“Hi Karl, I asked AI about our online marketing and it produced this report. Can you implement it all?”

As AI adoption grows, that email is no longer unusual. It is becoming routine.

At first, I felt mildly irritated. Then slightly threatened. Now it feels inevitable. But the more I reflect on it, the more I realise this is not fundamentally an online marketing problem. It is a leadership problem.

The real question beneath that email is not whether the recommendations are correct. It is this: what is your role when everyone suddenly has access to analysis?

If you are honest, the first reaction is rarely calm curiosity. It is defensiveness. You have years of experience. You have tested strategies, absorbed algorithm updates, and learned from campaigns that did not work. And now a language model, operating without business context, is offering advice on your work.

The instinct is to correct it quickly.

But when you attack the output, the sender often hears you attacking their initiative. Most clients forwarding AI suggestions believe they are being helpful. They are trying to contribute. If your tone becomes territorial, the conversation shifts away from online marketing and toward ego.

The more productive posture is simple, though not always easy. Validate before you evaluate.

A response such as, “Thanks for sending this over. There are a few ideas here worth exploring. I will review them and follow up with additional context,” accomplishes more than it appears to. It acknowledges effort, signals confidence, and creates space for thoughtful analysis. Only then do you begin evaluating the substance, not to prove the model wrong, but to improve the thinking.

To give you an example, recently, a B2B client forwarded AI analysis claiming competitors ranked because they focused their online marketing entirely on one solution category. The implied strategy was to narrow everything and build around a single flagship offering. There was a strategic idea buried in that recommendation. Category authority and positioning do matter. However, when I examined those competitor sites directly, it became clear that they ranked for their core solution while also publishing deeply across adjacent areas.

Their authority was not narrow. It was layered.

The response did not reject the model outright. It acknowledged the strategic insight while introducing evidence. The takeaway was not exclusive focus, but integrated authority. In these moments, you are not debating the AI output. You are upgrading the analysis.

At some point, and I tend to do this farly quickly, you also have to name the deeper issue. AI outputs are only as strong as the prompt and constraints behind them. Most clients at this stage provide none. They ask broad questions and receive broad patterns. Longer content. More pages. Expanded coverage. These are patterns, not strategy.

The real variable is context.

Once that became clear to me, the question shifted from “How do I respond to AI recommendations?” to “How do I shape the conditions under which AI produces recommendations?”

For each client, I am increasingly building a structured context layer. Not scattered notes in documents. Not knowledge that lives only in my head. But deliberate architecture that captures business model constraints, margin realities, sales cycle length, geographic focus, brand positioning, historical experiments, and internal resource limitations.

When that context exists, the prompts change.

Instead of asking, “What SEO recommendations do you have?” we ask:

“Given this margin structure, this sales cycle length, this brand positioning, and our historical organic performance data, what trade offs would you prioritise in our SEO strategy over the next two quarters?”

Now the model is forced to reason under constraint rather than generate surface level ideas.

Or instead of debating whether pages should exceed a certain word count, we might ask:

“Compare our page against the top five ranking competitors. Ignore raw word count. Evaluate information density, entity coverage, topical depth, and alignment with search intent. Where are we substantively thin relative to user decision stage?”

That single instruction removes one of the most common forms of lazy AI output.

In another case, we might prompt:

“Using the provided context, identify three plausible SEO initiatives. For each, outline the opportunity cost, required internal resources, and downside risk. Do not recommend all three. Conclude with which single initiative you would prioritise and why.”

AI is naturally expansive. Expertise is selective.

These prompts force selection.

They force trade offs.

They force prioritisation.

And that is where strategy lives.

What I am discovering is that adding context is not a one time setup. It is an ongoing process. Each quarter introduces new constraints, new experiments, and new lessons. Wins are recorded. Failures are documented. Assumptions are revised. When structured correctly, the context layer accumulates memory. It becomes iterative. It begins to reference its own history.

In a recent talk, Andrej Karpathy, an AI researcher known for his work at OpenAI and Tesla, described large language models as something closer to a new computing substrate, even likening them to an operating system where prompts function as programs and context acts as persistent state. That framing feels increasingly accurate in practice. The model itself is general. The differentiation happens in the state you maintain around it.

When we build a structured context layer per client, we are effectively creating persistent state. We are shaping the environment in which the model reasons. Over time, that state grows richer. It contains not just facts, but patterns of judgment. It reflects prior trade offs, failed experiments, and strategic boundaries.

In that sense, it begins to self evolve.

Not autonomously. Not mysteriously. But iteratively.

The more disciplined we are about encoding constraint and learning, the more aligned the outputs become. It starts to feel less like querying a generic intelligence and more like interacting with a system that understands how we think.

I increasingly suspect this is how practical, personal AGI may emerge. Not as a singular, all knowing system, but as deeply contextualized intelligence built around an operator, a team, or a company. Intelligence that is shaped by memory, constrained by reality, and refined over time.

The breakthrough is not raw capability.

It is structured context. In that sense, it begins to self evolve.

Not in a science fiction way. In a practical way.

The more accurately we encode constraint, history, and trade off logic, the more aligned the outputs become. Over time, the system reflects not just general SEO knowledge, but our specific judgment patterns. It starts to feel less like querying the internet and more like querying a structured extension of accumulated experience.

I increasingly feel this may be how personal AGI actually evolves in practice. Not as a singular superintelligence, but as deeply contextualised systems trained on the lived constraints of a specific operator or organisation. Intelligence that is not general, but grounded. Not omniscient, but aligned.

What surprised me most is how this same dynamic appears in interviews.

I see it when speaking with candidates for roles my clients are hiring for. On paper, candidates have never looked stronger. Resumes are polished. Accomplishments are quantified. Language is strategic. Yet in conversation, many responses feel frictionless. When I ask about a failed campaign, I often hear something that resembles a refined case study. There is growth and reflection, but little uncertainty. Little trade off. Little tension.

At times I find myself wondering whether I am speaking to a person or to an optimised response delivered through a person.

It is not dishonesty. It is convergence. We are all training on generated language. Everyone sounds analytical. Fewer people sound specific. And specificity is where judgment lives.

Because of this, I have changed what I listen for. I pay attention to miscalculations, to constraints that were underestimated, to trade offs that were consciously accepted. Real expertise includes scar tissue. It includes moments when outcomes were unclear and decisions carried risk.

We are entering a period where fluency is abundant. Clients sound strategic. Candidates sound polished. Outputs are structured and persuasive. But leadership is not fluency. It is discernment.

It is the ability to say that something sounds reasonable in theory but fails under real constraints.

The competitive edge is no longer access to ideas. It is judgment under limitation. And increasingly, the role of the modern operator is to encode that judgment into systems that sharpen rather than dilute it.

Clients will continue sending AI generated recommendations. Candidates will continue arriving optimised. We cannot compete with AI on fluency, and we should not try.

The advantage lies in context, trade off awareness, and disciplined interpretation.

In an AI saturated world, the rarest skill is not information.

It is discernment.

And discernment, whether human or systematised, must be built intentionally over time.