Key takeaway
When AI changes a shared workflow, define the owner, business result, quality guardrail, human rescue load, and review rhythm before treating it as permanent work.
Not every AI tool needs a new role. But once it changes how people work together, somebody has to define what changed, what good looks like, and how the business will learn.
Earlier this year, I joined a call expecting to talk about AI implementers.
Instead, a staffing operator corrected the way I was thinking about the job.
We had been comparing AI implementers with virtual staff. The logic sounded reasonable: bring in someone capable, point them at the company’s pain points, then let them find where AI can help.
He said the comparison had a problem.
When a law firm hires a case manager or legal assistant, the role already has a shape. The company generally knows what that person does, who manages them, and what acceptable performance looks like.
With AI, many of us are trying to hire or assign responsibility before the work has properly settled into a role.
His point was that even they were still figuring out what the job actually was.
I understood immediately because we were having the same conversations behind the scenes.
What exactly should an AI implementer own? How broad should the role be? Does the person build systems, train the team, maintain workflows, recommend tools, or somehow do all of it?
At first, this looks like a staffing problem.
But the staffing problem sits downstream of something bigger.
AI is changing the work before many companies have decided how that work should now be managed.
There is a useful way to look at this. The tool may not be useless, and the person may not be failing. The work may simply still be undefined. And undefined work can be designed.
Not every use of AI needs a new job
If you use ChatGPT to improve one email, you do not need to create an AI department.
You are using a tool inside work that already makes sense.
But AI can enter a business at several different levels. The confusion starts when we manage all of them as if they are the same thing.
Swipe sideways to see the full diagram.
At the first two levels, the existing person can usually own the work. They already know the goal. AI simply helps them complete a task.
At the third level, something important changes. The AI now affects handoffs, decisions, shared information, customer experience, or work that continues when the original user is not there.
That is the point where “everybody should use AI” stops being enough.
But the levels show how widely the work has changed, not how risky it is. One person using AI for a sensitive legal, financial, safety, privacy, or public-facing decision may need controls before the work ever crosses a team.
The moment AI stops being personal software
You do not need a complicated governance programme. You need to notice when an individual experiment has become shared work.
Several signs together are a strong signal, but one can be enough when the consequence of getting it wrong is serious.
If none of these are true, keep experimenting.
If several are true, you do not necessarily need a new employee. But you do need to manage the work.
An AI experiment asks, “Can this work?” Managed AI work asks, “Who keeps it useful after we discover that it sometimes can?”
When the work changes, five things need to become clear
Most AI projects begin with capability.
What can this model do? Can we connect it to the CRM? Can it answer customers, create content, build reports, and summarize meetings?
AI always has another answer when you ask, “What else?”
The business has to answer a different set of questions.
Swipe sideways to see the full diagram.
You still have a demonstration.
The output may look impressive, but nobody has changed what a real person does before, during, or after it.
You have a temporary experiment.
It continues while the builder has attention. When normal work becomes busy, the system slowly loses care.
You can see activity without learning.
The business knows how many outputs the AI produced but not whether the original work became better.
The same failure keeps returning.
A dashboard can record the problem. Only a review changes what happens next Monday.
The audit tells you where the gap is.
But filling the boxes badly is not much better than leaving them blank. So the next question is who should own the loop.
The owner is not automatically the person who knows the most AI
The builder understands the system. That does not always mean they understand the work well enough to own the business result.
A useful owner usually needs three things:
Swipe sideways to read the full table.
| Requirement | What it means | What happens when it is missing |
|---|---|---|
| Workflow proximity | They see where the work actually slows down, breaks, or requires judgment. | The company builds something technically interesting that does not fit Monday morning. |
| Decision authority | They can change the process, define exceptions, and ask other people to work differently. | They notice every problem but can only make suggestions. |
| Result accountability | The business outcome already matters to them, with or without AI. | AI activity becomes the goal because nobody is responsible for the actual result. |
Sometimes one person has all three. Often they do not.
In that case, separate the responsibilities instead of pretending one “AI person” can carry everything.
This is a fairer job for both people.
The workflow owner is not expected to become an AI engineer. The builder is not expected to invent business priorities from incomplete instructions.
Score the work, not how busy the AI looks
This is where many AI scorecards become strange.
We measure prompts sent, automations launched, agents created, articles generated, or hours the system ran.
We can use those numbers to see activity, but that still does not tell us whether the work became more valuable.
A useful scorecard needs more than one number because AI can improve speed while quietly damaging quality, creating more checking, or becoming impossible for anyone else to operate.
Swipe sideways to read the full table.
| Measure | Question | Examples |
|---|---|---|
| Business outcome | What should become meaningfully better if this works? | Time to complete the work, qualified replies, errors reduced, useful work reaching publication. |
| Quality guardrail | What must not deteriorate while the headline number improves? | Accuracy, customer experience, brand voice, compliance, work returned for correction. |
| Human rescue load | How much checking, correcting, escalating, and babysitting does the system still require? | Corrections per run, exceptions, review time, failures only the original builder can solve. |
| Repeatability | Can the workflow continue without depending on one person’s memory? | Another teammate can operate it, boundaries are documented, exceptions have a clear destination. |
The goal is not to create a perfect dashboard on day one.
Start with a baseline. What happens today without the AI? How long does it take? Where do errors appear? How much checking is already happening?
Then choose a change worth observing.
If you cannot describe what improvement would look like, the project may still be too broad.
The system that kept working through breakfast
I built a system to research competitors, find questions and keywords worth targeting, then create articles designed to rank.
It had a clear original job.
But whenever it finished something, I kept asking, “What else?”
AI always had another answer. Another feature. Another improvement. Another clever thing it could technically do.
Almost every morning, I carried my laptop to breakfast and connected it through my mobile hotspot because the program was still running.
The automation was supposed to save time.
Ironically, I actually rearranged how I had breakfast around it.
Eventually I asked the system:
“Your goal is supposed to get these things rank. Do you still remember or not? Are you actually working towards this goal? Or are you just writing articles?”
A version-one audit makes the problem visible:
Building became cheaper and easier, but complexity stays. The score is not there to prove the AI is clever. It helps us notice when the work is becoming more useful, and when we are simply becoming busier around it.
The review is where the operating model becomes real
You can have an owner and a score, then still learn nothing.
Somebody has to look at what happened and decide what changes next.
This does not need to become another long meeting. For one normal, low-risk workflow, a short weekly review is enough to begin.
Start with the result
Did the business outcome move, stay flat, or become harder to interpret?
Inspect the rescue work
Where did a human correct, check, reconnect, explain, or quietly complete the work themselves?
Find the repeating failure
Do not fix every strange output. Find the pattern creating the most cost, risk, or confusion.
Change one part of the workflow
Improve an input, boundary, prompt, handoff, exception rule, or responsibility. Then observe the next cycle.
Choose the direction
Keep it, narrow it, improve it, scale it, or stop it. Continuing unchanged is also a decision, so make it consciously.
If the workflow handles sensitive data, money, legal or safety decisions, or work published outside the company, add the controls it needs before release: human approval, privacy or security review, tighter monitoring, and a clear escalation path. The learning review sits above those safeguards. It does not replace them.
The review loop matters because AI work does not arrive fully formed.
The company learns what the role is by watching the work, noticing the exceptions, and improving the boundary between the person and the system.
So, do you actually need an AI role?
Maybe.
But hiring an “AI person” should not be the first reflex.
This is why I no longer think the question is simply, “Who owns AI?” The better question is:
What work changed because AI arrived, and what now has to be true for that work to become reliable, useful, and fair for the people carrying it?
Sometimes the answer is a new role. Sometimes it is an existing manager with clearer responsibility. And sometimes it is admitting that the experiment has not earned a permanent place yet. All three are better than hiring someone into a job the business cannot explain.
Start with one workflow this week
Do not redesign the entire company. Pick one repeated AI use and complete these nine lines.
You do not need all the answers before you begin.
You need enough clarity to give the person, the system, and the rest of the team a fair chance of learning what good work now looks like.
That, at least to me, is where AI starts becoming part of how a company actually works.
