ownership.
The point isn’t to collect an impressive dashboard of AI activity. The point is to improve the work.
The number of projects, automations, prompts, or licenses tells us very little about whether the organization is better off. A useful measure connects the system to an operating result: resolution time, error rate, rework, cost, customer retention, employee capacity, or another outcome the business already cares about.
If no meaningful measure should change, it is worth asking why the initiative exists.
Design for change without starting over
No organization should have to redesign itself every time a model or vendor changes.
Adaptability comes from separating the business workflow from the technology performing individual steps. The organization should retain its process knowledge, decision rules, evaluation criteria, data ownership, and escalation paths even when the underlying tools change.
Data is central to that resilience. AI performance depends on the quality, accessibility, and ownership of the information around it. If policies conflict, records are incomplete, or no one is responsible for maintaining the source material, changing the model won’t solve the underlying problem.
This is why data quality and governance are operating capabilities, not technical cleanup. Someone needs to own the information, understand where it comes from, decide when it is trustworthy, and keep it current.
The same is true of organizational knowledge. A company becomes more capable when its people learn how to assess a workflow, distinguish judgment from coordination work, evaluate risk, and decide what kind of solution fits. That capability remains useful even when the technology changes.
Vendor expertise can be purchased. Operational judgment has to be developed inside the organization.
The mission of AI transformation
The role of AI transformation is to design and continuously improve the system that allows people and AI to work together responsibly and at scale.
That includes choosing the right problems, fixing weak processes before accelerating them, defining human and AI roles, protecting attention, clarifying accountability, governing risk, improving data, measuring outcomes, and learning from results.
Deploying AI is one activity within that discipline. It is not the mission.
The mission is to build an organization that can use changing technology without losing sight of how work gets done, who owns the consequences, and whether the business outcome actually improves.