After two years of putting AI workflows into real marketing, strategy, content, and operational work, I stopped asking which AI tool was best and started deciding what each system should own.
Recently, I found myself moving between ChatGPT, Google Workspace, Gemini, Canva, and several other systems while working through a project.
That wasn’t unusual. What was different was how I was using them.
I wasn’t copying the same problem into five AI tools and seeing which one produced the best answer. ChatGPT was helping me think through the problem, challenge assumptions, structure the work, and determine what needed to happen next. Google Workspace held much of the information and finished business work. Gemini could execute within that environment. Canva had a defined production role. Other systems handled distribution, automation, publishing, or measurement.
I was still making the consequential decisions, but the systems around me had started assuming clearer responsibilities.
That was when something clicked.
For much of the past two years, I had been trying to get better at using AI tools. Now I was learning how to operate with AI systems.
Those are not the same thing.
A campaign made the difference visible
I could see the change clearly in a recent enterprise B2B marketing campaign.
The project had all the ingredients that make real marketing work more complicated than an AI demo. There was a business objective, an audience, product information, existing content, stakeholder input, campaign requirements, creative production, deadlines, measurement, and plenty of ideas that could easily consume time without making the campaign better.
A couple of years ago, I would have thought about AI tools primarily at the task level. I might have asked it to generate campaign ideas, rewrite a headline, summarize a document, or draft an article.
Those capabilities are useful. But none of them solves the larger problem of getting from an ambiguous business objective to coordinated work that people and systems can actually execute.
This time, I started with the decisions.
Before creating assets, I used ChatGPT to help interrogate the campaign logic. We worked through what the campaign needed to accomplish, what the buyer actually needed to understand, which assumptions were unsupported, what evidence we had, what belonged in the campaign, and what was merely an interesting distraction.
Only after those questions were sufficiently resolved did the work begin moving downstream.
The strategy informed the landing-page copy. Buyer questions influenced the supporting content. The campaign idea became creative direction another person could execute. Measurement requirements went into the marketing system responsible for tracking the campaign.
The important change wasn’t that AI created more of the campaign.
It was that I had begun using AI tools to help orchestrate the work required to produce it.
I had been solving the wrong problem
For a long time, I asked the same questions many people were asking. Which model is better? What can ChatGPT do that Gemini can’t? Should I use agents? What should I automate? Which new tool should I try?
Those were useful questions while I was learning because I needed to understand the capabilities.
Eventually, though, adding another capable system started producing diminishing returns.
You can have ChatGPT, Gemini, Claude, Canva, automation platforms, specialized AI applications, and dozens of other tools available and still have an operational problem. Information gets copied between systems. Different versions of documents appear. Applications perform overlapping jobs. Context disappears between steps. Eventually, the human becomes the integration layer, manually holding the whole thing together.
At that point, access to more intelligence isn’t the constraint.
The problem is orchestration.
The questions that matter become more operational: who should perform the work, where the authoritative information should live, what context needs to travel with the task, what should become repeatable, and where human judgment still creates value.
Those questions have become far more useful to me than debating which model wins the latest benchmark.
Building an AI buyer exposed another weakness
One of the most useful lessons came from something I built for my own marketing work.
I created an AI buyer simulator because I wanted another way to pressure-test marketing before releasing it. The idea was to put a message in front of a simulated economic buyer and expose unclear promises, objections, proof gaps, and reasons someone might ignore it.
Then I discovered that I had rigged my own test.
Not intentionally.
I had defined the simulated buyer in a way that made that person too likely to experience the problem my offer was designed to solve. The AI could produce detailed, persuasive feedback, but I had influenced the conclusion before the test even started.
So I rebuilt it.
The simulator now starts with a neutral buyer. Problems, urgency, budget, buying stage, and other important conditions stay unknown unless evidence supports them or I deliberately introduce them as test assumptions.
More importantly, I built a boundary into the process: the simulator’s reaction is a hypothesis, not customer evidence.
That lesson has influenced much more than the simulator.
AI can challenge the work without becoming the evidence for the work.
A model can explain why an idea is compelling. It can simulate objections, construct personas, compare alternatives, and deliver its conclusion with extraordinary confidence. None of that means a customer validated the assumption.
AI became more useful to me once I stopped confusing confidence of expression with confidence of evidence.
The tools became more useful when I gave them smaller jobs
The same principle started changing how I thought about the rest of my technology.
Google Workspace already contains much of the information and work I operate from, so Gemini has an advantage when execution belongs inside that environment.
Canva doesn’t need to become my strategy department. It needs enough strategic context and a sufficiently strong brief to perform the visual work it is designed to perform.
Descript doesn’t need to determine my content strategy before it edits or repurposes a video.
Automation shouldn’t decide what matters. It should reliably execute work that has already been determined to be repeatable.
And ChatGPT doesn’t need to become every one of those systems.
Its highest-value role for me has increasingly been somewhere else. I use it to interrogate objectives, structure complexity, explore alternatives, develop instructions, critique outputs, pressure-test decisions, and coordinate what should happen next. Then I send the work to the system best positioned to execute it.
That is a very different mental model from trying to find one AI platform that does everything.
ChatGPT isn’t actually my executive
I sometimes describe ChatGPT as my executive layer, but I’ve realized that shorthand gives the technology too much authority.
I am the executive layer.
I determine what matters, establish the objective, make consequential tradeoffs, approve important decisions, and remain accountable for the outcome.
ChatGPT has become something closer to an executive reasoning and orchestration layer. It can challenge my thinking and carry substantial amounts of work, but that doesn’t transfer accountability for the decision.
My operating model has gradually become straightforward: I decide and direct. ChatGPT helps me reason, challenge, structure, and orchestrate. Specialized systems execute the work they are best suited to perform. I review consequential decisions, and the results become evidence for what happens next.
That doesn’t mean keeping humans artificially inserted into every task.
Some work should become increasingly deterministic. When a process is repeatable, its inputs are understood, the rules are stable, and the consequences of an error are acceptable, there is little value in making a person manually perform every step.
Human attention is expensive.
The objective isn’t human involvement everywhere. It is human judgment where judgment creates value.
My test for new technology has changed
This has also changed how I evaluate new AI products.
I used to begin with what the tool could do. Now I start by asking where it belongs in the work.
That subtle change forces me to consider the business problem before the capability. What decision are we trying to improve? What information does the system need? Where does that information already live? What should happen after the AI generates its output? Who owns that next step? What requires review? What could become deterministic? What evidence will tell us whether the change actually helped?
If I can’t answer those questions, another AI subscription probably isn’t going to solve the problem.
This has also made me more willing to leave a tool alone.
An interesting capability does not automatically deserve a workflow, an integration, or a place in my operating environment.
Sometimes the right technology decision is subtraction.
AI adoption is becoming an operating-model problem
I suspect more organizations are approaching the same transition.
The first phase of generative AI was about access. Then came experimentation. People learned prompting, tried different models, generated content, created custom assistants, experimented with agents, and began automating work.
Those capabilities still matter.
But when an organization has several capable AI systems, the constraint starts moving. The harder problem becomes connecting customer needs, business objectives, people, information, processes, technology, and evidence.
That isn’t primarily a prompt-engineering problem.
It is an operating-model problem.
A company eventually has to decide who owns a decision, which system performs the work, where context lives, how work moves between systems, where controls belong, and how anyone will know whether the new process is actually better than the old one.
Those questions aren’t as impressive in a two-minute AI demonstration.
They are much closer to what businesses have to solve if they want AI to become part of how work actually gets done.
I learned my way into this
I didn’t sit down two years ago and design this operating model.
I learned my way into it.
I tried tools, wrote prompts, built workflows, overbuilt some of them, discarded others, and found places where automation genuinely helped. I also found places where technology simply moved complexity somewhere else.
Over time, I stopped treating every interesting AI capability as something that needed to become part of my system.
That progression reflects the way I increasingly think about learning itself: Learn. Apply. Transform.
Learn enough to understand what is possible. Apply it to actual work. Transform the process only after experience gives you evidence that the change is worthwhile.
That is where I am now.
I’m less interested in finding one AI that can do everything. I’m more interested in designing work so that people and specialized systems each do what they are best positioned to do.
Because the durable advantage probably isn’t access to an AI model nobody else can access. It isn’t collecting the largest number of applications, either. And it certainly isn’t automating every task just because automation is technically possible.
The advantage is knowing what deserves human judgment, what machines should execute, how those systems should work together, and how quickly what you learn can become better action.
The missing layer in my AI stack wasn’t another tool.
It was orchestration.
